minator Quote:

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

I was just adding the ARM performance claims from Cortex-A76 to Cortex-A720. I agree that the total performance gain should be compounded which is larger and further from the measured Geekbench benchmark. Benchmarks vary and ARM was more likely to be using the SPEC benchmark which I believe is better but the performance seems to be way off or ARM should qualify their claims better like "up to" or "on average with the SPEC benchmark". I believe ARM does compare using the same fab process but newer cores likely get better memory and more area where much of the gains come from. The 10% performance gain of the Cortex-A720 vs Cortex-A78 with the same area is more useful and honest than the other performance claims but still not qualified enough.

minator Quote:

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

I tried to point it out nicely as Hammer gave the same ARM data twice. After the 2nd time he finds the nice ARM pic showing what it means better than I can explain it and then acts like I was attacking the source of his data and him by extension.

minator Quote:

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.

If there was one size of Cortex-A720 cores in a SoC, I would expect the normal or "full configuration" from the pic, especially for the desktop. With 2 different sized Cortex-A720 cores, like the Orion O6 SoC uses, the smaller Cortex-A720 cores could be area optimized but I believe it makes more sense if they were power efficiency (performance/W) optimized and the larger Cortex-A720 cores were performance optimized, especially where higher performance Cortex-X cores could not be used for the desktop. Power/efficiency is usually what is wanted from the in-order cores as well. I am not so sure 3 different sized cores was a good idea as it requires a good understanding of what is ideal and it takes more time to tune.

I doubt area was a major concern or at least not enough to cause the smaller Cortex-A720 cores to be area optimized. The SoC includes a sizable NPU where Cortex-X CPU cores would have been more competitive for the desktop and a PCIe graphics card could have been used to add AI capabilities. The SoC not only has the NPU unit but dual 5Gbps Ethernet so more likely targets edge (of network) AI. Set top, multimedia/streaming and NAS boxes do not particularly need the AI although the networking capabilities are valuable. Good power efficiency and low power are more important than performance for these embedded uses but the Orion O6 does not look good with the high idle power already.

minator Quote:

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

I believe desktop ARM SoCs will continue to be repurposed embedded SoCs for a long time. The problem is the huge advantage x86-64 has in software, drivers and with a standard Windows OS. ARM AArch64 cores are no longer the small low power ARM cores of the past although they scale down further than x86-64 at the low end. There are large microoped OoO ARM cores rivaling x86-64 cores and not necessarily lower power or better power efficiency while ARM at least struggles to create designs that are competitive in performance and Apple and Qualcomm have closed or restricted development and hardware. Any advantage ARM has in area and power is due to less baggage from dropping compatibility but less compatibility may be a larger disadvantage than the baggage. ARM AArch64 also dropped much of the RISC philosophy while adopting many CISC advantages resulting in a hybrid architecture that is closer to CISC than the original minimalist ARM architecture RISC.

@codis

Quote:

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.

Look in the mirror.

ARM/Qualcomm-based Surface X with Microsoft SQ1 or SQ2 SoC is a flop.

Qualcomm is about shipping SoCs and running away from developer support!
https://www.youtube.com/watch?v=qKRmYW1D0S0

Last edited by Hammer on 05-Jun-2025 at 01:45 PM.
Last edited by Hammer on 05-Jun-2025 at 01:19 PM.

_________________
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

@minator

Quote:

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

What you missed is that A720's basic implementation remains the same regardless of "area optimised". For example, the non-area optimized has a larger cache.

https://www.jeffgeerling.com/blog/2025/orion-o6-itx-arm-v9-board-temper-your-expectations
But it doesn't beat the two-year-old Microsoft Windows Dev Kit 2023, with a Qualcomm Snapdragon 8cx Gen 3 SoC.

Orion o6 SoC doesn't beat Qualcomm Snapdragon 8cx Gen 3 SoC.

Cortex-A720's L2 cache can be configured from 128 KB to 512 KB, and L3 cache can be configured from zero to 32 MB. Orion o6 SoC has 12MB L3 cache.

Cortex A78's L2 and L3 cache can also be configured.

ARM PR is for an apples-to-apples comparison.

For Orion o6 SoC's CPU cache configuration from https://docs.radxa.com/en/orion/o6/hardware-design/hardware-interface

4xA720 higher clocked cluster
L1: 64KB inst + 64KB data per CPU core,
L2: Unified (shared) 512KB cache. (Effectively 128KB L2 per CPU core).

4xA720
L1: 64KB inst + 64KB data per CPU core,
L2: Unified (shared) 512KB cache. (Effectively 128KB L2 per CPU core).

4xA520
L1: 32KB inst + 32KB data per CPU core,

L3: Unified (shared) 12MB for CPU cluster. Desktop use case should have max out both of the quad core cluster A720 instead of down clocking it down i.e. shows a mobile phone mentality.

Orion o6 SoC's cache configuration is not optimized for desktop roles.

https://www.cpubenchmark.net/cpu.php?cpu=ARM+CIX+P1+CD8180+7+Core+2500+MHz&id=6561
ARM-based CIX P1 CD8180 7 Core 2500 MHz
Active Cores: 7 Threads: 7

Multithread Rating: 5505
Single Thread Rating: 1395

------------------

https://community.frame.work/t/arm-based-cpus/7203/163
Orion o6's Cix CD8180 SoC was discussed in Framework's forum in Feb 2025 along with RPi CM5 HAT and RK3588 for Framework 13.

https://community.frame.work/t/raspberry-pi-compute-module-5-cm5-carrier-board/60981/25
Devran_Uenal doing the work for getting ARM SoC for Framework 13.

Devran_Uenal is using DeepComputing's RISC-V mainboard for Framework 13 as the template to get an ARM SoC for Framework 13.

Last edited by Hammer on 05-Jun-2025 at 02:45 PM.
Last edited by Hammer on 05-Jun-2025 at 02:03 PM.
Last edited by Hammer on 05-Jun-2025 at 02:00 PM.
Last edited by Hammer on 05-Jun-2025 at 01:40 PM.
Last edited by Hammer on 05-Jun-2025 at 01:29 PM.
Last edited by Hammer on 05-Jun-2025 at 01:12 PM.

_________________
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:

What you missed is that A720's basic implementation remains the same regardless of "area optimised". For example, the non-area optimized has a larger cache.

No I didn't. The area optimised version is different:

From the Cortex-A720 TRM:

Quote:

Reduced area configuration You can configure the core to select the use of a variant of the product that has reduced logic and RAM structures.

The standard A720 is compared to the 2022 A715, whereas the area optimised version is compared to the 2020 A78.

OTOH there is no "basic implementation", it's sold as a soft core, so the implementation entirely down to the chip vendor. There's various configurable parameters in the CPU itself, but then there's all sorts of layout stuff they can mess with.

Quote:

Orion o6 SoC's cache configuration is not optimized for desktop roles.

It's not a desktop processor. Results are somewhat disappointing so far, but maybe it'll improve with better firmware.

_________________
Whyzzat?

@minator

Quote:

The standard A720 is compared to the 2022 A715, whereas the area optimised version is compared to the 2020 A78. OTOH there is no "basic implementation", it's sold as a soft core, so the implementation entirely down to the chip vendor. There's various configurable parameters in the CPU itself, but then there's all sorts of layout stuff they can mess with.

The area optimized A720 vs A720 has the same fundamental (basic) design to classify them to be "A720" instead of "A78" or "A520".

An example of the density optimised for AMD CPU is Zen 4C (compact) vs the larger normal Zen 4. Nothing is free, hence, Zen 4c has a lower clock speed potential when compared to normal Zen 4. Both Zen 4c and Zen 4 have the same pipeline length, with different high clock speed potential results. In terms of potential IPC, they are the same.

Quote:

It's not a desktop processor. Results are somewhat disappointing so far, but maybe it'll improve with better firmware.

https://radxa.com/products/orion/o6/#overview
https://radxa.com/orion/o6/overview/display-view.webp
Radxa promoted desktop Linux apps with Orion O6

Radxa Orion O6 has a PC form factor, a PC ATX power connector, PC's ATX back IO ports location, PC's full-size PCIe 16X physical slot (wired for 8X).

Firmware wouldn't change the CPU's cache design.

https://www.ami.com/blog/tag/cix-cd8180/
Radxa Orion O6 uses AMI's Aptio V UEFI Firmware.

https://www.cnx-software.com/2025/04/30/radxa-orion-o6-mini-itx-motherboard-gets-arm-systemready-sr-v2-5-certification/
Radxa Orion O6 mini-ITX motherboard powered by a CIX P1 12-core Cortex-A720/A520 SoC has just obtained Arm SystemReady SR (ServerReady) v2.5 certification.

Last edited by Hammer on 07-Jun-2025 at 02:44 AM.
Last edited by Hammer on 07-Jun-2025 at 02:35 AM.
Last edited by Hammer on 07-Jun-2025 at 01:48 AM.
Last edited by Hammer on 07-Jun-2025 at 01:46 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

@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.

https://www.cnx-software.com/2025/04/30/radxa-orion-o6-mini-itx-motherboard-gets-arm-systemready-sr-v2-5-certification/

For instance, a few users have already tested the V9.0.0 firmware with ACPI enabled by default, and they could install Windows 11, Ubuntu 24.10, and Fedora rawhide without major issues. However, while several graphics cards can be recognized and work to some extent, there are still some performance issues and quirks. We also know why the LITTLE Cortex-A520 cores have been disabled. If they are enabled in the BIOS, Windows 11/WinPE won’t boot properly, and instead show a BSOD. They should work fine in Linux.


You're not looking at the fine details.

_________________
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 doubt area was a major concern or at least not enough to cause the smaller Cortex-A720 cores to be area optimized. The SoC includes a sizable NPU where Cortex-X CPU cores would have been more competitive for the desktop and a PCIe graphics card could have been used to add AI capabilities. The SoC not only has the NPU unit but dual 5Gbps Ethernet so more likely targets edge (of network) AI.

A significant number of 2025 PC AM5 and Intel LGA 1851 motherboards include 5 / 10 Gbps Ethernet and multiple 10 / 20 / 40 Gbps USB ports.

Quote:

Set-top, multimedia/streaming and NAS boxes do not particularly need the AI although the networking capabilities are valuable. Good power efficiency and low power are more important than performance for these embedded uses but the Orion O6 does not look good with the high idle power already.

Radxa Orion O6's LPDDR5 RAM lacks ECC for an enterprise use case.

I use a cheapo ex-corporate decomssioned Intel Xeon Broadwell with ECC RDIMM DDR4 as my 24/7 file server.

For a new build enterprise workstation entry point, most AM4 and AM5 motherboards from ASUS and ASRock have ECC UDIMM support.

https://pcpartpicker.com/product/bWqBD3/asrock-a520m-hdv-micro-atx-am4-motherboard-a520m-hdv
This low-cost (US$68.48) ASRock A520M-HDV Micro ATX AM4 motherboard has ECC UDIMM support.

Intel places a premium on ECC RAM support via W680 (Alder Lake/Raptor Lake) and W890 (ArrowLake) chipsets.

Missing ECC RAM is just toys since it's a major difference between SUN/HP 68K workstations vs Commodore's A2500UX and A3000UX wannabe.

DDR5 without ECC is just SOHO play.

Major game server providers employ ECC RAM-capable PC motherboards.

https://www.micron.com/about/blog/company/innovations/lpddr5x-with-enhanced-ecc-for-automotive-rises-to-the-challenge
Automotive embedded employs ECC RAM. This example is LPDDR5X ECC.

_________________
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:

matthey wrote: 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).

NPUs and CPUs can't be compared: they have completely different use-cases.
Quote:

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

Correct: AI is currently a moving target, and there are plenty of different models/implementations.

And it looks like that SIMD isn't the best model. Rather, matrixes are.
Quote:

This website has one of the best reviews on the Orion O6 and includes a CPU benchmark.

Geekbench is pure crap: please don't use it as metric for CPUs comparisons.
Quote:

[big images]

Please don't use big images which corrupt the web site layout.

@matthey

Quote:

matthey wrote: I believe desktop ARM SoCs will continue to be repurposed embedded SoCs for a long time. The problem is the huge advantage x86-64 has in software, drivers and with a standard Windows OS. ARM AArch64 cores are no longer the small low power ARM cores of the past although they scale down further than x86-64 at the low end. There are large microoped OoO ARM cores rivaling x86-64 cores and not necessarily lower power or better power efficiency while ARM at least struggles to create designs that are competitive in performance and Apple and Qualcomm have closed or restricted development and hardware. Any advantage ARM has in area and power is due to less baggage from dropping compatibility but less compatibility may be a larger disadvantage than the baggage. ARM AArch64 also dropped much of the RISC philosophy while adopting many CISC advantages resulting in a hybrid architecture that is closer to CISC than the original minimalist ARM architecture RISC.

It's the exact opposite: the original ARM architecture was much more "CISC", whereas AArch64 is much more "RISC" (SIGH!).

ARM simplified A LOT AArch64, following a design similar to the Alpha/MIPS.

However, and to compensate, MANY instructions were added to reduce the number of executed instructions (which also benefitted the code density. In fact, it has a good density if we consider that it has only 32-bit opcodes).

@matthey

Quote:

matthey wrote: 14W at idle is more than the RPi 5 "power virus" max of 12W. So much for peace and quiet with low power ARM. Maybe the Orion O6 comes with a power virus since it is from China. Trumps tariffs will fix the problem though.

Fix what? In which sense? It's unclear.

@cdimauro

Quote:

Geekbench is pure crap: please don't use it as metric for CPUs comparisons.

There's clearly some timing bugs, but it's cross platform, runs a bunch of fairly representative tests, anyone can run it, and it seems to be widely used and respected.

What's wrong with it?

_________________
Whyzzat?

@minator

Quote:

minator wrote: @cdimauro Quote: Geekbench is pure crap: please don't use it as metric for CPUs comparisons. There's clearly some timing bugs, but it's cross platform, runs a bunch of fairly representative tests, anyone can run it, and it seems to be widely used and respected. What's wrong with it?

It's a synthetic benchmark with some algorithms implemented. And it's closed source.

I want to have real-world applications: the ones which are effectively used.

And with source code, so that I can compile and check the generated instructions and the hot spots in the code (so that I can understand where the execution time is most spent).

@cdimauro

https://www.geekbench.com/corporate/

Geekbench allows source code access via the Source License or Development License.

Most major vendors have source access geared towards them e.g. Geekbench 5 being "gamed" before Geekbench 6's release.

_________________
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: @cdimauro https://www.geekbench.com/corporate/ Geekbench allows source code access via the Source License or Development License. Most major vendors have source access geared towards them e.g. Geekbench 5 being "gamed" before Geekbench 6's release.

Only major vendors -> it's something like the SPEC benchmark -> you need to pay.

Which is anyway useless, since it's a just a collection of algorithms and not real-world applications.

@cdimauro

https://youtu.be/Xc8NhN7x6rU?t=1071
€450 Hackintosh in 2025 for MacOS X Sequoia on Intel Core i3-13100F + RX 5700 (36 CU, RDNA1) with real-world benchmarks against
Mac Mini M1 (8 CPU, 8 GPU cores),
Mac Mini M2 (8 CPU, 10 GPU cores),
Mac Mini M4 (10 CPU, 10 GPU cores).


Lightroom AI denoise (less is better)
Intel Core i3-13100F + RX 5700 = 14 seconds
Mac Mini M1 = 1m 31 seconds
Mac Mini M2 = 62 seconds
Mac Mini M4 = 36 seconds


Lightroom CC Export (less is better)
Intel Core i3-13100F + RX 5700 = 6.54 seconds
Mac Mini M1 = 15.9 seconds
Mac Mini M2 = 13.4 seconds
Mac Mini M4 = 8.83 seconds


BlackMagic RAW Speed Test Part 1 (higher is better)
Intel Core i3-13100F + RX 5700 = 39
Mac Mini M1 = 17
Mac Mini M2 = 25
Mac Mini M4 = 35


BlackMagic RAW Speed Test Part 2 (higher is better)
Intel Core i3-13100F + RX 5700 = 147
Mac Mini M1 = 45
Mac Mini M2 = 59
Mac Mini M4 = 72


Final Cut Pro Export (less is better)
Intel Core i3-13100F + RX 5700 = 5m 58s
Mac Mini M1 = 14m 4s
Mac Mini M2 = 9m 12s
Mac Mini M4 = 8m 32s


Shadow of the Tomb Raider Mac version (higher is better)
Intel Core i3-13100F + RX 5700 = 119 fps
Mac Mini M1 = 25 fps
Mac Mini M2 = 31 fps
Mac Mini M4 = 47 fps



BlackMagic Disk Speed Test Read, Write in MB/s (higher is better)
Intel Core i3-13100F + RX 5700 =5200, 4800 (via WD Black SN770 1TB NVMe PCIe 4.0)
Mac Mini M1 = 2800, 3000
Mac Mini M2 = 1510, 1590
Mac Mini M4 = 2900, 2065

It shows inferior storage performance from Apple.

Last edited by Hammer on 12-Jun-2025 at 07:28 AM.
Last edited by Hammer on 12-Jun-2025 at 07:18 AM.
Last edited by Hammer on 12-Jun-2025 at 07:17 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: @cdimauro https://youtu.be/Xc8NhN7x6rU?t=1071 €450 Hackintosh in 2025 for MacOS X Sequoia on Intel Core i3-13100F + RX 5700 (36 CU, RDNA1) with real-world benchmarks against Mac Mini M1 (8 CPU, 8 GPU cores), Mac Mini M2 (8 CPU, 10 GPU cores), Mac Mini M4 (10 CPU, 10 GPU cores). Lightroom AI denoise (less is better) Intel Core i3-13100F + RX 5700 = 14 seconds Mac Mini M1 = 1m 31 seconds Mac Mini M2 = 62 seconds Mac Mini M4 = 36 seconds Lightroom CC Export (less is better) Intel Core i3-13100F + RX 5700 = 6.54 seconds Mac Mini M1 = 15.9 seconds Mac Mini M2 = 13.4 seconds Mac Mini M4 = 8.83 seconds BlackMagic RAW Speed Test Part 1 (higher is better) Intel Core i3-13100F + RX 5700 = 39 Mac Mini M1 = 17 Mac Mini M2 = 25 Mac Mini M4 = 35 BlackMagic RAW Speed Test Part 2 (higher is better) Intel Core i3-13100F + RX 5700 = 147 Mac Mini M1 = 45 Mac Mini M2 = 59 Mac Mini M4 = 72 Final Cut Pro Export (less is better) Intel Core i3-13100F + RX 5700 = 5m 58s Mac Mini M1 = 14m 4s Mac Mini M2 = 9m 12s Mac Mini M4 = 8m 32s Shadow of the Tomb Raider Mac version (higher is better) Intel Core i3-13100F + RX 5700 = 119 fps Mac Mini M1 = 25 fps Mac Mini M2 = 31 fps Mac Mini M4 = 47 fps

You're mostly testing the GPU, whereas I'm only (and clearly) interested on testing only the CPU.

@cdimauro

https://youtu.be/Xc8NhN7x6rU?t=1050
My YouTube link has Geekbench 6 CPU running MacOS X Sequoia for Intel Core i3-13100 and Apple Mac Mini M1/M2/M4.

Major creative apps use the GPU, hence it's platform vs platform.

_________________
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: @cdimauro https://youtu.be/Xc8NhN7x6rU?t=1050 My YouTube link has Geekbench 6 CPU running MacOS X Sequoia for Intel Core i3-13100 and Apple Mac Mini M1/M2/M4. Major creative apps use the GPU, hence it's platform vs platform.

What's not clear to you that I'm NOT interested on GeekBench (because it's a synthetic benchmark) and I'm ONLY interested on PURE CPU benchmarks (and open source, so that I can disassemble the hot spots and check their instructions patterns)?

Getting more on topic, there's been a leak of the much rumoured Nvidia Arm CPUs.

https://videocardz.com/newz/nvidia-n1x-20-core-cpu-performance-leak-reveals-high-end-ambitions


It's Nvidia so probably expensive, but it's another option.
It appears to be similar to the 20 core GB10 CPU in the DGX Spark (10x Cortex-X925 + 10 x Cortex-A725).

Geekbench 6 scores:

3096 single core
18837 multi core

Last edited by minator on 14-Jun-2025 at 04:17 PM.

_________________
Whyzzat?

@minator: it looks good and quite competitive.

nVidia is entering the market with a great product (because it also integrates its awesome GPU).