Speculation: Radeon will become the "x86" of GPUs

[tamz_msc]

AMDs GPU IP is not open source.

This thread is about whether Intel will license and adopt AMDs Radeon IP, which is proprietary.

It is a virtual certainty that Intel will NOT be adopting AMDs IP, but will be pursuing their own unique designs/ISA for their in house GPUs.

GPU ISA is different from x86 ISA.

For graphics HLSL(DirectX) and GLSL(OpenGL) are mapped to the GPU's(NVIDIA/AMD) ISA via the drivers. Forget graphics for a moment and focus on compute. Given how CUDA has entrenched itself in GPGPU during the last 10 years, it'll be suicide for Intel's GPU to not be able to compile CUDA. Meaning that Intel supporting OpenCL only would be counter to their intentions to enter the GPU space in the first place.

If that is indeed the case, that is Intel doesn't support CUDA, then NVIDIA has very little to worry about.

AMD doesn't figure anywhere in this, and even if Intel doesn't support OpenCL at the moment to the extent AMD does(which is unlikely), wasting resources in bringing their GPUs on par with it and taking it to the fight against CUDA would be a losing battle.

If anything, Intel should be focusing more on getting their GPUs to support CUDA instead.

 

[24601]

GPU ISA is different from x86 ISA.

For graphics HLSL(DirectX) and GLSL(OpenGL) are mapped to the GPU's(NVIDIA/AMD) ISA via the drivers. Forget graphics for a moment and focus on compute. Given how CUDA has entrenched itself in GPGPU during the last 10 years, it'll be suicide for Intel's GPU to not be able to compile CUDA. Meaning that Intel supporting OpenCL only would be counter to their intentions to enter the GPU space in the first place.

If that is indeed the case, that is Intel doesn't support CUDA, then NVIDIA has very little to worry about.

AMD doesn't figure anywhere in this, and even if Intel doesn't support OpenCL at the moment to the extent AMD does(which is unlikely), wasting resources in bringing their GPUs on par with it and taking it to the fight against CUDA would be a losing battle.

If anything, Intel should be focusing more on getting their GPUs to support CUDA instead.

CUDA is updated for every generation of GPU along with new functions specifically for the new generation, etc.

Supporting CUDA as anyone (GPU hardware vendor) outside of Nvidia is simply not feasible.

The entire reason why people love CUDA is due to it mapping almost directly to Nvidia's hardware/architecture of each specific generation of cards.

 

[prtskg]

CUDA is updated for every generation of GPU along with new functions specifically for the new generation, etc.

Supporting CUDA as anyone (GPU hardware vendor) outside of Nvidia is simply not feasible.

The entire reason why people love CUDA is due to it mapping almost directly to Nvidia's hardware/architecture of each specific generation of cards.

Indeed. Gunning for CUDA will not benefit AMD or Intel, IMO. They should go the OpenCl route and ensure every programs work as intended.

 

[tamz_msc]

CUDA is updated for every generation of GPU along with new functions specifically for the new generation, etc.

Supporting CUDA as anyone (GPU hardware vendor) outside of Nvidia is simply not feasible.

The entire reason why people love CUDA is due to it mapping almost directly to Nvidia's hardware/architecture of each specific generation of cards.

I'm aware of that - what Intel should do is to provide tools to migrate CUDA to make it run on their hardware. Like AMD's HIP but with at least 10x the developer effort than what AMD can spare in catching up with every iteration of CUDA.

Intel thinks it can turn an industry whose ships already set sail 10+ years ago towards it's direction. Pipe dream, IMO.

 

[tamz_msc]

Indeed. Gunning for CUDA will not benefit AMD or Intel, IMO. They should go the OpenCl route and ensure every programs work as intended.

It's not about making every program work - it's about ensuring that researchers who use GPGPU spend more time doing actual research than figuring out ways to run their code on unfamiliar hardware and software.

NVIDIA goes out of their way to make sure that the latter doesn't happen.

 

[senseamp]

The quicker Intel and AMD get away from x86 mentality, the better off they will be. The reality is they got lucky with x86, where through a very unique series of events Intel ended up with a government granted monopoly on the foundational ISA of a paradigm computing called PC, and AMD got lucky to get in on the action as a second supplier. One should not confuse luck with skill.

 

[urvile]

with the obvious end goal of turning the GPU into the central focal point of the system. Don't believe me? Check out Huang's speech

Of course he said that he is the CEO of the company. Sounds like hyperbole....

 

[Vattila]

ROCm compute stack on mainline Linux 4.17

"Long story short, it's looking like all the pieces are falling into place where with the Linux 4.17 kernel cycle in a few months, the ROCm/OpenCL user-space could finally be playing nicely on the mainline kernel! That is with the discrete GPUs where as up until recently they were more focused on getting all the APU bits upstreamed first. It's been a long time coming and I can't wait for this milestone so the open-source ROCm code can be more widely deployed and easier to use with no longer needing out-of-tree kernel modules"

https://www.phoronix.com/scan.php?page=news_item&px=AMDKFD-dGPU-GPUVM

AMD's open-source alternative to CUDA is progressing. Will Intel support it?

 

[prtskg]

ROCm compute stack on mainline Linux 4.17

"Long story short, it's looking like all the pieces are falling into place where with the Linux 4.17 kernel cycle in a few months, the ROCm/OpenCL user-space could finally be playing nicely on the mainline kernel! That is with the discrete GPUs where as up until recently they were more focused on getting all the APU bits upstreamed first. It's been a long time coming and I can't wait for this milestone so the open-source ROCm code can be more widely deployed and easier to use with no longer needing out-of-tree kernel modules"

https://www.phoronix.com/scan.php?page=news_item&px=AMDKFD-dGPU-GPUVM

AMD's open-source alternative to CUDA is progressing. Will Intel support it?

Intel contributes signifanctly in open source. Though they do it for their own benefit, vulkan coding is an example where it has helped others. I hope something like that happens in OpenCL too.