@bhabbott

Quote:

My latest retro PC build was on a cardboard box. Sadly when I put the 386SX motherboard in a proper PC case it died after running for less than a minute.

For retro, I would rather use Amiga's Workbench GUI 1st with two-button mouse approach (WHDload) due to a closer similarity with my gaming PC's Windows 11 GUI 1st with two-button mouse instead of PC DOS text UI.

I would rather play with Amiga's Dark Forces port over PC's Dark Forces DOS version since the Amiga version just works e.g. no mucking with config with SoundBlaster, Autoexec.bat, and Config.sys.

I prefer that Commodore release an Amiga version of Mac LC-475 /Quadra 605 (68LC040 @ 25Mhz) for about US$1000 in Q4 1993. Imagine, A1000+ with 68LC040 @ 25Mhz with CD32 chipset for US$1000.

Quote:

I still think $24 million for an X68000 style case that can't even take real 5.25" floppy drives is a waste. Why don't they just release the CAD files so anyone can 3D print their own case? I guess they already took the $24 million so they have to deliver something.

The topic starter forgets that Sharp didn't design Motorola's 68000 i.e. Sharp selected an existing 68K ASIC CPU and designed a value-added custom multimedia chipset and case as a system integrator vendor.

This website is not 68Kworld.net.

Commodore's PA-RISC selection was open to existing licensed designs from Hitachi. Hint: cache structure.

The key asset for the Vortex X86 clone vendor is its ASIC engineers.

Last edited by Hammer on 26-May-2025 at 12:27 PM.
Last edited by Hammer on 26-May-2025 at 12:19 PM.
Last edited by Hammer on 26-May-2025 at 12:13 PM.

_________________

Hammer Quote:

That Mechanical Keyboard Amiga 1200's €185.50 has the usual Amiga tax since the Corsair K70 core (MX Red) has €99.99 via Corsair's online store. https://de.pcpartpicker.com/product/NkRwrH/corsair-k70-core-rgb-wired-gaming-keyboard-ch-910991e-de Corsair K70 core (MX Red switches) has €82.00 via Amazon DE. PS; I prefer MX Red switches.

There are plenty of options with 68k Amiga hardware currently, which is good, but hardware value is too low to attract mass sales. At least the Amiga 1200 MiSTer+MiSTress hardware is very compatible/accurate and it is possible to use completely new hardware which allowed the former Amiga user in the video to return to the 68k Amiga.

Amiga 1200 MiSTer+MiSTress
+ compatible/accurate
+ upgradeable compatibility/accuracy
+ very versatile (supports other FPGA cores)
+ more modern I/O
+ completely new parts available or original Amiga 1200 case and keyboard can be used
+ quality looks good
- mediocre performance (downgrade for 68040 and 68060 users)
- not expandable (64 MiB max memory, no CPU or 3D upgrades, no clock ports, no parallel port, etc.)
- mediocre value/price
- assembly required

It is a nice 68k Amiga hardware option but is not cheap and easy enough for the 68k Amiga masses to return.

Hammer Quote:

https://liliputing.com/odroid-h4-is-a-mini-pc-board-with-intel-alder-lake-n-that-sells-for-99-and-up/ ODROID-H4 is a mini PC board with Intel Alder Lake-N that sells for US$99. Intel is competing against the likes of RPi 5.

There is crossover although the RPi 5 is lower power than the ODROID-H4 x86-64 hardware.

SBC | process | TDP | peak power | power supply
RPi3 40nm 4W? 6W 12.5W
RPi4 28nm 5.5W? 8W 15W
RPi5 16nm 8W? 12W 27W
OD-H4 7nm 12W 18W? 60W (133W with SATA)

The RPi powers in the table above include the whole SBC where the ODROID-H4 is just the Intel SoC which is a multi-chip package (MCP) with multiple dies on one chip and it is not clear whether the power is just for the CPU/GPU die or I/O die but the larger power supply requirement hints at a significantly higher power requirement for the SBC. The ODROID-H4 CPU cores can be clocked down to save power but the price efficiency (performance/$) is reduced and the price is high using the smaller process. The recommended customer price in 1,000 unit quantities for the N97 SoC was $128.00. The price has no doubt come down and larger quantities purchased but this SoC is likely more expensive than the RPi 5 SoC and will likely have a shorter life before being discontinued by Intel.

https://www.intel.com/content/www/us/en/products/sku/233090/intel-processor-n97-6m-cache-up-to-3-60-ghz/specifications.html
https://www.intel.com/content/www/us/en/products/sku/231805/intel-core-i3n305-processor-6m-cache-up-to-3-80-ghz/specifications.html

The N97 and i3-N305 CPU cores and GPUs likely greatly outperform the RPi 5 SoC with the latter mobile CPU cores approaching 10 year old desktop core performance. A large library of x86-64 software including games can be used on the x86-64 hardware although using Windows legally would add to the cost where Linux is free. The RPi 5 can be used passively with a large heat sink in cooler environments where the ODROID-H4 hardware would be handicapped trying to use it with passive cooling. There are a little bit lower power x86-64 SoC options down to about 6W TDP but it looks like they do so by lowering the clock speed.

https://en.wikipedia.org/wiki/Gracemont_(microarchitecture)

The Atom x7211E only has 2 cores but can still turbo boost to 3.2GHz with 6W TDP and a recommended customer price of only $39, less than 1/3 of the N97 SoC price. This would be fine for retro emulation but x86-64 game performance would drop with half the cores and the weaker GPU. The majority of low end x86-64 customers likely choose the hardware for gaming as ARM has the lower power embedded market. Intel used to have lower power in-order Bonnell microarchitecture Atom x86-64 cores which had better performance and power efficiency (performance/W) than in-order ARM cores but Intel gave up on them or they used too much area/transistors in comparison to then lower priced ARM cores. The in-order Atom SoC SBCs could be passively cooled using a 45nm process circa 2009.

http://www.linuxtech.net/features/intel_atom_pineview_motherboards_overview.html Quote:

The market for low power draw (below 20 Watts), passively cooled, small form factor motherboards was until recently dominated by rather expensive VIA boards and to a smaller extent AMD (with the Geode). Even last years' first generation Intel Atom cpu did not fit into this category, as first generation Intel Atom motherboards (with the commonly used power-hungry Intel 945GC chipset or even with the Nvidia ION) typically consumed 20-35 Watts and therefore usually needed a fan on the chipset (see our detailed TDP comparison).

Bonnell Atom SoCs were further die shrunk to 32nm and used PowerVR GPUs with the SBC only using a 36W power supply.

http://www.linuxtech.net/reviews/intel_DN2800MT_cedarview_atom_power_draw.html Quote:

PSU: ELV 87586 36W power brick (out: 12V DC, in: 230V AC 50Hz, efficiency level IV)

There were Bonnell cores that were under 1W TDP like the Silverthorne Z515@1.1GHz using a 45nm process.

https://en.wikichip.org/wiki/intel/microarchitectures/bonnell

These in-order cores could likely be running at twice the clock speed or more while dissipating only 1W on more modern silicon like the 7nm N97 SoC cores. The x86-64 ISA is fat and most customers want up scaled cores to take advantage of the large software library. It would be nice if there was a good CISC ISA that was not as fat, lower power and had a large software library that did not require as much performance. Passive cooling is a major cost advantage as more expensive power supplies, fans and cases are avoided. This is why the RPi 4 is more practical than the RPi 5, especially for embedded use like the PiStorm.

Hammer Quote:

The topic starter forgets that Sharp didn't design Motorola's 68000 i.e. Sharp selected an existing 68K ASIC CPU and designed a value-added custom multimedia chipset and case as a system integrator vendor. This website is not 68Kworld.net. Commodore's PA-RISC selection was open to existing licensed designs from Hitachi. Hint: cache structure. The key asset for the Vortex X86 clone vendor is its ASIC engineers.

Commodore was planning to produce a 68k Amiga SoC which would reduce the cost by $100 USD for Amiga computers and consoles according to Commodore internal documentation. PA-RISC systems and peripherals without the 68k Amiga SoC and 68k Amiga compatibility would not have been Amigas. Commodore started with commodity CPUs but was not limited to commodity CPUs. They licensed PA-RISC and licensed the 68k or planned to license it. They failed to integrate the 68k Amiga fast enough which is a major reason why they failed. A-EonKit failed to cost reduce their hardware enough resulting in poor sales and having to sabotage the competition like the Mirari board and RGL THEA1200 to protect their products. They should have invested in more competitive hardware.

@matthey

That's a flawed comparison when ODROID H4 has more ports, 64-bit DDR5 4800 up to 48 GB SODIMM and a sizeable integrated GPU i.e.

1 x HDMI 2.0,
2 x DisplayPort 1.2,
2 x USB 3.0,
2 x USB 2.0,
1 x 2.5 GbE LAN (Intel I226-V),
1 x 3.5mm audio out,
1 x 3.5mm audio in,
1 x SPDIF out,
M.2 2280 connector,

iGPU: Intel UHD 24 execution units up to 1.2 GHz (460 GFLOPS FP32, 960 GFLOPS FP16 pack math).

Raspberry Pi 5 utilizes a 32-bit LPDDR4X-4267 memory controller. Slow Genshin Impact performance.

https://youtu.be/pXo7GOyb7MY?t=303
Raspberry Pi 5 running Doom 3 from Xbox 1 era.

https://www.techpowerup.com/gpu-specs/uhd-graphics-24eu-mobile.c4146
Intel UHD 24 execution units at 1200 MHz have 460.8 GFLOPS FP32, 15 watts TDP.

https://www.youtube.com/watch?v=7kp_ANA2vMk
Intel Processor N97 + Intel UHD Graphics 24EUs (Alder Lake-N) with 10 modern games gameplay test including Counter-Strike 2 (Source Engine 2), Dota 2 (Source Engine 2) and Genshin Impact (Unity 3D from 2017). This example has the 1.2 Ghz UHD 24 IEU model.

For modern DX12-class Vulkan games, the Intel solution crushed them all.


https://youtu.be/UBQosbjl9Jw?t=629
Raspberry Pi 5 running Half Life 2 (Source Engine 2) with jerky frame rates

There are lower wattage PSUs with Intel N97 low-end ultrabooks e.g.
https://www.kogan.com/au/buy/blackviewofficialstore-blackview-laptops-acebook-8-156-in-intel-12th-n97-16gb512gb-windows-11-notebook-european-charger-acebook8-grey-eu
This Intel N97 based laptop bundles with a 36W Charger Adapter.

Desktop deployment needs to budget enough PSU beyond the fixed laptop power budget.

The full desktop PEG (PCIe for graphics) slot has 75 watts specification standard. Any desktop solution that targets PEG (PCIe for graphics) use case must be able to provide up to 75 watts from the PCIe slot.

Quote:

Commodore was planning to produce a 68k Amiga SoC

Commodore-Amiga Inc's proposed A1200-like SoC has a copy-n-paste 68EC020 CPU IP from Motorola and in-house value-added 2D GPU design.

This is equivalent to Broadcom's copy-n-paste ARM CPU IP from ARM and in-house value-added GPU (Video Core family) design ability.

On semi-conductor business, Commodore is roughly equivalent to Broadcom's ability.

Your argument glossed over the additional work required to synthesized FPGA TG68 into ASIC.

Prior to RPI 2040 "cut-n-paste" SoC, RPi was purely system integrator play with SoC coming from Broadcom.

Commodore-Amiga Inc's proposed A1200-like SoC proposal wasn't proven since CSG's best in-house production-ready CPU/2D GPU result is the C65 chipset via CSG's 2 microns fabs. AA Lisa requires 1.5 micron fabs via HP.

CSG's GMT Microelectronics successor reached 1 micron fabs.

Remember, Motorola blocked a Japanese 68000 1 IPC redesign with MUST be fabricated with Motorola's fabs requirement. You wouldn't get the X86 CPU clone market with Motorola's tight IP control!

Softbank(JP)'s ARM tried the tight IP control against Qualcomm's independent AArch64 implementation. Qualcomm successfully defended its use of ARM ISA in its chips, and the jury sided with Qualcomm on the core issue of whether the company had breached its licensing agreement with ARM.

SiFive's RISCV designs are commercial RISCV implementation and they must be paid by SoC integrator like StarFive.

Last edited by Hammer on 28-May-2025 at 04:27 AM.
Last edited by Hammer on 28-May-2025 at 04:25 AM.
Last edited by Hammer on 28-May-2025 at 04:19 AM.
Last edited by Hammer on 28-May-2025 at 03:48 AM.
Last edited by Hammer on 28-May-2025 at 03:40 AM.
Last edited by Hammer on 28-May-2025 at 03:25 AM.
Last edited by Hammer on 28-May-2025 at 03:10 AM.
Last edited by Hammer on 28-May-2025 at 02:51 AM.
Last edited by Hammer on 28-May-2025 at 02:16 AM.
Last edited by Hammer on 28-May-2025 at 02:13 AM.

_________________

@matthey

https://gadgetversus.com/processor/intel-processor-n97-vs-broadcom-bcm2712/
GeekBench compared Intel Processor N97 vs Broadcom BCM2712

Intel Processor N97 (quad-core Intel Gracemont CPU):
Geekbench 5 single core: 927
Geekbench 5 multi-core: 2,549
Geekbench 6 single core: 1,031
Geekbench 6 multi-core: 3,068

Broadcom BCM2712 (quad-core ARM Cortex A76 CPU):
Geekbench 5 single core: 609
Geekbench 5 multi-core: 1,626
Geekbench 6 single core: 748
Geekbench 6 multi-core: 1,507

----------
Intel Processor N97 GPU: 460.8 GFLOPS FP32 and 921.6 GFLOPS FP16 via Intel UHD Graphics 730 (Raptor Lake GT1).

Broadcom BCM2712 GPU: 51.2 GFLOPS FP32 and 102.4 GFLOPS FP16 via Broadcom Video Core VII.

This is like clubbing a baby seal.
----------

The Intel N97 processor is likely to have a supply lifetime of approximately seven years, according to ASUS. https://www.asus.com/networking-iot-servers/aiot-industrial-solutions/single-board-computers/n97s-im-aa/

Intel Skymont (included in Arrow Lake and Lunar Lake) is already displaced Gracemont (included in Alder Lake) and Cresmont (included in Meteor Lake).


https://en.wikipedia.org/wiki/Gracemont_(microarchitecture)
Intel's lowest asking price for a working quad-core Gracemont with 24 IEU GPU is Atom x7425E's $58 at a 1000-unit order. x7211E SKUs are recovered chips with fabrication defects, and their asking price is $39 at a 1000-unit order. Above the asking price is Intel's extra profit. My point, Intel still has room to maneuver.

As shown with the PPC Mirari desktop project, overclocking (e.g. 1.4 GHz to 1.6 GHz) is in play since there's a higher clocked variant.


https://videocardz.com/newz/amd-reportedly-won-contract-to-design-playstation-6-chip-outbidding-intel-and-broadcom
Both Intel and Broadcom participated in Sony's PS6 contract bid process, with AMD winning the PS6 contract. Broadcom was eliminated from the final round. Broadcom's offer would be an ARM CPU and an ARM GPU, since Broadcom's in-house GPU record is not sufficient.


https://www.tomshardware.com/video-games/playstation/intel-reportedly-lost-playstation-6-chip-design-contract-to-amd-in-2022-the-dollar30-billion-deal-failed-probably-because-arc-gpus-have-a-lot-to-prove
For the PS6 project, Intel's asking profit margin is higher than AMD's.

Last edited by Hammer on 28-May-2025 at 07:53 AM.

_________________

@matthey

Quote:

matthey wrote: There are plenty of options with 68k Amiga hardware currently, which is good, but hardware value is too low to attract mass sales. At least the Amiga 1200 MiSTer+MiSTress hardware is very compatible/accurate and it is possible to use completely new hardware which allowed the former Amiga user in the video to return to the 68k Amiga.

For me MiSTer fails out of the gate because it uses a DE10 Nano.


Quote:

Amiga 1200 MiSTer+MiSTress ... - mediocre performance (downgrade for 68040 and 68060 users) - not expandable (64 MiB max memory, no CPU or 3D upgrades, no clock ports, no parallel port, etc.) - mediocre value/price - assembly required

Such high standards Amiga fans have today. 'only' 64MB, 'downgrade from 040', and 'no 3D upgrade'. I feel so inadequate having 'only' 32MB and a 50MHz 030 in my A1200.

Quote:

It is a nice 68k Amiga hardware option but is not cheap

Not cheap, and also not an Amiga. It's not even a clone, just some addon boards for a DE10 nano.

Quote:

SBC | process | TDP | peak power | power supply RPi3 40nm 4W? 6W 12.5W RPi4 28nm 5.5W? 8W 15W RPi5 16nm 8W? 12W 27W OD-H4 7nm 12W 18W? 60W (133W with SATA)

RPis were infamous for drawing more power than expected compared to other boards people were used to, and famously didn't provide a specification for power draw.

Quote:

The RPi powers in the table above include the whole SBC where the ODROID-H4 is just the Intel SoC which is a multi-chip package (MCP) with multiple dies on one chip and it is not clear whether the power is just for the CPU/GPU die or I/O die but the larger power supply requirement hints at a significantly higher power requirement for the SBC.

Intel have always sucked at making low power devices. But nobody cares about that. What matters is that it smokes (perhaps literally) any real Amiga. 100, 1000, 10,000x faster? At some point you must be satisfied, right?

Quote:

The ODROID-H4... will likely have a shorter life before being discontinued by Intel.

Yet another product you won't be able to get.

That's why I have resolved to only make stuff with standard TTL gates from now on.

Quote:

Commodore was planning to produce a 68k Amiga SoC which would reduce the cost by $100 USD for Amiga computers and consoles according to Commodore internal documentation. PA-RISC systems and peripherals without the 68k Amiga SoC and 68k Amiga compatibility would not have been Amigas.

How lucky we are that Commodore went down when it did, before bastardizing the Amiga.

Last edited by bhabbott on 28-May-2025 at 08:06 AM.

Hammer Quote:

That's a flawed comparison when ODROID H4 has more ports, 64-bit DDR5 4800 up to 48 GB SODIMM and a sizeable integrated GPU i.e. ... For modern DX12-class Vulkan games, the Intel solution crushed them all.

I said the ODROID-H4 performance was roughly like a 10 year old desktop system, which is good, and it uses much less power, which is better yet. I expect passively cooling is not practical and there is additional expense for a power supply and case with airflow. The RPi 5 SBC may be more like a 20 year old desktop but it may be possible to use passive cooling in some environments and it can be powered from a 27W USB-C power supply which is cheaper than a 60W or greater dedicated power supply. The total system price is cheaper for the RPi 5 which also includes the memory. An under $100 USD lower memory RPi 5 system is possible while the cheapest ODROID-H4 system is going to be more like $150 USD.

ODROID-H4 SBC $99
8GiB memory $40
active cooler $10
60W power supply $20?
---
$169 for 8GiB system ($149 for 4GiB system, $139 for 2GiB system)

RPi 5 SBC $40 (base price but no zero memory option)
8GiB memory $40 ($5/GiB)
active cooler $10 (passive cooler $5)
45W USB-C power supply $17
---
$107 for 8GiB system ($87 for 4GiB system, $77 for 2GiB system)

The ODROID-H4 may offer more value if using x86-64 software but it is more expensive for a system and uses more power. RPis may offer more value for some uses like embedded use and retro emulation. They may be adequate for low end productivity software requirements much like the 68k Amiga for many years with less than the original $35 RPi specs and closer to the $5 RPi Zero or RPi Pico 2 specs (68060&AA+ use fewer transistors than the ARM SoCs of these SBCs). The 68k Amiga did not continue to be integrated or enhanced resulting in the high prices today.

Hammer Quote:

This Intel N97 based laptop bundles with a 36W Charger Adapter. Desktop deployment needs to budget enough PSU beyond the fixed laptop power budget. The full desktop PEG (PCIe for graphics) slot has 75 watts specification standard. Any desktop solution that targets PEG (PCIe for graphics) use case must be able to provide up to 75 watts from the PCIe slot.

If I/O is limited, the N97 SoC may be able to get away with as cheap of power supply as the RPi 5. The RPi 5 can get away with a cheaper power supply by limiting the I/O too. The ODROID-H4 does not have PCIe slots although it may be possible to add them. The 75W limit of PCIe from the PCIe slot is high for a low power SBC and low for a x86-64 GPU. Actually, it is possible to have capable 1080P 3D from the PCIe 75W slot in a low profile package.


https://yestonstore.com/products/yeston-rtx-3050-6gb-gddr6-gaming-graphics-card

The NVidia RTX 3050 low profile is perhaps the best performance from a GPU powered from the 75W PCIe slot but they usually sell for roughly $200 USD, more than a low power SBC with integrated GPU. The SBC to support them including power supply is usually several hundred USD as well. The PPC Mirari board has better I/O options than the ODROID-H4 SBC in my opinion with 2x NVMe without hats and 2x SATA while the last generation of PPC 28nm T series SoCs for the boards target a 30W envelope (likely peak power so maybe 20W TDP) which is less than twice that of the N97 SoC. The Mirari SBC could possibly be mass produced for under $500 USD but add the PPC SOC, a GPU, memory, power supply, active cooler and case and this system will likely cost closer to $1,000 USD. Reduce the CPU power (process improvement in the case of N97), integrate the GPU, eliminate the PCIe slots and reduce the SBC board size and the sub $100 ODROID-H4 SBC is possible with mass production. The only problem is that newer chip fab processes are expensive but an older fab process and lower power in-order CPU cores can be passively cooled for cheaper as the older Atom SBCs with PowerVR GPUs demonstrated.

http://www.linuxtech.net/reviews/intel_DN2800MT_cedarview_atom_power_draw.html

The power and area savings can be used for caches and the GPU which are more important for modern game performance.

CPU/core | pipeline | process | transistors
68060 8-stage 500nm 2,500,000
Pentium/P54C 5-stage 500nm 3,300,000
Atom/Silverthorne 16-19-stage 45nm 47,212,207
Atom/Moorsetown 16-19-stage 45nm 140,000,000 (Silverthorne with integrated GPU)
N97 14-19-stage 7nm 1,000,000,000+ (integrated x86-64 OoO CPU with 4 cores and GPU)

The x86-64 SoCs are often scaled up to gain performance and more x86-64 software support but they are fat and expensive as they use a smaller fab process to reduce power. Scaling x86-64 cores down into ARM territory is possible as demonstrated by in-order Atom/Bonnell cores but x86-64 software compatibility is lost and the cores are larger and use more power than ARM cores despite being more powerful and having better power efficiency than ARM in-order cores when they were released. Intel Bonnell in-order cores are weak performance compared to Intel low power OoO cores like the N97 uses but the cost advantage when reaching passive cooling is large and the expense between the Bonnell 45nm/32nm and 7nm/5nm necessary to reduce the OoO core power to that of the in-order cores is expensive.

Hammer Quote:

Commodore-Amiga Inc's proposed A1200-like SoC has a copy-n-paste 68EC020 CPU IP from Motorola and in-house value-added 2D GPU design. This is equivalent to Broadcom's copy-n-paste ARM CPU IP from ARM and in-house value-added GPU (Video Core family) design ability.

The CPU+chipset clock speed would have improved from 14MHz to 28MHz and 57MHz with double to quadruple the performance including the memory bandwidth. The 68020&AGA@14MHz were slow because of the low clock speed and memory bandwidth. The perception of AGA using SRAM on FPGA devices is not the same as Commodore failing to advance the Amiga chipset adequately.

Hammer Quote:

On semi-conductor business, Commodore is roughly equivalent to Broadcom's ability. Your argument glossed over the additional work required to synthesized FPGA TG68 into ASIC.

A TG68 core would not be a good candidate to turn into an ASIC. It is not as advanced as a 68EC020 core.

Hammer Quote:

Prior to RPI 2040 "cut-n-paste" SoC, RPi was purely system integrator play with SoC coming from Broadcom. Commodore-Amiga Inc's proposed A1200-like SoC proposal wasn't proven since CSG's best in-house production-ready CPU/2D GPU result is the C65 chipset via CSG's 2 microns fabs. AA Lisa requires 1.5 micron fabs via HP. CSG's GMT Microelectronics successor reached 1 micron fabs.

Outsourcing like the AGA Lisa chip to HP was the way to go, especially where Commodore did not have the capital to invest in their fab business. They should have sold their fab business which would have given them much needed cash for operations, R&D, debt, etc.

Hammer Quote:

Remember, Motorola blocked a Japanese 68000 1 IPC redesign with MUST be fabricated with Motorola's fabs requirement. You wouldn't get the X86 CPU clone market with Motorola's tight IP control! Softbank(JP)'s ARM tried the tight IP control against Qualcomm's independent AArch64 implementation. Qualcomm successfully defended its use of ARM ISA in its chips, and the jury sided with Qualcomm on the core issue of whether the company had breached its licensing agreement with ARM.

The 68k was still #1 in embedded 32-bit sales when Motorola bought the pipelined Japanese 68000 design, even though they were investing profits in PPC development. There is no 68k market to protect today if that was the reason then. Licensing 68k cores today should be relatively cheap as production of 68k chips does not compete with Motorola 68k chip production. The situation difference is only about 80 million 68k chips per year from then to now.

Hammer Quote:

SiFive's RISCV designs are commercial RISCV implementation and they must be paid by SoC integrator like StarFive.

SiFive is paid a license but it looked like there were no royalties when I looked into it. I did not do any licensing inquiries lacking a business but it looked like the licenses were reasonable too. It looked like they had royalty free reasonable licenses for low cost SoCs for SBCs. Replace the RISC-V cores with 68k cores and license a GPU like StarFive to produce a 68k VisionFive 2 SBC for under $100 USD, maybe less with mass production which I am not so sure the RISC-V VisionFive 2 has. Are they even selling a million units a year, especially with the RP2350/Raspberry Pi Pico 2 RISC-V cores dividing the small low end RISC-V market? Does RPi not want the RISC-V competition so they throw in a couple of nothing area RISC-V cores even though they are less powerful than the ARM Cortex-M33 cores?

bhabbott Quote:

For me MiSTer fails out of the gate because it uses a DE10 Nano.

It is possible to simulate the logic with the original timing of circuits on old silicon. Accurate simulation of the AGA chipset is possible but not 68040 or 68060 CPUs. I am ok with simulation where original timing can be maintained. You would probably like to have ASICs of the original chips but it is not practical as the original chip processes are no longer available and making multiple ASICs is much more expensive than making one 68k SoC ASIC. Even then, FPGA logic may be more practical for chipsets.

bhabbott Quote:

Such high standards Amiga fans have today. 'only' 64MB, 'downgrade from 040', and 'no 3D upgrade'. I feel so inadequate having 'only' 32MB and a 50MHz 030 in my A1200.

You are not using an original Amiga 1000 68000+OCS with 256kiB of memory? Are you a hardcore Amiga fan or not?

bhabbott Quote:

Not cheap, and also not an Amiga. It's not even a clone, just some addon boards for a DE10 nano.

If the "Amiga" cases are licensed from the Amiga IP owner, then it is at least partially an Amiga. Maybe the FPGA logic and timing are not 100% as they are recreations instead of using the original schematics but you would likely complain even if the logic came from the schematics or logic definitions used to create the original chips. The 68k CPU, Amiga chipset and AmigaOS are highly compatible which is what matters to most Amiga fans.

bhabbott Quote:

RPis were infamous for drawing more power than expected compared to other boards people were used to, and famously didn't provide a specification for power draw.

I used the "power virus" workload power as it should be close to worst case. The complexity of modern SoCs with many power gated units makes peak or worst case power not as useful anymore. It used to be that the power draw of CPUs executing various workloads/benchmarks could be measured with the peak power being the highest power reading. Today, SoCs and even CPU cores rarely have all units powered up at once during normal operation. A virus or malicious program could look for exploits to power on as much as possible in hopes of a meltdown. Many advanced chips have temp sensors and throttle performance but that may be hacked too. It becomes difficult to accurately give a power rating. TDP is not well defined and peak power is difficult to achieve and measure. There is usually some estimation necessary to come up with a power rating and recommendation for a power supply power rating. Perhaps some marketing advantage plays into this. TDP was introduced by Intel when peak power was widely used and the smaller TDP number made high power x86 chips look lower power.

bhabbott Quote:

Intel have always sucked at making low power devices. But nobody cares about that. What matters is that it smokes (perhaps literally) any real Amiga. 100, 1000, 10,000x faster? At some point you must be satisfied, right?

I would like an inexpensive low end 68k SoC using in-order cores that could be clocked between 1-2GHz. Right now the 68k is getting smoked by sub $1 ARM SoCs that use more transistors than a 68060&AA+ Amiga. Am I too greedy?

bhabbott Quote:

Yet another product you won't be able to get. That's why I have resolved to only make stuff with standard TTL gates from now on.

It is not too difficult to make boards for SoCs. The PPC Mirari board is another example of a board for old SoCs. Starting with an old SoC design usually shortens the product life though.

bhabbott Quote:

How lucky we are that Commodore went down when it did, before bastardizing the Amiga.

Hombre/PA-RISC was separate from the 68k Amiga and may have ended up being shorter lived than the 68k Amiga. Integer fixed point 3D viability did not last long. Integer fixed point 3D only lasted from about 1995 to 2000 and if late in Commodore fashion, it may not have been successful. PA-RISC implements integer SIMD in the integer unit where adding floating point SIMD would reduce integer performance (the Apollo Core does the same as it is cheap). When a floating point capable SIMD unit was wanted, it may have been easier to return to the 68k and add it. PA-RISC, 88k and AC are the only ISAs I can think of that added SIMD into the integer unit and none of them added floating point SIMD even though it is more important than integer SIMD today.

@Hammer

Quote:

GeekBench compared Intel Processor N97 vs Broadcom BCM2712 Intel Processor N97 (quad-core Intel Gracemont CPU): Geekbench 6 single core: 1,031 Geekbench 6 multi-core: 3,068 Broadcom BCM2712 (quad-core ARM Cortex A76 CPU): Geekbench 6 single core: 748 Geekbench 6 multi-core: 1,507

Those RPi numbers look pretty old. The SoC in the RPi is run very conservatively, but over time they tweak the settings and it gets faster.

You can also tweak the setting yourself and get considerably better numbers - it is very overclockable. Some people have got it well over 3GHz with crazy cooling setups.

I modified the memory timing on mine and changed the PCIe setting for the SSD. These were my results from a few months ago:

This was at 2.4GHz with memory tweak:
Geekbench 6 / Linux:
Single core 890
Multi core 2108
- About the same as a 10 year old Mac mini (Core i5 2.6GHz)


Overclocked to 2.8GHz with memory tweak:
Geekbench 6 / Linux:
Single core 1018
Multi core 2283
- About the same as an 8 year old MacBook Pro (Core i7 2.6GHz)

_________________
Whyzzat?

@minator

Memory bandwidth remains largely as is.

N97 power limits can be increased, and it will self-overclock.
https://www.minipcunion.com/download/file.php?id=2467&sid=45681ec06b1da9954abec9cb8bb94a04
Change power limits from the BIOS.

https://forum.odroid.com/download/file.php?id=19185&sid=456e389c90fffa2f3434588dbdbd8e41
3.8 GHz example.

Haswell is old and not in production.

For modern X64 CPUs, keeping temps down (e.g., using Honeywell PTM7950 over MX-6 paste, undervoltage) allows the boost clock to overclock itself.

Using Passmark e.g. game physics
https://www.cpubenchmark.net/cpu.php?cpu=Broadcom+BCM2712&id=6672
Broadcom BCM2712 with 2.4 Ghz (Quad core ARM Cortex A76)
Physics: 142 fps
Extended Instructions: 2,492 Million Matrices/Sec


https://www.cpubenchmark.net/cpu.php?cpu=Intel+N97&id=5337
Intel N97
Physics: 416 fps
Extended Instructions: 3,098 Million Matrices/Sec

Last edited by Hammer on 29-May-2025 at 12:50 AM.
Last edited by Hammer on 29-May-2025 at 12:43 AM.
Last edited by Hammer on 29-May-2025 at 12:39 AM.
Last edited by Hammer on 28-May-2025 at 11:41 PM.
Last edited by Hammer on 28-May-2025 at 11:28 PM.

_________________

@matthey

Quote:

ODROID-H4 SBC $99 8GiB memory $40 active cooler $10 60W power supply $20? --- $169 for 8GiB system ($149 for 4GiB system, $139 for 2GiB system) RPi 5 SBC $40 (base price but no zero memory option) 8GiB memory $40 ($5/GiB) active cooler $10 (passive cooler $5) 45W USB-C power supply $17 --- $107 for 8GiB system ($87 for 4GiB system, $77 for 2GiB system)

That's flawed. For NVMe, you haven't included M.2 HAT for RPi 5, while it's included with ODROID-H4.

ODROID-H4 has 2.5 Gigabit Ethernet while RPi 5 has 1 Gigabit Ethernet.

From https://pcpartpicker.com/products/memory/#ff=ddr5_sodimm&sort=price&page=1
Crucial CT8G48C40S5 (DDR5-4800 SODIMM) 8 GB has $20.99 retail.

RPi 5 is a dead end with RAM expansion.

https://www.hardkernel.com/shop/odroid-h4/
32 GB and 48 GB DDR5 SO-DIMMs 4800 or 5600 MT/s are validated.

From https://www.hardkernel.com/shop/odroid-h4/
H4's PSU is $9.40

Memory bandwidth is important for iGPU performance in games and other 3D workloads. You keep avoiding Amiga's original use case: games! Intel N97 has a larger iGPU.

ODROID-H4 SBC = $99
8GB DDR5-4800 SODIMM = $20
ODROID H4 PSU (15Vx4A) = $9.40
--------------------------------------------
$128.40

For an external GPU i.e. AMD Radeon cards:
RPI 5 has PCIe 2.0 x1 interface enabled and can enable a PCIe 3.0 x1 interface.
H4 has a PCIe 3.0 x4 interface via NVMe. 4 lanes can be split up with an adapter.

You stacked flawed prices against H4, which is flawed.

The only reason why RPi is used for C= Amiga is PiStorm, and that's only up to PRi 4B or CM4. Both H4 and RPi 5 would run full UAE.

Last edited by Hammer on 29-May-2025 at 01:58 AM.
Last edited by Hammer on 29-May-2025 at 01:42 AM.

_________________

@matthey

Quote:

If I/O is limited, the N97 SoC may be able to get away with as cheap of power supply as the RPi 5. The RPi 5 can get away with a cheaper power supply by limiting the I/O too. The ODROID-H4 does not have PCIe slots although it may be possible to add them. The 75W limit of PCIe from the PCIe slot is high for a low power SBC and low for a x86-64 GPU. Actually, it is possible to have capable 1080P 3D from the PCIe 75W slot in a low profile package.

Both RPi 5 and ODROID-H4 would need an external GPU solution that includes a separate PSU.

Both may not be cost-effective at this point when compared to budget LGA1700 or AM4/AM5 motherboards.

ODROID-H4 has PCIe 3.0 4 lanes.
RPi 5 has up to PCIe 3.0 1 lane.

Quote:

The NVidia RTX 3050 low profile is perhaps the best performance from a GPU powered from the 75W PCIe slot but they usually sell for roughly $200 USD, more than a low power SBC with integrated GPU.

Not the best GPU family for Linux. Why limit yourself to PEG's 75 watts limit when the solution is the external GPU?

Quote:

The SBC to support them including power supply is usually several hundred USD as well. The PPC Mirari board has better I/O options than the ODROID-H4 SBC in my opinion with 2x NVMe without hats and 2x SATA while the last generation of PPC 28nm T series SoCs for the boards target a 30W envelope (likely peak power so maybe 20W TDP) which is less than twice that of the N97 SoC.

PPC Mirari is not a US$99 price range SBC, and PPC e5500 lacks 128-bit SIMD hardware. PPC e6500 has Altivec 2.0 SIMD.

Power consumption is not a major concern for desktops. Power is a major concern for mobility.

There are other solutions beyond US$99.

N97 solution can run on 37-watt PSU, as the laptop example shows. Unlike the laptop example, ODROID H4's power limit setting is exposed in the BIOS i.e. end users can override 12 watts TDP into 30 watts TDP, hence higher auto-overclocking boost.

Quote:

The Mirari SBC could possibly be mass produced for under $500 USD but add the PPC SOC, a GPU, memory, power supply, active cooler and case and this system will likely cost closer to $1,000 USD. Reduce the CPU power (process improvement in the case of N97), integrate the GPU, eliminate the PCIe slots and reduce the SBC board size and the sub $100 ODROID-H4 SBC is possible with mass production.

PPC Mirari does not target the laptop use case.

I would like to see other CPU ISA options for the Framework 13 mainboard form factor beyond SiFive U74/P550, Intel 11th/12th/13th/Meteor Lake, and AMD Phoenix Point/Strix Point. These are the options for the Framework 13 mainboard form factor.

The price for T1042 can be rapidly reduced by 100 unit batch orders.

https://pcpartpicker.com/products/power-supply/#sort=price&page=1
National PC PSU price list for the USA.

https://pcpartpicker.com/products/case/#sort=price&page=1
National PC case price list for the USA.

https://pcpartpicker.com/products/memory/#ff=ddr3_sodimm&sort=price&page=1
National SODIMM DDR3 price list for the USA.



Quote:

The only problem is that newer chip fab processes are expensive but an older fab process and lower power in-order CPU cores can be passively cooled for cheaper as the older Atom SBCs with PowerVR GPUs demonstrated.

From the factory configuration, ODROID-H4 is passively cooled. Power limit settings are in the BIOS.

ODROID-H4's bundled passive-cooled heatsink just needs a cheap fan for active cooling.

Quote:

The x86-64 SoCs are often scaled up to gain performance and more x86-64 software support but they are fat and expensive as they use a smaller fab process to reduce power. Scaling x86-64 cores down into ARM territory is possible as demonstrated by in-order Atom/Bonnell cores

It was re-demonstrated by Intel Lunar Lake against Qualcomm Elite X.

AMD needs to move towards a 3 nm process node TSMC like Intel Lunar Lake and Qualcomm Elite X. Mobility is purpose-designed.

AMD already has three Zen 5 uarch grades i.e., removing excess pipelines, compacting the layout at the expense of high clock speed, and removing excess L3 cache.

https://www.compactpc.com.tw/products/item/99
This fanless embedded computer is powered by an Intel Tiger Lake i5-1135G7 Processor


Quote:

CPU/core | pipeline | process | transistors 68060 8-stage 500nm 2,500,000 Pentium/P54C 5-stage 500nm 3,300,000 Atom/Silverthorne 16-19-stage 45nm 47,212,207 Atom/Moorsetown 16-19-stage 45nm 140,000,000 (Silverthorne with integrated GPU) N97 14-19-stage 7nm 1,000,000,000+ (integrated x86-64 OoO CPU with 4 cores and GPU)

That's flawed when 68060's front end is very weak with higher latency DDRx memory types, and the FPU is not pipelined and does not have multiple FPU pipelines.

Warp1260 developers added an external 64KB L2 cache as their attempt to mitigate against DDR3's higher latency.

N97's CPU core has multiple multipipelined FPU and SIMD units.
Intel Gracemont has six x86-64 decoders + microcode bottlenecked by 5 macro-ops. The instruction decoders are fed by 32-byte per cycle links.

Gracemont doesn't have a separate decode/micro-op cache and decoded instruction dispatch.

Gracemont integer:
Port 0: ALU,
Port 1: ALU, INT MUL, INT DIV LEA,
Port 2: ALU, INT MUL, INT DIV, PDEP,
Port 3: ALU,
Port 4: Branch,
Port 5: Branch,

Gracemont vector/FPU:
Port 6: FStore.
Port 7: FStore.
Port 8: 128bit ALU, 128bit FMA, 128bit, FADD, 128bit, FMUL, AES, FDIV, SHA.
Port 9: 128bit ALU, 128bit FMA, 128bit FADD, 128bit FMUL, AES.
Port 10: 128bit ALU.

256bit AVX2 SIMD is spread across multiple 128bit SIMD units like Zen 1.x.

Gracemont Address Generation
Port 11: Load AGU,
Port 12: Load AGU,
Port 13: Store AGU,
Port 14: Store AGU,
Port 15: Store Data,
Port 16: Store Data,

These are 128-bit wide i.e. double-pump with 256-bit AVX2.

Gracemont's design approach is flat and wide.

These are improved upon with Skymont, which is included with Arrow Lake and Lunar Lake.

68060 has weaker MMU performance for modern OS like recent Linux. Hint: TLB cache size.

For 2D or 3D games, 68060 @ 1.5 GHz will be crushed by modern CPUs with dual FP64/FMA3 pipelines, e.g. ARM Cortex A53 supports 128-bit NEON SIMD when it gangs dual FP64 pipelines.

68060 wasn't designed to drive OpenGL's FP32 efficiently. SiFive U74 has one FPU pipeline attached to the integer B pipeline. SiFive U7's design is primitive from the 1990s past.

ARM Cortex A53's instruction control feeds the following ports,
Port 0: ALU,
Port 1: ALU, INT MUL, Branch
Port 2: AGU,
Port 3: FADD, FMUL, FMA3, 1st half NEON.
Port 4: FADD, FMUL, FMA3, 2nd half NEON.

Cortex A53 is biased towards FPU and NEON SIMD i.e. multimedia SIMD!. LOL.


SiFive P550's ROB (reorder buffer) feeds the following ports
Port 0: ALU, Branch,
Port 1: ALU,
Port 2: ALU, INT MUL,
Port 3: Load AGU,
Port 4: Store AGU,
Port 5: FMA3,
Port 6: FMA3,
P550 has 3 decoders with 12 bytes per cycle from the L1 instruction cache and up to 13-stage pipeline. P550 doesn't support vector instructions i.e. no "multimedia" stream instructions.


ARM Cortex A75
Port 0: ALU, INT MUL,
Port 1: ALU, INT DIV,
Port 2: Branch
Port 3: AGU, LD/ST,
Port 4: AGU, LD/ST,
Port 5: FMA, 128-bit ALU, AES, NEON.
Port 6: FMA, 128-bit ALU, NEON.
Port 7: NE / FP Store

68K needs to be re-implemented with modern uarch's flat and wide design. Even AC68080 is not good enough. It needs to be 68K assimulated P550.

Intel went crazy with Gracemont's flat and wide approach, and it's wider with Skymont.


Quote:

The x86-64 SoCs are often scaled up to gain performance and more x86-64 software support but they are fat and expensive as they use a smaller fab process to reduce power. Scaling x86-64 cores down into ARM territory is possible as demonstrated by in-order Atom/Bonnell cores

Wrong. It was re-demonstrated by Intel Lunar Lake against Qualcomm Elite X.

Quote:

but x86-64 software compatibility is lost and the cores are larger and use more power than ARM cores despite being more powerful and having better power efficiency than ARM in-order cores when they were released. Intel Bonnell in-order cores are weak performance compared to Intel low power OoO cores like the N97 uses but the cost advantage when reaching passive cooling is large and the expense between the Bonnell 45nm/32nm and 7nm/5nm necessary to reduce the OoO core power to that of the in-order cores is expensive.

Wrong. It was re-demonstrated by Intel Lunar Lake against Qualcomm Elite X.

Atm, Gracemont is Intel's low-end budget uArch atm.

Quote:

The CPU+chipset clock speed would have improved from 14MHz to 28MHz and 57MHz with double to quadruple the performance including the memory bandwidth.

That's speculative. The goal for A1200 SoC is backward compatibility with A1200/CD32.

Quote:

The 68020&AGA@14MHz were slow because of the low clock speed and memory bandwidth. The perception of AGA using SRAM on FPGA devices is not the same as Commodore failing to advance the Amiga chipset adequately.

For a pure 32-bit Chip RAM configuration, the A1200's bottleneck is with Budgie since it implemented the A3000's interface I/O between the 32-bit CPU and the 32-bit Chip RAM.

The 68EC020 CPU couldn't access the A1200's 32-bit Chip RAM like AA Lisa due to Budgie's A3000's CPU 32-bit Chip RAM I/O implementation.

A1200 *is* AA3000plus's guts with some cost reduction e.g. 24-bit address line AA Gayle instead of 32-bit address line Fat Gary.

Budgie replaced Bridgette and combined some Buster and Ramsey memory controller features. Akiko encapsulates most of the functions from Budgie, Galyle, and two CIAs. Paula, Alice, and Lisa are the remaining chips.

Changing memory timings can break backwards compatibility.

Access Ltd's Budgie replacement allows the CPU to have 2X access rate with 32-bit Chip RAM like Lisa. Unknown game compatibility result.

Quote:

The 68k was still #1 in embedded 32-bit sales when Motorola bought the pipelined Japanese 68000 design, even though they were investing profits in PPC development.

Prove it for 68060. You need a mass use case sales driver for 68060.

PPC was engaged in a clock speed race against X86 and DEC Alpha with the workstation/desktop use case sales driver.

ARM's high clock speed gain was with DEC's StrongARM (1 IPC) and started PDA handheld compute power arms race. With ARM9T's +100Mhz, the early loser was a 33 MHz to 60 MHz 68000-based Dragon Ball VZ.

3rd party's 1 IPC 68000 re-implementation effectively competes against Motorola's 68040.

The main point with a 3rd party 1 IPC 68000 re-implementation clone is to get away from Motorola's pricing policies.

Quote:

There is no 68k market to protect today if that was the reason then. Licensing 68k cores today should be relatively cheap as production of 68k chips does not compete with Motorola 68k chip production. The situation difference is only about 80 million 68k chips per year from then to now.

Prove cheap with 68060 when the 68040 cloner's asking price is expensive.

Last edited by Hammer on 29-May-2025 at 06:37 AM.
Last edited by Hammer on 29-May-2025 at 06:18 AM.
Last edited by Hammer on 29-May-2025 at 06:15 AM.
Last edited by Hammer on 29-May-2025 at 06:04 AM.
Last edited by Hammer on 29-May-2025 at 05:27 AM.

_________________

@matthey

Quote:

SiFive is paid a license but it looked like there were no royalties when I looked into it. I did not do any licensing inquiries lacking a business but it looked like the licenses were reasonable too.

From https://www.sifive.com/about/business-model

SiFive provides intellectual property (IP) rather than making semiconductor chips. We license our RISC-V-based instruction set architecture (ISA) and processor technology

(skip)
While RISC-V is an open standard, our designs are proprietary and help you develop chips optimized for your use cases. Revenue comes from upfront licensing fees and royalties based on the chip's selling price.

(skip)

Beyond the CPU
Beyond CPUs, we develop components like memory management units (MMUs), debug and interrupt handling systems, and JTAG support. We also focus on integrating hardware IP with software platforms to accelerate the development of both chips and application software.

_________________

Hammer Quote:

That's flawed. For NVMe, you haven't included M.2 HAT for RPi 5, while it's included with ODROID-H4.

The ODROID-H4 makes more sense than a RPi 5 if using NVMe drives but the drive itself pushes up the price. The RPi 5 can still makes sense at the low end using microSD/USB drives, 2GiB or 4GiB of memory, using less power and at a lower price. The ODROID-H4+ has a large performance advantage but makes less sense with less than 8GiB of memory and a NVMe drive to access the x86-64 software library.

Hammer Quote:

ODROID-H4 has 2.5 Gigabit Ethernet while RPi 5 has 1 Gigabit Ethernet.

Not a big deal for low end hardware except for specific uses like high resolution streaming.

Hammer Quote:

From https://pcpartpicker.com/products/memory/#ff=ddr5_sodimm&sort=price&page=1 Crucial CT8G48C40S5 (DDR5-4800 SODIMM) 8 GB has $20.99 retail. RPi 5 is a dead end with RAM expansion. https://www.hardkernel.com/shop/odroid-h4/ 32 GB and 48 GB DDR5 SO-DIMMs 4800 or 5600 MT/s are validated.

There is a 16GiB RPi 5 option but it does not make much sense. Even the 8GiB RPi 5 option is borderline practical compared to x86-64 SBC options. The 2GiB RPi 5 is limited with the code density of AArch64 too but memory saving AArch32 options are still available on the RPi 5 and can be used for the RPi OS where 32-bit remains popular for the memory saving ability. Ironically, ARM is deprecating ARM 32-bit ISAs when with 2GiB of memory and a slower microSD swap, a 32-bit OS makes more sense. Without it, the 4GiB RPi 5 may be the only practical option and it does not make use of the 64-bit address space which is why 64-bit was added in the first place.

https://www.reddit.com/r/raspberry_pi/comments/smv9h3/raspberry_pi_os_32bit_vs_64bit_performance_review/
tinkerterrapin Quote:

In one of my projects 64 bit mode runs about 35% faster and takes 45% more memory. Tested on the same Zero 2 W with fresh Raspberry Pi OS 32/64bit installs. 64bit - Linux raspberrypi 5.10.92-v8+ #1514 SMP PREEMPT Mon Jan 17 17:39:38 GMT 2022 aarch64 GNU/Linux Memory Used: 465088 Time taken: 240.090733ms Time taken: 239.601585ms Time taken: 241.56121ms Time taken: 238.481583ms Time taken: 239.770699ms 32bit - Linux raspberrypi 5.10.92-v7+ #1514 SMP Mon Jan 17 17:36:39 GMT 2022 armv7l GNU/Linux Memory Used: 320888 Time taken: 373.907552ms Time taken: 387.10974ms Time taken: 374.230208ms Time taken: 374.560729ms Time taken: 374.627604ms

The cost of ARM64/AArch64 is 45% more memory is needed! With ARM 32-bit ISAs disappearing from application processors, that means 256MiB, 512MiB and 1GiB SBCs are no longer practical with 64-bit Linux. The RPi 4 already comes with 2GiB minimum leaving the 64-bit RPi 3 with 1GiB and the RPi Zero 2 W with 512MiB but most embedded customers likely use 32-bit support, dedicated embedded applications can get away with using less memory than a general purpose OS and RTOS options exist that use less memory than Linux. The deprecated ARM 32-bit support is forcing new RPi SBCs to scale up and compete with 64-bit x86-64 SBCs which have better performance and a large x86-64 software library.

SBC | memory | storage
RPi5 2GiB microSD (cheap option but 32-bit OS and apps may be better with slow drive swap)
RPi5 2GiB NVMe (not much sense as adding more memory is cheaper than adding a NVMe drive)
RPi5 4GiB microSD (faster 64-bit OS and apps with reduced swap but 64-bit addressing not used)
RPi5 4GiB NVMe (faster but added NVMe drive cost now competes with x86-64 SBCs)
RPi5 8GiB microSD (borderline worthwhile competing against x86-64 SBCs)
RPi5 8GiB NVMe (better off with x86-64 SBCs due to price & spec opening up x86-64 software library)
RPi5 16GiB microSD (does not compete against x86-64 SBCs)
RPi5 16GiB NVMe (does not compete against x86-64 SBCs)

RPi can continue to use older cores and chip fab processes but that may become difficult. The other option is to scale up 32-bit Cortex-M cores for embedded application like uses which may be what they are doing with their fabless semiconductor development. ARM has created an opportunity for a 32-bit CPU ISA with good code density to fill the void they have created by removing 32-bit compatibility to better compete against 64-bit x86-64. ARM likely assume that no 32-bit architecture with Thumb-2 like code density could have good performance and that no new 64-bit ISA competition would have better code density and performance than AArch64 but I believe neither of these assumptions are true.

Hammer Quote:

From https://www.hardkernel.com/shop/odroid-h4/ H4's PSU is $9.40

There are cheaper prices for RPi accessories too. RPi switched from making their profit on the SBCs to making their profit on accessories like cables, coolers, etc. The change was from dirt cheap ARM SoCs that likely cost less than $1 to expensive ARM SoCs in the RPi 5 that are the majority of the cost of the SBC leaving minimal margin for the SBC. The ODROID-H4 x86-64 SoC is also most of the SBC cost. There was little ARM competition for low end "standard" embedded/hobby hardware with a divided ARM market, much like the current 68k Amiga market, but x86-64 hardware is more competitive down to where it scales with so much baggage.

Hammer Quote:

Not the best GPU family for Linux. Why limit yourself to PEG's 75 watts limit when the solution is the external GPU?

There are good reasons to be limited to the 75W PCIe max. An additional GPU power connector would require a larger power supply. Low profile small form factor and old business hardware can be used which often do not have additional power connectors. The performance is more than adequate for 1080P gaming and likely outperforms most x86-64 SBCs with integrated GPUs. The GPU performance/$ is not as good as higher end GPUs but the higher price offsets cheap x86-64 used hardware costs. The price would be lower if AMD would bring out a competitive card where there Radeon RX 6400 is not competitive in performance but is much cheaper. Integrated GPUs have a cost and performance advantage but x86-64 SoCs usually reduce the price and power instead of pushing performance but the x86-64 SoCs and integrated GPUs are getting better and usable for gaming. Consoles use them even though the most powerful GPUs continue to be discreet GPUs.

Hammer Quote:

PPC Mirari is not a US$99 price range SBC, and PPC e5500 lacks 128-bit SIMD hardware. PPC e6500 has Altivec 2.0 SIMD. Power consumption is not a major concern for desktops. Power is a major concern for mobility. There are other solutions beyond US$99. N97 solution can run on 37-watt PSU, as the laptop example shows. Unlike the laptop example, ODROID H4's power limit setting is exposed in the BIOS i.e. end users can override 12 watts TDP into 30 watts TDP, hence higher auto-overclocking boost.

The base hardware spec is not so different between a sub $100 SBC system and a several hundred dollar desktop system. The desktop system has more I/O and power for the I/O but the additional cost can be several times the cost of the sub $100 SBC system. The sub $100 SBC system is adequate for most users as the RPi 4, RPi 5 and ODROID-H4 are more powerful than the PPC Mirari board for a fraction of the cost. The Mirari board can upgrade the GPU further but it is a very expensive option that is limited by the weaker PPC CPU performance.

Hammer Quote:

From https://www.sifive.com/about/business-model SiFive provides intellectual property (IP) rather than making semiconductor chips. We license our RISC-V-based instruction set architecture (ISA) and processor technology (skip) While RISC-V is an open standard, our designs are proprietary and help you develop chips optimized for your use cases. Revenue comes from upfront licensing fees and royalties based on the chip's selling price. (skip) Beyond the CPU Beyond CPUs, we develop components like memory management units (MMUs), debug and interrupt handling systems, and JTAG support. We also focus on integrating hardware IP with software platforms to accelerate the development of both chips and application software.

I am aware of what the SiFive site says which is the same as it was back when I was investigating. There may be royalties on some of their products or pass through royalties but people posting on forums made it sound like there were no royalties for their development products like ARM charges, just a set license fee. It is not possible to avoid royalties altogether with a modern SoC. HDMI requires a small royalty and a good integrated GPU would likely require royalties. SerDes and some other standard IP blocks may require royalties. It may be possible to create a MCU without 3D GPU and DVI 15-pin HD out for the Amiga chipset with USB 2.0 and 100Mbit Ethernet. The V4SA avoids using SerDes even though the HDMI maybe should with a license and royalties paid. There are Cyclone V FPGAs with SerDes for not much more.

@matthey

Quote:

The ODROID-H4 makes more sense than a RPi 5 if using NVMe drives but the drive itself pushes up the price. The RPi 5 can still makes sense at the low end using microSD/USB drives, 2GiB or 4GiB of memory, using less power and at a lower price. The ODROID-H4+ has a large performance advantage but makes less sense with less than 8GiB of memory and a NVMe drive to access the x86-64 software library.

Linux ARM can access x86-64 software library (e.g. Valve Steam) via Box64 with JIT emulation overheads.

RPi 5 has Windows 11 ARM, which includes x86 JIT emulation.

Quote:

Not a big deal for low end hardware except for specific uses like high resolution streaming.

2.5 Gbit Ethernet is useful for a low-end NAS role.

Quote:

There is a 16GiB RPi 5 option but it does not make much sense. Even the 8GiB RPi 5 option is borderline practical compared to x86-64 SBC options. The 2GiB RPi 5 is limited with the code density of AArch64 too but memory saving AArch32 options are still available on the RPi 5 and can be used for the RPi OS where 32-bit remains popular for the memory saving ability.

16 GB RPi 5 makes sense when used as a desktop computer or a larger language AI model use cases.

RPi 5 has an official AI Kit incl M.2 HAT+, 26 TOPS, and Hallo AI acceleration module. Expect the RPi's next SoC selection with an integrated NPU.

Embedded AI sales are driving resurgent MIPS64 SoCs. MIPS64 also gained 16-bit compressed instructions and assimilated NPU. MIPS64 camp has modernized its SoC solutions.

For 68K, it would take a lot of work to follow MIPS64's modernization. Business development is a major factor.

"Embedded AI" in disposable missiles/drones is in great demand. War is good business when you're the supplier and not directly involved in a hot war.

Business development for rebooting the 68K family is a major factor that is outside of Retro Amiga's scope.

Quote:

There are good reasons to be limited to the 75W PCIe max. An additional GPU power connector would require a larger power supply.

PC PSUs with 400 to 500 watts are cheap.

Quote:

The performance is more than adequate for 1080P gaming and likely outperforms most x86-64 SBCs with integrated GPUs. The GPU performance/$ is not as good as higher end GPUs but the higher price offsets cheap x86-64 used hardware costs. The price would be lower if AMD would bring out a competitive card where there Radeon RX 6400 is not competitive in performance but is much cheape

AMD is currently busy with the game console B2B. Intel BattleMage has a better "bang per buck" since they lost Xbox Next and PS6 contracts. When AMD released RX 5700 XT, AMD was busy with Xbox Series X, Xbox Series S, and PS5 R&D. RX 9700 XT's release is the same pattern as RX 5700 XT's release.

https://www.youtube.com/watch?v=Tv0o6505JAc
Intel's B580 GPU, tested on Linux.

Quote:

Low profile small form factor and old business hardware can be used which often do not have additional power connectors. The performance is more than adequate for 1080P gaming and likely outperforms most x86-64 SBCs with integrated GPUs.

RTX 3050 8GB was rivaled by AMD Strix Point R9's and Intel Lunar Lake's Xe 56/64 IEU integrated GPUs. Reminder, modern PC laptops employ custom x86-64 SBCs with mobile x86-64 SoCs. PC laptop is the major form factor and use case for PC SBCs.

Nvidia has discontinued RTX 3050 8 GB's production. Nvidia told its partners to shift to the lower-cost 6 GB variant. NVIDIA effectively lost a performance segment.

AMD Strix Point R9's and Intel Lunar Lake's Xe 56/64 IEU integrated GPUs reserve video memory settings can exceed 8 GB for larger language AI models e.g. 8GB, 16GB, 32GB, and 64 GB. It's also useful for brain-dead games that look at video memory storage size with texture storage allocation.

Quote:

The base hardware spec is not so different between a sub $100 SBC system and a several hundred dollar desktop system. The desktop system has more I/O and power for the I/O but the additional cost can be several times the cost of the sub $100 SBC system.

US$99 is one of the "magical" price targets. It was used in the past marketing from Commodore, Sinclair Research, and many others.

Quote:

The sub $100 SBC system is adequate for most users as the RPi 4, RPi 5 and ODROID-H4 are more powerful than the PPC Mirari board for a fraction of the cost. The Mirari board can upgrade the GPU further but it is a very expensive option that is limited by the weaker PPC CPU performance.

PPC Mirari has a 64-bit PPC with standard PPC FPU i.e. there are cult followers for the PPC camp. I'm okay with Mirari's PPC approach since it's better than the "expert advice" that gave the custom FPU with PPC e500 design for desktops.

Last edited by Hammer on 30-May-2025 at 05:10 AM.
Last edited by Hammer on 30-May-2025 at 05:06 AM.
Last edited by Hammer on 30-May-2025 at 05:03 AM.
Last edited by Hammer on 30-May-2025 at 04:49 AM.
Last edited by Hammer on 30-May-2025 at 04:46 AM.
Last edited by Hammer on 30-May-2025 at 04:31 AM.

_________________

@Hammer

Quote:

RPi 5 is a dead end with RAM expansion.

I've been using Laptops with fixed RAM for probably around a decade now, never had a problem.

My RPi 5 could probably have used 4GB but I went for 8GB just to avoid using swap.

Quote:

The only reason why RPi is used for C= Amiga is PiStorm, and that's only up to PRi 4B or CM4. Both H4 and RPi 5 would run full UAE.

Goes like a rocket running Amiberry.

Quote:

Power consumption is not a major concern for desktops.

If you're using a desktop for music production or if you live in a hot environment, yes, power is absolutely a concern.

Last year I spend a *lot* of money on a Mac laptop to run local Ai models.

An equivalent PC - if it could even be built - would be in the region of 1.5 KiloWatts.
Given how hot it can get here in the summer, and the sheer amount of noise it'll make, it was simply not an option.

Quote:

It was re-demonstrated by Intel Lunar Lake against Qualcomm Elite X.

Lunar lake is on 3nm, X-Elite is a first implementation on 4nm.
Lunar lake is close to the X-Elite but doesn't even touch the M4.

Quote:

AMD needs to move towards a 3 nm process node TSMC like Intel Lunar Lake and Qualcomm Elite X.

The rumour is AMD will jump to 2nm in 2026 and 1.4nm straight afterwards. They're kicking Intel while they're down.

They are also supposedly doing an Arm laptop processor called Sound Wave.

_________________
Whyzzat?

@minator

Quote:

I've been using Laptops with fixed RAM for probably around a decade now, never had a problem.

Soldered RAM is problematic enough, which led to the DELL-driven CAMM standard.

JEDEC CAMM (Compression Attached Memory Module) is a new memory module form factor for laptops, initially developed by Dell and now a JEDEC standard.

Quote:

Goes like a rocket running Amiberry.

Reminder, WinUAE supports PPC.

Quote:

If you're using a desktop for music production

Pure music production doesn't need high-performance gaming GpGPUs, but I did use two RTX 3080 Ti GPUs for AI-related bulk sound work conversion and tested the said software on RTX 4080 and RTX 4090.

I later sold the two RTX 3080 Ti GPUs in early 2025.

Quote:

or if you live in a hot environment, yes, power is absolutely a concern.

I don't recall this being a concern for single external power cable RTX 4060 cards in countries like the Philippines or the Australian state of Queensland.

Quote:

Last year I spend a *lot* of money on a Mac laptop to run local Ai models.

I see, an Apple.

Quote:

An equivalent PC - if it could even be built - would be in the region of 1.5 KiloWatts.

Bullshit, that's false.

Apple's AI performance is a comical joke compared to RTX 4070, let alone RTX 4090.

Geekbench AI
https://browser.geekbench.com/v6/compute/search?utf8=%E2%9C%93&q=M4+max
Mac Pro M4 Max's score ranges, and that's with vendor-specific Metal API.

https://browser.geekbench.com/v6/compute/search?utf8=%E2%9C%93&q=RTX+4080
Desktop RTX 4080 easily beats Mac Pro M4 Max with vendor-neutral OpenCL API.

Running AI workload doesn't use RTX's raster and RT hardware. You're bullshiting.

RTX 4080 and RTX 4090 crushed Apple's best on raytracing results
https://opendata.blender.org/benchmarks/query/?compute_type=OPTIX&compute_type=CUDA&compute_type=HIP&compute_type=METAL&compute_type=ONEAPI&group_by=device_name&blender_version=4.4.0


My gaming PC-01 with Ryzen 9 7950X + RTX 4080 OC (350 watts, Gigabyte brand) has an 850 watts PSU ATX 3.0. NVMe 4TB at PCIe 4.0.
https://www.techpowerup.com/review/gigabyte-geforce-rtx-4080-super-gaming-oc/41.html
Power Consumption for Gigabyte GeForce RTX 4080 Super Gaming OC i.e. typical gaming 312 watts, 20 ms spike 393 watts. Gigabyte recommends a 700-watt PSU.


My gaming PC-02 with Ryzen 9 9950X3D + RTX 4090 OC (600 watts, ASUS brand) has a 1000-watt PSU ATX 3.0. NVMe 4TB at PCIe 5.0.
https://www.techpowerup.com/review/asus-geforce-rtx-4090-strix-oc/39.html
Power Consumption for ASUS GeForce RTX 4090 STRIX OC i.e. typical gaming 360 watts, 20 ms spike 547 watts. ASUS recommends a 900-watt PSU.


Your "would be in the region of 1.5 KiloWatts" is a load of bullshit. If your argument is true, my gaming PC-02 wouldn't be working.

I also spent money on 7 KiloWatts of solar panels.

https://www.reddit.com/r/macbookpro/comments/1hj3m0p/m4_max_is_draining_167w_powersystem_total/
"M4 Max is draining 167w power". This is the walled power DTR level.

On Blender 4.4, Apple M4 Max (GPU - 40 cores) is equivalent to RTX 4070.

Last edited by Hammer on 31-May-2025 at 09:00 AM.
Last edited by Hammer on 31-May-2025 at 08:44 AM.
Last edited by Hammer on 31-May-2025 at 08:42 AM.
Last edited by Hammer on 31-May-2025 at 07:53 AM.
Last edited by Hammer on 31-May-2025 at 07:47 AM.

_________________

@Hammer

Quote:

Pure music production doesn't need high-performance gaming GpGPUs,

It doesn't need GpGPUs at all, it just needs to be quiet.

Quote:

Bullshit, that's false.

False for a gaming PC.
I didn't want a gaming PC. I wanted something that could run large AI models. A single 4090 has nowhere near enough memory, I'd need 3 or 4 of them to load and run a 70b Q8 model.

That's a big, hot, and expensive system, if it's even possible, I don't know if it's possible to fit more then 2 4090s in one machine.
You can that level of memory with the Nvidia professional AI cards but Nvidia's prices make Apple look like a budget brand.

I ended up getting a Mac laptop with 128GB. It runs the AI plenty fast enough for my needs and I can run Logic with a ton of plug-ins.
I keep it in low power mode, so the fans rarely ever switch on.

This was a year ago. Today AMD does an APU that can load those size of models for a lot less money.

Last edited by minator on 31-May-2025 at 12:40 PM.

_________________
Whyzzat?

@Hammer

Quote:

MiSS-1200 is not a full A1200 clone due to missing C= A1200 expansion capability. Need to combine the efforts from MiSS-1200 and ACA500's A1200 expansion port capability.

Neither do any of the other alternatives mentioned, like A1200NG, so why do you bring that up? If you need that, then there’s really just one option, a Re:A1200 (of which there are a few variants, but all use original chips.

@Matthey
Minimig/MiSTer is not limited to 64MB of RAM, it has hundreds (256, 384) of MB (DDR2) available. MiST has max 64MB these day… Sidi128 had 128MB. And with non-DE10 MiSTer clones appearing, more RAM and features may pop up too.

_________________
B5D6A1D019D5D45BCC56F4782AC220D8B3E2A6CC

@minator
Quote:

It doesn't need GpGPUs at all, it just needs to be quiet.

Any noise issue can be mitigated. I have my two-floor story house, I'm not limited by living space issues.

Virtue signaling climate change credentials via low power consumption while promoting readily all-in-one e-waste designs is LOL i.e. at the end of their use life, the entire machine is ditched e.g. Mac Pro with M4 Max uses SSD storage modules that are soldered to the logic board. SSD is a consumable like batteries.

Quote:

False for a gaming PC. I didn't want a gaming PC. I wanted something that could run large AI models. A single 4090 has nowhere near enough memory, I'd need 3 or 4 of them to load and run a 70b Q8 model.v

Reminder, NVIDIA RTX PRO 6000 Blackwell has 96GB VRAM and RTX 6000 Ada Generation has 48GB VRAM.

I haven't changed my C= Amiga era game-centric focus i.e. gaming 1st, other workload comes later.

Instructure ML solutions have a larger APU, such as AMD's MI300A (228 CU CDNA 3 GPU + 24 Zen 4 cores) or larger HBM VRAM-equipped NVIDIA server CUDA GPUs. AMD's mistake is a separation between server and gaming GPU uarch, and to be solved in the next generation UDNA uarch.

These are larger than Apple M4 Max (iGPU with 40 cores) and AMD's Strix Halo (iGPU with 40 CU). AMD's Strix Halo APU has the same frozen RAM configuration as Apple's M4 Max.

For B2B, AMD already delivered PS5 Pro APU with 64 CU (RDNA 3 + RDNA 4 RT) iGPU and fast unified memory.

Like many others, if I develop ML software, it most likely targets NVIDIA's infrastructure GPUs, and I would use my NVIDIA gaming GPUs to develop the software.

NVIDIA RTX PRO 6000 Blackwell can play games at the flagship GB202 GPU level i.e. RTX 5090.

https://www.tomshardware.com/pc-components/gpus/rtx-pro-6000-crowned-the-new-gaming-king-but-its-usd10-000-price-tag-makes-the-all-gold-dhabab-rtx-5090-seem-cheap
RTX Pro 6000 crowned the ‘new gaming king’ i.e. faster than RTX 5090.

The 64GB of GDDR7 costs the manufacturer about $200. The lack of competition in high-end PC gaming led to NVIDIA's price gouging.


Llama-3.3-70B-Instruct-Q4_0-GGUF is 39.97 GB in size.

Last edited by Hammer on 09-Jun-2025 at 07:54 AM.
Last edited by Hammer on 31-May-2025 at 03:13 PM.
Last edited by Hammer on 31-May-2025 at 03:10 PM.
Last edited by Hammer on 31-May-2025 at 03:09 PM.
Last edited by Hammer on 31-May-2025 at 03:07 PM.

_________________

@Hammer

Quote:

Hammer wrote: @cdimauro Quote: Whatever: that's not the/a problem. The same "Virtual Machine Platform" dependency needs to be enabled. You're making a big deal about the Windows Subsystem for Linux and Google Play Games name difference. For normal end users, the end goal with Windows Subsystem for Linux is Windows Subsystem for Android, which MS has ended support for. The primary driver is running 3rd-party apps and games. Google Play Games for Windows is the official method from Google Inc. Quote: Irrelevant It's relevant since it has the Linux foundation with a different middleware focus, i.e., targeting Windows DirectX closed-source games. Android also has Linux foundations with Google's middleware. Linux port + Box64 port + Radeon driver port enables little-endian hardware platforms to run Steam games. This is relevant for this topic's source link's fallback feature. Quote: Again, irrelevant: there are plenty of retro projects which used the same SoC of this "new" X68000, which are very very cheap. Which was the point of this part of the discussion. For *this* topic, this project is not purely "retro" when this Japanese project includes a mainstream OS platform fallback feature, which is clearly stated in the source website. Your excuse is that you can't read Japanese. This project has a retro X68000 theme with modern mini-PC system integration. Mini-PC form factor is a major virtue in the Japanese market. Most gamers don't care about the hardware guts, they only care about the delivered gaming experience within their budget. Only the extremists care about the hardware guts. Unlike Zuiki's projects, the specific ASIC C64 joystick is effectively a dead-end eWaste toy since it has no fallback use case to run a mainstream desktop OS. ---------------------------------- For Zuiki's initial 2022 project: https://retrolike.net/2023/10/29/the-zuiki-x68000z-the-minimized-powerhouse/ The cost for X68000Z tiers and packs On June 8, 2023, Zuiki opened pre-orders for the consumer-level X68000Z through various platforms. Including Amazon Japan, BEEP, and select department stores and retailers. They offer three distinct packages: Starter Pack (¥29,935 / $200 US): This basic package includes the computer, a joypad, and copies of Gradius and Human 68K Version 1.01. Despite its simplicity, all the computers in this tier are identical, regardless of which package you choose. Basic Pack (¥65,780 / $439): The Basic Pack comprises everything from the Starter Pack, adding a keyboard and mouse specifically designed for this computer. The keyboard is a faithful replica of the original down to the buttons and weight. This package provides a more comprehensive experience but comes at a higher price. Complete Pack (¥87,780 / $586): For those seeking the ultimate X68000Z experience. The Complete Pack includes everything mentioned in the other sets. Most interesting addition is is the five-inch monitor replicating the 1987 original 14″ version. This replica monitor maintains the 4:3 aspect ratio of the original CRT screen but features an HDMI connection and a resolution of 1024×768 pixels. It can also be used with other hardware that supports this resolution. Additionally, this package comes with a UART cable, a serial I/O interface for connecting original accessories. Bundled X68000 games have proper licenses. Production scale can be estimated from revenue generation vs unit price. US$24,131,274.13 / $439 = about 54,968 unit production scale. US$24,131,274.13 / $200 = about 120,656 unit production scale. Zuiki developed a semi-custom SoC with ARM-based Z7213 SoC (System-on-a-Chip), which is based on Allwinner R16 + Mali GP-400 SoC. For 2022, the Japanese market can tolerate Zuiki's initial offer i.e. "what the market will bear" basic economics 101. It's similar to Mac 68K's audience, which they can tolerate Apple's higher prices, being driven by Mac's consistent GUI design, productivity software, and baseline stable high resolution i.e. they are not Amiga's demographics. The majority of Amiga's demographics are in Europe, with weaker buying strength. I'm on topic, you're not.