Good luck with that as intel pretty spent year with arc justing getting the drivers up to snuff
Gamebird8
Is it only targeting Mobile GPUs hence why they aren’t trying to compete with AMD?
No, it’s just that OP added to the title of the article
imaginary_num6er
>From friends to foe: ARM is rumored to be developing a gaming graphics card competing with NVIDIA
Top 10 anime betrayals
Comfortable-Exit8924
intel still making Arc drivers compatible for half life 2 so yea , DOA
AejiGamez
AMD might finally lose the title of „worst GPU drivers“ Albeit its just a myth with AMD nowadays, these ARM drivers are gonna be ass at the beginning
nicky94
Excellent news! more competition the better.
erebuxy
> and Intel
lol
zmunky
lol even if it reaches AMD levels of performance, Nvidia is literally coasting with new and fast gpus to whip out since they are so far ahead.
particlemanwavegirl
wow TIL ARM is actually a company that makes chips, I had kind of just assumed it was more like a professional standards organization that just defined the spec for other manufacturers to follow.
SailorMint
From a Duopoly, to a Triumvirate to a Quartel!
WeakDiaphragm
This is good. Nvidia needs competition. I give intel 3 more generations to be competing with AMD. And I hope ARM will need less than 5 generations to compete with Intel
Hattix
ARM’s coming from the same place Intel and Qualcomm are, and it’s not easy to do.
Intel’s Alchemist architecture is the easiest place to start. It’s a derivation of Intel Gen10/11 (and Tiger Lake’s XE) and so its entire memory heirarchy is designed as being a client of a L3$ it doesn’t control itself. While it was reworked to become a client of its own memory controllers and their L2$, internal organisation remains something designed for a small, low power GPU.
It makes perfect sense for Intel’s IGPs to be as low power as possible, if your package power limit is 120 watts and the GPU takes 40 watts of that, the CPU’s now got to somehow fit in 80 watts. Meanwhile, even low end junk like the RTX 4060 has 120 watts all to itself!
So Alchemist retains the “EU” architecture (renamed to “Vector Engine”), where each EU has eight FP32 lanes which was developed for Ivy Bridge’s Gen7 (though in a 2×4 FP32 config). The smallest unit of EUs, comparable to Nvidia’s SM or AMD’s WGP, is a “subslice”, made of two vector engines. An “Xe Core” contains 16 EUs and 8 FP32 lanes per EU as we’ve already seen. Ampere has four streaming multiprocessors as its most basic group, each with a pair of 16x FP32s per SMSP. RDNA2 is organised into WGPs, each with four SIMDs and 32x FP32s per SIMD. Already we see Ampere and RDNA2 are *much* bigger in their basic execution engines than Alchemist is.
This really tiny arrangement means Alchemist has problems wiring it all up [and it isn’t able](https://chipsandcheese.com/2022/10/20/microbenchmarking-intels-arc-a770/) to saturate its memory bandwidth very easily at all. RDNA 2 in the 6700XT can absolutely saturate its entire bandwidth with just 6 workgroups dispatched per clock. Ampere (and some Ada, Ada is a different beast and scales down very poorly) can reach peak utilisation between 10 and 16 workgroups dispatched. Alchemist peaks at 16, then drops between 17 and 31, then peaks again at 32 – Its memory heirarchy is a mess, a direct result of being designed to be small and low power.
These are the challenges ARM’s architecture will have facing it. ARM Valhall (e.g. in Mali G710) is like a shrunk down Alchemist. Like Alchemist, its basic design was made assuming it was a client of a large last-level cache (LLC) and it had to save power wherever it could and it’s designed for configurability, so L2$ in Valhall is a client of LLC and closely coupled to shader cores. Valhall’s shader cores are dual-issue (similar to RDNA’s WGPs) with 16-wide FP32. At this kind of level it looks weirdly similar to AMD’s CDNA architecture, if AMD’s compute units didn’t actually exist and each SIMD was just sitting there exposed. Oh, and it had hardly any of them.
ARM has its work cut out. It isn’t known for very good shader compilers (Intel was far better than ARM, and still had a lot of work to do with Alchemist) but that’s the crux of modern DX12 and Vulkan performance.
stepping_
i can see them doing much better than intel, but all i can see them doing for the next 7 years is competing well with AMD and barely putting a dent on nvidias profits. at least in the price to performance department we are gonna have good options.
BrotherMichigan
Ah yes, the two large gaming GPU players, NVIDIA and Intel.
CactusDoesStuff
ARM is one of the things that I’m most excited about. hypes me up thinking that we’ll have super efficient CPUs that are fully compatible with Windows soon enough
stormdraggy
> compete against Nvidia and intel ~~and AMD~~
Lmao
JuiceLittle6160
Have they even designed a GPU over 1 tflops yet?
Wolfgod_Holo
but will this threaten Jensen’s leather jacket money?
H0vis
It’ll take years but there’s a huge gap in the market for something at what used to be the mid-range but is now budget. Nvidia’s greed has left a big gap for new players, and I wish them all well.
IndexStarts
I guess Intel is more noteworthy than AMD in the GPU market according to the article’s title lmao
Taterthotuwu91
Arm is prob more interesting for laptops and consoles than desktops 🤔
ImSo_Bck
Or rather, that it doesn’t need to?😂
lokisbane
I read this as AMD several times and was about to ask if everyone is just joking and being sarcastic about AMD. Lol
corgiperson
I guess I’ve never really thought about this but what instruction set does a GPU use? Would ARM be making an ARM GPU for potential gamers to slot in?
25 Comments
Good luck with that as intel pretty spent year with arc justing getting the drivers up to snuff
Is it only targeting Mobile GPUs hence why they aren’t trying to compete with AMD?
No, it’s just that OP added to the title of the article
>From friends to foe: ARM is rumored to be developing a gaming graphics card competing with NVIDIA
Top 10 anime betrayals
intel still making Arc drivers compatible for half life 2 so yea , DOA
AMD might finally lose the title of „worst GPU drivers“ Albeit its just a myth with AMD nowadays, these ARM drivers are gonna be ass at the beginning
Excellent news! more competition the better.
> and Intel
lol
lol even if it reaches AMD levels of performance, Nvidia is literally coasting with new and fast gpus to whip out since they are so far ahead.
wow TIL ARM is actually a company that makes chips, I had kind of just assumed it was more like a professional standards organization that just defined the spec for other manufacturers to follow.
From a Duopoly, to a Triumvirate to a Quartel!
This is good. Nvidia needs competition. I give intel 3 more generations to be competing with AMD. And I hope ARM will need less than 5 generations to compete with Intel
ARM’s coming from the same place Intel and Qualcomm are, and it’s not easy to do.
Intel’s Alchemist architecture is the easiest place to start. It’s a derivation of Intel Gen10/11 (and Tiger Lake’s XE) and so its entire memory heirarchy is designed as being a client of a L3$ it doesn’t control itself. While it was reworked to become a client of its own memory controllers and their L2$, internal organisation remains something designed for a small, low power GPU.
It makes perfect sense for Intel’s IGPs to be as low power as possible, if your package power limit is 120 watts and the GPU takes 40 watts of that, the CPU’s now got to somehow fit in 80 watts. Meanwhile, even low end junk like the RTX 4060 has 120 watts all to itself!
So Alchemist retains the “EU” architecture (renamed to “Vector Engine”), where each EU has eight FP32 lanes which was developed for Ivy Bridge’s Gen7 (though in a 2×4 FP32 config). The smallest unit of EUs, comparable to Nvidia’s SM or AMD’s WGP, is a “subslice”, made of two vector engines. An “Xe Core” contains 16 EUs and 8 FP32 lanes per EU as we’ve already seen. Ampere has four streaming multiprocessors as its most basic group, each with a pair of 16x FP32s per SMSP. RDNA2 is organised into WGPs, each with four SIMDs and 32x FP32s per SIMD. Already we see Ampere and RDNA2 are *much* bigger in their basic execution engines than Alchemist is.
This really tiny arrangement means Alchemist has problems wiring it all up [and it isn’t able](https://chipsandcheese.com/2022/10/20/microbenchmarking-intels-arc-a770/) to saturate its memory bandwidth very easily at all. RDNA 2 in the 6700XT can absolutely saturate its entire bandwidth with just 6 workgroups dispatched per clock. Ampere (and some Ada, Ada is a different beast and scales down very poorly) can reach peak utilisation between 10 and 16 workgroups dispatched. Alchemist peaks at 16, then drops between 17 and 31, then peaks again at 32 – Its memory heirarchy is a mess, a direct result of being designed to be small and low power.
These are the challenges ARM’s architecture will have facing it. ARM Valhall (e.g. in Mali G710) is like a shrunk down Alchemist. Like Alchemist, its basic design was made assuming it was a client of a large last-level cache (LLC) and it had to save power wherever it could and it’s designed for configurability, so L2$ in Valhall is a client of LLC and closely coupled to shader cores. Valhall’s shader cores are dual-issue (similar to RDNA’s WGPs) with 16-wide FP32. At this kind of level it looks weirdly similar to AMD’s CDNA architecture, if AMD’s compute units didn’t actually exist and each SIMD was just sitting there exposed. Oh, and it had hardly any of them.
ARM has its work cut out. It isn’t known for very good shader compilers (Intel was far better than ARM, and still had a lot of work to do with Alchemist) but that’s the crux of modern DX12 and Vulkan performance.
i can see them doing much better than intel, but all i can see them doing for the next 7 years is competing well with AMD and barely putting a dent on nvidias profits. at least in the price to performance department we are gonna have good options.
Ah yes, the two large gaming GPU players, NVIDIA and Intel.
ARM is one of the things that I’m most excited about. hypes me up thinking that we’ll have super efficient CPUs that are fully compatible with Windows soon enough
> compete against Nvidia and intel ~~and AMD~~
Lmao
Have they even designed a GPU over 1 tflops yet?
but will this threaten Jensen’s leather jacket money?
It’ll take years but there’s a huge gap in the market for something at what used to be the mid-range but is now budget. Nvidia’s greed has left a big gap for new players, and I wish them all well.
I guess Intel is more noteworthy than AMD in the GPU market according to the article’s title lmao
Arm is prob more interesting for laptops and consoles than desktops 🤔
Or rather, that it doesn’t need to?😂
I read this as AMD several times and was about to ask if everyone is just joking and being sarcastic about AMD. Lol
I guess I’ve never really thought about this but what instruction set does a GPU use? Would ARM be making an ARM GPU for potential gamers to slot in?
I guess ARM can’t compete with AMD. Okay.