Thanks for this.Here is 12700T vs w5-3435X, Both based on Golden Cove, the latter having AMX: https://browser.geekbench.com/v6/cpu/compare/5330257?baseline=4656765
I didn't pick the fastest scores, but checked both are using the same GB version.
The effect of AMX can be seen on Object Detection. It also looks like Background Blur is affected.
It’s only an issue when Apple has a huge lead.Extremely funny how these large increases in object detection weren’t an issue until now.
Look Intels adding more neural extensions to Lunar and Arrow lake?
Intel confirms Arrow Lake-S & Lunar Lake CPUs will support instructions for AVX-VNNI, SHA512, SM3, SM4 and LAM - VideoCardz.com
Intel Arrow Lake-S appears in instruction set reference guide Intel has just updated its “Architecture Instruction Set Extensions and Future Features” last month, which confirms what instructions will be supported by upcoming processors. The guide confirms that both series will support a range...videocardz.com
I guess we gotta check the IPC improvement for Intels upcoming CPUs without all these extensions now…
AgreedFine by me. I know you're trying to make some kind of point about people being unfair to Apple, and maybe there is some of that going on; I think my posting history makes clear I'm not one of those - but there absolutely is value, a lot of it, in knowing what a core is likely to do with a random existing application that isn't implementing the buzzword of the week and isn't affected by specialized acceleration for it.
This applies to everyone - Intel, AMD, Apple, IBM, anyone else.
Completely agree. Just having some fun with the GB6 charts I been seeing on weibo and Twitter. It’s been real fun learning all these extensions.Fine by me. I know you're trying to make some kind of point about people being unfair to Apple, and maybe there is some of that going on; I think my posting history makes clear I'm not one of those - but there absolutely is value, a lot of it, in knowing what a core is likely to do with a random existing application that isn't implementing the buzzword of the week and isn't affected by specialized acceleration for it.
This applies to everyone - Intel, AMD, Apple, IBM, anyone else.
SME is not only for matrix operations, it also includes a subset of HPC-oriented SVE instructions. Apples implementation uses 512-bit vectors, so comparisons to AVX-512 are appropriate as long as one stays within the HPC domain.
So basically, GB6 has been using Intel's AMX all along. While Apple had AMX, it wasn't utilized by GB6 because you have to use Apple's CoreML library to even target it. CoreML will automatically run your code through NPU, AMX, or GPU so there was no way to guarantee it. Hence, it was left out of GB6.
This is just Apple/ARM having more feature parity with x86 instructions. x86 vs M4 scores for GB6 are valid then.
Looks like AMD also benefits from AVX512-VNNI?
Zen 4:
Zen 3:Micro-Star International Co., Ltd. MS-7E16 - Geekbench
Benchmark results for a Micro-Star International Co., Ltd. MS-7E16 with an AMD Ryzen 7 7700X processor.browser.geekbench.com
Gigabyte Technology Co., Ltd. B550M DS3H - Geekbench
Benchmark results for a Gigabyte Technology Co., Ltd. B550M DS3H with an AMD Ryzen 7 5700X processor.browser.geekbench.com
Huge uplift in Object Detections section. So anyone going to calculate IPC without Object detection now for Zen 4 cause of AI shenanigans.
GB is getting more and more useless with each update. They need to stop with the ST/MT crap scores since the MT score is moot anyway, divide the tests into categories and derive a single composite score from the category scores. Composite score seems too high? Dive into the category score comparison to see which one is the highest. Category score seems too high? Compare the subtests in that category.No one should be taking a GB overall score as a standard for "IPC" anyway.
Check the license notifications.
It’s not based on DXVK.Which is based on and includes code from DXVK. Yet again Apple taking advantage of permissive license morons and not distributing their derivative source code (of course they have no legal obligation, but simple non-American ethics follows a take and give back model).
License disclosure indicates they are in fact using some of DXVK's code. They could open source it to show to what extent they are using it but they won't because Apple.
*All* benchmarks need digging into subtests. The most stupid one is AnTuTu that aggregates CPU/GPU/MEM/UX scores as a single figureGB is getting more and more useless with each update. They need to stop with the ST/MT crap scores since the MT score is moot anyway, divide the tests into categories and derive a single composite score from the category scores. Composite score seems too high? Dive into the category score comparison to see which one is the highest. Category score seems too high? Compare the subtests in that category.
Intel had SPEC results invalidated by overly specific compiler optimizations earlier this very year. So it's an ongoing problem.This issue already happened with SPEC results where some compilers from Sun and Intel were "breaking" some of the tests
They also apparently got permission to include Apple's game porting toolkit in CrossOver, despite licensing terms suggesting this should not be allowed.At least in this case CodeWeavers got a nice paycheck (presumably) for doing the dirty work.
No, that's definitely allowed - all the LGPL and GPL components have their source available.They also apparently got permission to include Apple's game porting toolkit in CrossOver, despite licensing terms suggesting this should not be allowed.
It’s clear IPC shouldn’t be derived from one benchmark.Zen 4 IPC wasn't calculated from GB6, so no need to recalculate it.
Since M4 is the first ARM SoC with SME, I do think it will take some for real world apps to implement SME. I’m not aware of any application (apart from GB) right now that supports SME so we can’t test the uplift from M3 to M4 if an application uses SME.M4 with SME seems to have the biggest advantage in this sub test for the single core run, even more so than SPR with AMX.
No, what annoys us is Apple's walled garden. If they would let us run Windows and our current games library with proper DX12/Vulkan drivers, why would we be angry?
Fine by me. I know you're trying to make some kind of point about people being unfair to Apple, and maybe there is some of that going on; I think my posting history makes clear I'm not one of those - but there absolutely is value, a lot of it, in knowing what a core is likely to do with a random existing application that isn't implementing the buzzword of the week and isn't affected by specialized acceleration for it.
This applies to everyone - Intel, AMD, Apple, IBM, anyone else.
It’s clear IPC shouldn’t be derived from one benchmark.
Since M4 is the first ARM SoC with SME, I do think it will take some for real world apps to implement SME. I’m not aware of any application (apart from GB) right now that supports SME so we can’t test the uplift from M3 to M4 if an application uses SME.
Hopefully PyTorch. Mat lab etc support it soon
I may be mistaken, but IIRC Apple license said that one must not embed the game porting toolkit in a product, and/or that the toolkit was for development only.No, that's definitely allowed - all the LGPL and GPL components have their source available.
Oh I misunderstood you. In any case, Apple too can offer difference license terms to different people. We can't, but CodeWeavers can integrate it.I may be mistaken, but IIRC Apple license said that one must not embed the game porting toolkit in a product, and/or that the toolkit was for development only.
But the terms may have changed.
The licence disclosure lists DXVK because GPTK (as a whole) uses wine/crossover. D3DMETAL has nothing to do with that, as the engineer behind it plainly stated. Besides, DXVK is an DX11 to Vulkan translator, it doesn’t do DX12. Apple DID give codeweavers permission to bundle D3DMETAL with crossover though.License disclosure indicates they are in fact using some of DXVK's code. They could open source it to show to what extent they are using it but they won't because Apple.
Edit: It could be that it's entirely in the (L)GPL Wine components but since they bundle it all together who knows. But even still if shows that as usual Apple will build proprietary solutions on top of open ones and refuse to open that. At least in this case CodeWeavers got a nice paycheck (presumably) for doing the dirty work.
In closed form & in exchange for access to their patches over Wine. Meanwhile Mesa has Microsoft contributors and has for years. Apple's not really changing at all. Use wherever possible, give back as little as possible and avoid GPL like the plague. The other American megacorps at least do play the open source game for non-core components. And can you think of anything less core to Apple's success than keeping D3DMetal closed? But they do it anyway.Apple DID give codeweavers permission to bundle D3DMETAL with crossover though.
But it's supported on GB6, which is what these comparisons have been. See these posts below:AMX isn't supported on client CPUs at this time, which is where all the comparisons have been.
It’s seems AMX support is available in GB6.
Here is 12700T vs w5-3435X, Both based on Golden Cove, the latter having AMX: https://browser.geekbench.com/v6/cpu/compare/5330257?baseline=4656765
I didn't pick the fastest scores, but checked both are using the same GB version.
The effect of AMX can be seen on Object Detection. It also looks like Background Blur is affected.