@codis

Quote:

Interesting. But the ITX form factor suggests it is targeted at the desktop environment. I think they can achieve a substantial market share, with enough performance for the casual and average user.

It's just a Linux ARM desktop box, which itself is a tiny minority.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

@AmigaBlitter

FYI, Cortex A720 is a "medium" CPU core relative to ARM Cortex X4. Cortex A720 is 10 percent improved over Cortex A78.

Last edited by Hammer on 02-Jun-2025 at 12:46 AM.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

@matthey

Quote:

A 60 Teraflops GPU uses 80 billion transistors. A NPU may more efficiently use transistors and TOPs are a measure of simpler integer operations instead of floating point operations but a 30 TOPs NPU could use hundreds of millions if not billions of transistors. In comparison, a 68060 CPU uses 2,530,000 transistors, could be mass produced for less than $1 USD today and this Orion O6 board could maybe emulate a 68060@2GHz (we may soon find out as michalsc has a board). Personally, I would rather have a more powerful GPU that can improve GPU performance and perform AI work even though it may not be as efficient for some AI datatypes. The CPU SIMD units can also use most AI datatypes. I consider AI specific hardware to be wasteful considering the direction of AI is unknown and AI lies. HW ray tracing is not efficient compared to traditional 3D rendering but at least it has some clear benefits like for lighting and can be used sparingly where most effective.

Hardware ray tracing has superior effects accuracy compared to pure raster. AI is useful for ray reconstruction instead of compute shader-based denoise.

Hardware BVH raytracing gets bogged down with the bounding box and triangle test, hence it's geometry processing, and NVIDIA GPUs are king with geometry processing.

For GPU AI,
NV Blackwell tensors have FP6.
NV ADA and Ampere tensors have FP8.
AMD RDNA 4 matrix cores have FP8.

RT cores with Shader Execution Reordering (SER) improve data and execution divergence.

Back in 2023 https://medium.com/@opinali/tracing-rdnas-future-1b82b0e7cea3

The absence of SER in RDNA 3 was a disappointment. Ray Tracing is fundamentally unfit to GPUs, but sorting / reordering rays does some rounding of the square pegs we're pushing into round SIMD holes. Even Intel got the memo, their first-generation Arc Alchemist has a Thread Sorting Unit and theirs don’t require any change in games, although one can optimize for it. I was surprised in fact that NVidia’s SER requires game engine support. Meanwhile, AMD didn’t make this a priority for Navi 3.


PS; NAVI 3 = RDNA 3. Raedon 7900 XTX gets gimped in CyberPunk 2077.

AMD gained Shader Execution Reordering (SER) with RDNA 4(e.g. RX 9600/9700 series), which is included with NVIDIA's ADA Lovelace generation. DXR's Tier 1.2's two extra (Opacity micromaps, Shader execution reordering) features are already enabled in NVAPI, and these improve path tracing workloads.

Saturated hardware BVH raytracing is NVIDIA's new saturated "tessellation" (saturated geometry) gimping the competition. The same heavy geometry playbook from NVIDIA's hardware T&L (killed 3DFX, killed SGI, killed 3DLabs) and saturated "tessellation" (gimped ATi/AMD) tactics.

A reminder to AMD (and others), a GPU is not a DSP. For NVIDIA, it's a good thing that NVIDIA's competitors are out of touch with "what is a GPU?". NVIDIA keeps wreaking competitors with saturated geometry playbook multiple times.

Last edited by Hammer on 02-Jun-2025 at 07:43 AM.
Last edited by Hammer on 02-Jun-2025 at 07:38 AM.
Last edited by Hammer on 02-Jun-2025 at 07:38 AM.
Last edited by Hammer on 02-Jun-2025 at 07:31 AM.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

@Hammer

Quote:

Hammer wrote: It's just a Linux ARM desktop box, which itself is a tiny minority.

Right now it's just a Linux ARM desktop box. Soon it might prove to be an excellent candidate for an emu68 standalone box. Or potentially an AxRT box.

Price:Performance ratio is very good. Plus it pains amigappc1, so that's always a bonus.

_________________
All the way, with 68k

@agami

Quote:

Right now it's just a Linux ARM desktop box. Soon it might prove to be an excellent candidate for an emu68 standalone box. Or potentially an AxRT box.

AxRT runs on modern Linux on either ARM or X86 or Windows 11 WSL2.

AxRT doesn't run retro 68K Amiga software. AxRT doesn't include Box68K.

Quote:

Right now it's just a Linux ARM desktop box. Soon it might prove to be an excellent candidate for an emu68 standalone box. Or potentially an AxRT box.

The main point with Emu68 is speeding up 68K CPU performance with retro C= Amiga chipset.

Another Amithlon kills retro Amiga games.

Quote:

Price:Performance ratio is very good. Plus it pains amigappc1, so that's always a bonus.

That's debatable.

https://www.cpubenchmark.net/cpu_lookup.php?cpu=ARM+Cortex-A720+8+Core+2600+MHz&id=6722
Ranking for ARM Cortex A720 with 8 cores @ 2600Mhz
$211.90 for 8GB
$461.90 for 64GB

https://www.cpubenchmark.net/cpu.php?cpu=ARM+Cortex-A720+8+Core+2600+MHz&id=6722

ARM Cortex A720 with 8 cores @ 2600Mhz
Multithread Rating: 6141
Single Thread Rating: 1455



https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i5-12400&id=4677
Intel Core i5-12400
Multithread Rating: 19100
Single Thread Rating: 3489


https://www.cpubenchmark.net/cpu.php?cpu=AMD+Ryzen+5+8500G&id=5841
AMD Ryzen 5 8500G
Multithread Rating: 21631
Single Thread Rating: 3893


-------------------------------------------------
Intel camp:
8GB RAM
https://pcpartpicker.com/list/QMktBq
CPU: Intel Core i5-12400 2.5 GHz 6-Core Processor ($143.57 @ Amazon)
Motherboard: ASRock B760M-H2/M.2 Micro ATX LGA1700 Motherboard ($93.67 @ Amazon)
Memory: Kingston FURY Beast 8 GB (1 x 8 GB) DDR5-6000 CL30 Memory ($39.99 @ Amazon)
Total: $277.23


64GB RAM
https://pcpartpicker.com/list/MkNmHW
CPU: Intel Core i5-12400 2.5 GHz 6-Core Processor ($143.57 @ Amazon)
Motherboard: ASRock B760M-H2/M.2 Micro ATX LGA1700 Motherboard ($93.67 @ Amazon)
Memory: Silicon Power XPOWER Pulse Gaming 32 GB (2 x 16 GB) DDR5-6000 CL36 Memory ($67.49 @ Newegg Sellers)
Total: $304.73

-------------------------------------------------
AMD camp:
8GB RAM
https://pcpartpicker.com/list/4Rz6GP
CPU: AMD Ryzen 5 8500G 4.1 GHz 6-Core Processor ($149.00 @ B&H)
Motherboard: ASRock B650M PG Lightning Wifi Micro ATX AM5 Motherboard ($119.99 @ Amazon)
Memory: Kingston FURY Beast 8 GB (1 x 8 GB) DDR5-6000 CL30 Memory ($39.99 @ Amazon)
Total: $308.98

https://www.aliexpress.com/w/wholesale-B650.html?g=y&SearchText=B650&sortType=price_asc
Since Radxa is PRC company, PRC branded AM5 B650 motherboards starts from $72.88 USD.
Total: 261.87


16GB RAM
https://pcpartpicker.com/list/xsThJn
CPU: AMD Ryzen 5 8500G 4.1 GHz 6-Core Processor ($149.00 @ B&H)
Motherboard: ASRock B650M PG Lightning Wifi Micro ATX AM5 Motherboard ($119.99 @ Amazon)
Memory: Silicon Power XPOWER Storm RGB 64 GB (2 x 32 GB) DDR5-6000 CL38 Memory ($127.99 @ Newegg Sellers)
Total: $396.98

PC RAM and PC motherboards can be cheaper from PRC.

Last edited by Hammer on 02-Jun-2025 at 07:23 AM.
Last edited by Hammer on 02-Jun-2025 at 07:13 AM.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

@Hammer

Quote:

AxRT doesn't run retro 68K Amiga software. AxRT doesn't include Box68K.

To my knowledge there's nothing stopping it to happen, it haven't happened yet simply because of a) man power b) people applying to it,

Regarding b) one would think that if that deemed so important, people would jump at the opportunity and help doing it, but sadly no.

_________________
Indigo 3D Lounge, my second home.
The Illusion of Choice | Am*ga

@michalsc

Quote:

michalsc wrote: I have it on my desk, but waiting for EFI updates before I continue with my experiments.

Have you ever thought about an 'Amiga Emulator' where the emulation of AGA/ECS runs on one ARM core where another core does the 68k emulation?

All this new ARM systems have multiple cores that are totally wasted on most Amiga / 68k emulations. You have a lot experience with ARM systems and Amiga, maybe you can program an Amiga emulator that runs better on ARM than any UAE version could.

@OneTimer1


Quote:

All this new ARM systems have multiple cores that are totally wasted on most Amiga / 68k emulations. You have a lot experience with ARM systems and Amiga, maybe you can program an Amiga emulator that runs better on ARM than any UAE version could.

Only now that is becoming a reality on WinUAE 6, it would be amazing, but let keep the expectations reasonable

_________________
Indigo 3D Lounge, my second home.
The Illusion of Choice | Am*ga

@OneTimer1

Quote:

OneTimer1 wrote: Have you ever thought about an 'Amiga Emulator' where the emulation of AGA/ECS runs on one ARM core where another core does the 68k emulation? All this new ARM systems have multiple cores that are totally wasted on most Amiga / 68k emulations. You have a lot experience with ARM systems and Amiga, maybe you can program an Amiga emulator that runs better on ARM than any UAE version could.

Note that Eum68 also uses multiple ARM cores i.e. individual CPU cores for translation, I/O (e.g. fetch), and debug.

Like the separate RTG thread improvement, WinUAE 6 beta's Amiga chipset emulation processing is moved to secondary CPU threads. WinUAE 6 beta also moved to the X86-64 v3 (AVX2) instruction set target due to potentially useful instructions for emulation.

Last edited by Hammer on 03-Jun-2025 at 02:09 AM.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

@pixie

Quote:

To my knowledge there's nothing stopping it to happen, it haven't happened yet simply because of a) man power b) people applying to it, Regarding b) one would think that if that deemed so important, people would jump at the opportunity and help doing it, but sadly no.

Major AROS maintainers don't follow the MorphOS direction, which includes userland 68K emulation.

AROS 68K still has issues with 68K AmigaOS compatibility. AROS 68K still has a long way to go to displace AmigaOS 3.1 or 3.X.

AROS PPC still has a long way to go to displace AmigaOS 4.1 FE, A-EON System 54, and MorphOS.

Last edited by Hammer on 03-Jun-2025 at 02:22 AM.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

Hammer Quote:

FYI, Cortex A720 is a "medium" CPU core relative to ARM Cortex X4. Cortex A720 is 10 percent improved over Cortex A78.

ARM claims would be 30% performance improvement from Cortex-A78 to Cortex-A720 but from the RPi 5 Cortex-A76 to Cortex-A720 is more interesting for this thread.

+15% Cortex-A715_to_Cortex-A720 https://en.wikipedia.org/wiki/ARM_Cortex-A720
+5% Cortex-A710_to_Cortex-A715 https://en.wikipedia.org/wiki/ARM_Cortex-A715
+10% Cortex-A78_to_Cortex-A710 https://en.wikipedia.org/wiki/ARM_Cortex-A710
+ 7% Cortex-A77_to_Cortex-A78 https://en.wikipedia.org/wiki/ARM_Cortex-A78
+25% Cortex-A76_to_Cortex-A77 https://en.wikipedia.org/wiki/ARM_Cortex-A77
---
+62% performance from Cortex-A76_to_Cortex-A720

Geekbench single core benchmark performance of Orion O6 with Cortex-A720 is 45% better than RPi 5 Cortex-A76 and 34% after adjusting for clock speed difference. ARM claims should be with the same process where the Orion O6 has an advantage with a claimed 6nm process compared to the RPi 5 16nm process. The Orion O6 also has an advantage with LPDDR5 memory compared to the RPi 5 LPDDR4X. The newer ARM cores like the Cortex-A720 are not getting as large of a performance gain as earlier improvements and it looks like most gains come from silicon and memory upgrades rather than major core design improvements.

The 10% gain over a Cortex-A78 that you used likely comes from the Wiki.

https://en.wikipedia.org/wiki/ARM_Cortex-A720 Quote:

Architecture changes in comparison with ARM Cortex-A715 o Update to ARMv9.2 o 15% peak performance improvement over the Cortex-A715 o Can down to same size as Cortex-A78 with 10% performance improvement o Area optimize configuration for no area cost vs Cortex-A78 o Down L2 cache hit latency to 9 cycles (from 10 cycles) o Down mispredict latency to 11 cycles (from 12 cycles)[6] o x2 L2 bandwidth o DSU-120 o Up to 14 cores (up from 12 cores) o Up to 32 MiB of shared L3 cache (increased from 16 MiB)

I read it as the Cortex-A720 can be scaled down to the area of a Cortex-A78 and still has a 10% performance gain. Most Cortex-A720 cores use a newer fab process that allows more transistors/area and tune for more performance. Of the 30% performance gain from the Cortex-A78 to the Cortex-A720, 10% comes from design improvements and likely 20% from larger SRAM caches and buffers (modern SoCs can be 90% SRAM or more). SRAM transistor density is no longer scaling with the process node size though.

https://fuse.wikichip.org/news/7343/iedm-2022-did-we-just-witness-the-death-of-sram/
https://www.tomshardware.com/tech-industry/sram-scaling-isnt-dead-after-all-tsmcs-2nm-process-tech-claims-major-improvements

The TSMC 2nm process is able to improve SRAM density after the TSMC 3nm process had practically none but this may not continue. Would ARM be able to continue selling new core designs if three generations of core designs only gives a 10% performance improvement? Is the advantage of code density for caches increased with SRAM scaling slowing?

More active transistors use more power and generate more heat while ARM relies on newer processes to reduce power. It is also possible to move older cores like the Cortex-A76 with fewer transistors to newer processes where they can be clocked up more to close the performance gap and they may be lower power than the newer cores as they use fewer transistors. The Orion O6 vs RPi 5 has a 46% single core performance advantage with 2.6GHz vs 2.4GHz cores.

Radxa Orion O6 2.6GHz vs Raspberry Pi 5 2.4GHz
https://browser.geekbench.com/v6/cpu/compare/11810701?baseline=12249204

Radxa Orion O6 2.6GHz vs Raspberry Pi 5 3.0GHz
https://browser.geekbench.com/v6/cpu/compare/11810701?baseline=12255615

If the RPi 5 is clocked at 3.0GHz though, the single core performance advantage of the Orion O6 is down to 22%. The RPi 5 Cortex-A76 uses fewer transistors generating less heat and fewer transistors use less area lowering the production cost, increasing the number of chips per wafer and improving yields. The increased power from higher frequency may offset the decreased power from fewer active transistors. This is why power efficiency (performance/W) is important rather than trading power and area for performance as some of ARMs newer core designs do, which is especially bad for their small and mid cores. That leads to the question of why did the Orion O6 not use ARM big cores where they are using three different sizes of cores. The big cores have more than double the performance. Is it that the big core area more than triples for a Cortex-X925 over a Cortex-A720, the much higher power and heat, the cost to license new big cores, restrictions on China having the newest technology?

The following link has a nice chart comparing newer ARM cores which includes the Cortex-A720 which is more like the Apple little core or e-core.

https://www.reddit.com/r/hardware/comments/1gvo28c/latest_arm_cpu_cores_compared_performanceperarea/

With a 6nm fab process, ARM mid cores with only a 2.6GHz clock speed is underwhelming. I wish I could say the Radxa Orion O6 was super low power for NAS and multimedia boxes but it idles at 14W. The price is ok and the small board size nice but is it worth it to get a PCIe slot that could support modern discreet GPUs if there were drivers? Is it that much better than a simpler and older tech Radxa ROCK 5 ITX+ without PCIe that uses the I/O and SerDes for other things like a 2nd M.2 M slot that can provide a 2nd NVMe drive or SATA?

https://radxa.com/products/rock5/5itxp

@matthey

Quote:

ARM claims would be 30% performance improvement from Cortex-A78 to Cortex-A720 but from the RPi 5 Cortex-A76 to Cortex-A720 is more interesting for this thread. +15% Cortex-A715_to_Cortex-A720 https://en.wikipedia.org/wiki/ARM_Cortex-A720 +5% Cortex-A710_to_Cortex-A715 https://en.wikipedia.org/wiki/ARM_Cortex-A715 +10% Cortex-A78_to_Cortex-A710 https://en.wikipedia.org/wiki/ARM_Cortex-A710 + 7% Cortex-A77_to_Cortex-A78 https://en.wikipedia.org/wiki/ARM_Cortex-A78 +25% Cortex-A76_to_Cortex-A77 https://en.wikipedia.org/wiki/ARM_Cortex-A77 --- +62% performance from Cortex-A76_to_Cortex-A720 Geekbench single core benchmark performance of Orion O6 with Cortex-A720 is 45% better than RPi 5 Cortex-A76 and 34% after adjusting for clock speed difference. ARM claims should be with the same process where the Orion O6 has an advantage with a claimed 6nm process compared to the RPi 5 16nm process. The Orion O6 also has an advantage with LPDDR5 memory compared to the RPi 5 LPDDR4X. The newer ARM cores like the Cortex-A720 are not getting as large of a performance gain as earlier improvements and it looks like most gains come from silicon and memory upgrades rather than major core design improvements. The 10% gain over a Cortex-A78 that you used likely comes from the Wiki.

You got my attention.

https://community.arm.com/arm-community-blogs/b/announcements/posts/arm-cortex-a720-and-cortex-a520-cpus

Cortex-A720 can be configured at implementation time to match the area footprint of the Arm Cortex-A78, but provides 10 percent more performance²





Your "The 10% gain over a Cortex-A78 that you used likely comes from the Wiki." is FALSE.

Come at me.

Quote:

If the RPi 5 is clocked at 3.0GHz though, the single core performance advantage of the Orion O6 is down to 22%. The RPi 5 Cortex-A76 uses fewer transistors generating less heat and fewer transistors use less area lowering the production cost, increasing the number of chips per wafer and improving yields.

Both Cortex A76 and A78 (recycled for A720) have 4-way decoders.
There's a major difference in the load and store unit design.

Cortex A76 has two 16-byte loads and a single 32-byte store. 32byte per cycle L1-L2 bandwdith.

Cortex A78 has three 16-byte loads and a single 32-byte store. 64byte per cycle L1-L2 bandwdith.

Similar to major Zen 2 to Zen 3 improvements are with load-store design.

https://www.cpubenchmark.net/cpu.php?cpu=ARM+Cortex-A76+4+Core+3000+MHz&id=5739
ARM Cortex-A76 4 Core @ 3000 MHz
Multithread Rating: 2578
Single Thread Rating: 1299

https://www.cpubenchmark.net/cpu.php?cpu=BCM2712&id=6054
BCM2712, Arm Cortex-A76 @ 2400 Mhz
Multithread Rating: 3284
SingleThread Rating: 1464


Quote:

The increased power from higher frequency may offset the decreased power from fewer active transistors. This is why power efficiency (performance/W) is important rather than trading power and area for performance as some of ARMs newer core designs do, which is especially bad for their small and mid cores. That leads to the question of why did the Orion O6 not use ARM big cores where they are using three different sizes of cores. The big cores have more than double the performance. Is it that the big core area more than triples for a Cortex-X925 over a Cortex-A720, the much higher power and heat, the cost to license new big cores, restrictions on China having the newest technology?

Since ARM Holdings (owned by SoftBank from Japan) doesn't fully control ARM China, it may not have access to ARM's Cortex X series. Europe effectively lost control over ARM since the UK government is a loser.

For X925 generation, A725 displaced A720.

This is similar to AMD licensing Zen 1.x technology for China's Hygon Dhyana project.

Like Japan's 1960 to 1980 era, "local partner" protectionist policies are required in China.

The US's ARM clone champions are Qualcomm's Oryon and Apple's M series.

Last edited by Hammer on 03-Jun-2025 at 08:19 AM.
Last edited by Hammer on 03-Jun-2025 at 08:12 AM.
Last edited by Hammer on 03-Jun-2025 at 07:56 AM.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

@Hammer

Quote:

Hammer wrote: @agami Quote: Price:Performance ratio is very good. Plus it pains amigappc1, so that's always a bonus. That's debatable.

I apologise. Let me qualify that with stating that Price:Performance is very good for big-endian computing.

You know how funny some of our cohort get surrounding the topic of x86/x64. Other than to ppcamiga1, ARM(64) represents a palatable target for moving away from PowerPC (Power ISA). Offering hardware from the sub $100 USD bracket (for the masses), and in every bracket up to $3,000 plus systems like the Thelio Astra from System76 (for the elites compensating inadequacies). Including laptops, tablets, and even an open source "PinePhone" smartphone from Pine64.

I am a big fan of what AMD have been doing since the the X5, their 3DNow SIMD extension, Athlon/Opteron architecture, x86-64 extension, early multi-core CPUs, acquisition of ATi, innovation in APUs, standardising the gaming console market, Ryzen, power efficiency for portable computing and now handheld gaming.

As much as I would love to see and would likely be a Day One user of MorphOS for x64, ARM represents a more versatile hardware ecosystem.

Last edited by agami on 03-Jun-2025 at 07:40 AM.
Last edited by agami on 03-Jun-2025 at 07:38 AM.

_________________
All the way, with 68k

@Hammer

if you say so...

and for what do you need the PPC version?

Last edited by OlafS25 on 03-Jun-2025 at 08:58 AM.

@Hammer

Quote:

It's just a Linux ARM desktop box, which itself is a tiny minority.

No.
You can run an ARM version of M$ Win11 on it - not that I personally want that.
And it is just one of those boards released or announced lately.

You seem a bit too focussed on the past.

Hammer Quote:

You got my attention. https://community.arm.com/arm-community-blogs/b/announcements/posts/arm-cortex-a720-and-cortex-a520-cpus Cortex-A720 can be configured at implementation time to match the area footprint of the Arm Cortex-A78, but provides 10 percent more performance² ... Your "The 10% gain over a Cortex-A78 that you used likely comes from the Wiki." is FALSE. Come at me.

I was not implying that your source of data was wrong. Most data on Wiki is correct as it is in this case. I do question your understanding of the data you spam sometimes. The ARM pic is good and illustrates my point. Three generations of ARM mid core design improvements provide a 10% performance gain without increasing the core area which is mostly for SRAM caches and buffers. Actually, it could be worse as part of that 10% performance gain may be due to ISA improvements from ARMv8.2 to ARMv9.2. It sounded like you were expecting a 10% performance gain from the Cortex-A78 to the Cortex-A720 but the big Cortex-A720 cores in the Orion O6 SoC are likely more like the "Full configuration" in the ARM pic than the "Area optimized" configuration which shows a larger performance gain. By using two different configured Cortex-A720 cores, the big core could be a "Performance optimized" Cortex-A720 which has more performance than the "Full configuration" and the mid core Cortex-A720 may use a "Power optimized" configuration.

Another take away from the ARM pic is that the area roughly doubled from a Cortex-A78 to a "Full configuration" Cortex-A720. This is for ARM's mid core OoO designs while their big/p-core OoO designs are likely increasing more and faster. ARM stopped giving transistor counts, which is a process independent rough measure of area, with the OoO Cortex-A57.

Year | CPU | transistors
1975 6502 3,500
1979 68000 68,000
1984 68020 190,000
1985 ARM1 25,000
1985 80386 275,000
1986 ARM2 30,000
1987 68030 273,000
1990 68040 1,170,000
1993 Pentium 3,100,000 superscalar in-order 2-way
1994 68060 2,530,000 superscalar in-order 2-way
1994 ARM7 250,000
1995 PentiumPro 5,500,000 OoO uop
2002 ARM11 7,500,000 (RPi 1 core)
2008 Nehalem 731,000,000 (1st gen Core i7 with 4 cores) 64 bit OoO uop
2011 Cortex-A7 10,000,000 superscalar in-order 2-way (RPi 2 has 4 cores)
2012 Cortex-A53 12,500,000 64-bit superscalar in-order 2-way (RPi 3 has 4 cores)
2012 Cortex-A57 75,000,000 64-bit OoO 3-way big.LITTLE companion of Cortex-A53

ARM OoO big/mid core design progression:
Cortex-A57, Cortex-A72, Cortex-A73, Cortex-A75, Cortex-A76, Cortex-A77, Cortex-A78 ,Cortex-A710, Cortex-A715, Cortex-A720

If the transistors doubled every 3 generations of ARM designs as the ARM pic shows for ARM-Cortex-A78 to Cortex-A720, the transistors would have increased from the Cortex-A57 75 million transistors per core to 600 million transistors per Cortex-A720 core (8 times the transistors). Maybe a Cortex-A720 core with 600 million transistors could emulate a high clocked 68060 with 2.5 million transistors but the 68060 on newer silicon could be mass produced for $0.50 USD while the Cortex-A720 SoC has 8x 600 million transistor cores, 4x Cortex-A520 in-order cores larger than Cortex-A53 cores above and a GPU that may use more transistors than all these CPU cores combined which may cost $50 USD per SoC to mass produce. Not only are billions of transistors required for the Orion O6 SoC instead of millions for a 68060 and perhaps tens of millions for a 68k SoC that costs less than a meal, but the 68k SoC may allow another $10 in memory cost savings. The Orion O6 SoC is not even competitive in performance against x86-64 SoCs or even smart phone and tablet ARM SoCs with more powerful p-cores and it does not seem to trade performance for much lower power and price either.

I do not want to get in some kind of competition with you and what I write is not just directed at you but the forum thread of what I see as the big picture which I try to put in perspective for others to understand. I am no tech genius and appreciate when others can fill in the gaps in my understanding but please do not try to make me wrong when I am right. Also, please try to focus your points and tech data to the topic.

Hammer Quote:

Since ARM Holdings (owned by SoftBank from Japan) doesn't fully control ARM China, it may not have access to ARM's Cortex X series. Europe effectively lost control over ARM since the UK government is a loser. For X925 generation, A725 displaced A720. This is similar to AMD licensing Zen 1.x technology for China's Hygon Dhyana project. Like Japan's 1960 to 1980 era, "local partner" protectionist policies are required in China. The US's ARM clone champions are Qualcomm's Oryon and Apple's M series.

The UK government had conditions for SoftBank to buy ARM and Japan is an ally. Yes, Japan has protectionist policies but they are nothing like China's required ownership of businesses and requirements of not just businesses but Chinese individuals to surrender info to the Chinese government. Russia, China and North Korea are also in a proxy war with the West which may be WWIII in the early stages. It would be worse if ARM was Chinese owned like former UK tech business Imagination Technologies.

https://en.wikipedia.org/wiki/Imagination_Technologies Quote:

Imagination Technologies Group Limited is a British semiconductor and software design company owned by Canyon Bridge Capital Partners, a private equity fund based in Beijing that is ultimately owned by the Chinese government. With its global headquarters in Kings Langley, England, its primary business is in the design of PowerVR mobile graphics processors (GPUs), neural network accelerators for AI processing, and networking routers. The company was listed on the London Stock Exchange until it was acquired by Canyon Bridge in November 2017.

The Chinese government forces the ARM China subsidiary to hand over ARM IP and operates independently of ARM including performing their own R&D. This led to ARM trying to sidestep ARM China.

https://www.tomshardware.com/pc-components/cpus/arm-wants-to-sell-directly-to-chinese-customers-sidestep-arm-china

China is a large part of ARM's business which is problematic for ARM that has been under performing recently, some of which may be because of increased sanctions on tech to China and ARM limiting their IP available to ARM China in order to protect it from the Chinese government. The Neoverse V-series CPU core designs for datacenters and supercomputers are mentioned in the ban for China but maybe ARM is choosing to protect other highest tech ARM IP in China.

Last edited by matthey on 03-Jun-2025 at 06:41 PM.

@matthey

Quote:

+15% Cortex-A715_to_Cortex-A720 https://en.wikipedia.org/wiki/ARM_Cortex-A720 +5% Cortex-A710_to_Cortex-A715 https://en.wikipedia.org/wiki/ARM_Cortex-A715 +10% Cortex-A78_to_Cortex-A710 https://en.wikipedia.org/wiki/ARM_Cortex-A710 + 7% Cortex-A77_to_Cortex-A78 https://en.wikipedia.org/wiki/ARM_Cortex-A78 +25% Cortex-A76_to_Cortex-A77 https://en.wikipedia.org/wiki/ARM_Cortex-A77 --- +62% performance from Cortex-A76_to_Cortex-A720

It's compound, so it should be should be +78%.

Quote:

It sounded like you were expecting a 10% performance gain from the Cortex-A78 to the Cortex-A720

A fair assumption given that it's *exactly* what he wrote:
Quote:

Quote: The ARM Cortex-A720 is based on the ARM Cortex-A78, with a 10 percent improvement.

He completely missed that this was for the area optimised version.

Quote:

Geekbench single core benchmark performance of Orion O6 with Cortex-A720 is 45% better than RPi 5 Cortex-A76 and 34% after adjusting for clock speed difference.

I thought that number was way out, but if they've used the area optimised core it's almost exactly where it should be. That said, it's not clear what clock the CPU is actually running at and the firmware is flaky to say the least.

However, it could be that it's an area-optimised low end chip, probably destined for Android TV boxes, Chromebooks and the like. Their Arm based competitors are still selling A76 based chips so they can charge what they want in mini desktops for now.

I think it'll get more interesting once proper Arm desktop chips start to appear.

_________________
Whyzzat?

@codis

Quote:

codis wrote: @Hammer Quote: It's just a Linux ARM desktop box, which itself is a tiny minority. No. You can run an ARM version of M$ Win11 on it - not that I personally want that. And it is just one of those boards released or announced lately. You seem a bit too focussed on the past.

Windows 11 is free, hence, you seem a bit too focused on the past with your M$.

Without a valid Windows 11 product key, the personalization feature is disabled, and a nagging watermark appears.

Windows 11 ARM flopped hard.
https://www.youtube.com/watch?v=dnC_b9nhnQ8
Higher return rate on ARM PCs.

https://www.youtube.com/watch?v=qKRmYW1D0S0
ARM Architecture is good, but Qualcomm's Snapdragon is not our friend! Wendell covers the disappointing reality of Microsoft's collaboration.

Last edited by Hammer on 05-Jun-2025 at 04:45 AM.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

@matthey

Quote:

I was not implying that your source of data was wrong. Most data on Wiki is correct as it is in this case. I do question your understanding of the data you spam sometimes.

Your "The 10% gain over a Cortex-A78 that you used likely comes from the Wiki" assertion is not correct.

ARM's PR slide is another issue.

ARM Cortex-A720 is slightly enhanced from Cortex-A78, which itself is slightly enhanced from A77 and A76.

Quote:

The UK government had conditions for SoftBank to buy ARM and Japan is an ally. Yes, Japan has protectionist policies but they are nothing like China's required ownership of businesses

Japan changed its protectionist policy in the 1980s.

Europe focuses on RISC-V when the RISC-V Foundation relocates to Switzerland.

Refer to https://escholarship.org/content/qt6nm902k9/qt6nm902k9.pdf?t=n4ovvz

2. Protection

Until the 1970s, Japan's domestic semiconductor market was
substantially closed by tariffs, quotas, and other barriers. Foreign
investment, when permitted, was predicated on transfers of technology to Japanese firms.
Once Japan approached technological parity with the United States, however, it gradually removed controls, eliminating duties entirely in 1985.

Keiretsu Purchasing Practices

Each of Japan's top six semiconductor manufacturers is a
member of a keiretsu, one of the giant corporate families of Japan. Keiretsu are interrelated companies typically grouped around a common bank and a common trading company.

Users of semiconductors and makers of semiconductor equipment are likewise keiretsu members.

Linked by corporate ties, long-term supply arrangements, and an ongoing chain of reciprocal business favors, members feel compelled to buy from a keiretsu sibling even when an outside product offers a more economically justifiable choice.

Thus, the argument goes, the business culture of Japan can thwart sales of U.S. semiconductors even if they are freely imported and packed with comparative
advantage



https://www.usitc.gov/publications/332/working_papers/the_south_korea-japan_trade_dispute_in_context_semiconductor_manufacturing_chemicals_and_concentrated_supply_chains.pdf

Japan’s semiconductor and semiconductor manufacturing equipment industries developed in tandem.

Beginning in the 1960s, Japan’s semiconductor industry engaged in license agreements with thenleading U.S. firms, which resulted in a transfer of intellectual property and key technologies. In 1965, Tokyo Electron Laboratories (now known as “TEL,� Japan’s largest semiconductor manufacturing equipment (SME) firm) partnered with Fairchild Semiconductor (United States) to act as Fairchild’s distributer of testing systems in Japan and later partnered with Lam Research (United States) in the 1980s.

At the same time, the Japanese government engaged in policies preventing leading U.S. semiconductor firms from directly investing in Japan, protecting nascent Japanese semiconductor firms from competition. Japanese firms leveraged knowledge spillovers derived from international partnerships and a favored position in the domestic market in the 1980s to develop dynamic random-access memory (DRAM), with Japanese firm market share increasing from less than 30 percent to nearly 75 percent between 1978 and 1986.


This 2019 essay examines how the US is concerned with the decline of the Japanese semiconductor industry and the trade conflict between South Korea and Japan.

This is not a simple company vs company when state power is involved.

"Japan Inc." ordered Panasonic to withdraw 3DO M2 (IBM PPC, 3DO GPU) from the games console market i.e. don't compete against Sony PlayStation (LSI Logic MIPS, Toshiba-JP GPU), Sega (Hitachi-JP SuperH, Sega GPU), and Nintendo (SGI MIPS and GPU for N64).

University of California's RISC-V's improved code density is a lesson learned from MIPS (Stanford University). MIPS gained 16-bit instructions in 1996.

Both the US and Japan later recognized that the survival of their respective semiconductor industries is important for mutual defense capability i.e. a war of attrition would consume all the industrial capacity from both nations, and that's not even enough i.e. South Korea and Taiwan later joined the defense group.

NATO-EU should cover the European-Russian front. Europeans running out of smart weapons during the Libyan adventure wasn't smart.

Trump can be noisy like a New Yorker, but logistics can't be ignored.

_________________
Amiga 1200 (rev 1D1, KS 3.2, PiStorm32/RPi CM4/Emu68)
Amiga 500 (rev 6A, ECS, KS 3.2, PiStorm/RPi 4B/Emu68)
Ryzen 9 7950X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB

@Hammer

Quote:

Windows 11 is free, hence, you seem a bit too focused on the past with your M$. Without a valid Windows 11 product key, the personalization feature is disabled, and a nagging watermark appears.

Now I get it why some other posters are engaged with you in certain mudslinging threads, where you permanently hurl ad hominems at each other.

Good luck with that !