These efforts to support CUDA on non-Nvidia hardware seem to me misguided. If all you want is to be able to easily use non-NVidia hardware then high level tools like PyTorch already let you do that (and torch.compile uses Triton for target-specific optimization). OTOH if you want to be programming close to the metal to achieve top performance then you are probably not using CUDA in the first place, and using some CUDA translation layer on non-NVidia hardware would be an even worse idea.
There's nothing wrong to run CUDA on non-Nvidia hardware. CUDA has an interface that is reasonably well-designed, well-documented/reverse-engineered, and battle-tested for decades. What we need is not to invent another interface just under the name of 'open standard', but to implement the same interface. ROCm is exactly doing this, and so are other hardware SDKs such as MooreThread and Alibaba T-Head.
Someone needs to stand up a benchmark suite for ROCM, this, and everyone else attempting it to really get the ball rolling here. SemiAnalysis could have a blast with this.
> Ease of programming and a giant leap in performance is one of the key reasons for the CUDA platform’s widespread adoption
This, so much. Other platforms continue to ignore developer UX, but it's one of the main things that get's new users onboard and keeps old users around.
Including stuff like Fortran, Haskell, Java, .NET via PTX, Python JIT, IDE tooling integration with major IDEs, graphical GPU debugging and profiling, libraries and co?
A guess would be some time next year — since our public launch our focus has generally been on API coverage and increasingly recently, on performance.
While performance improvements will always remain a target, we're soon at full coverage of the core CUDA APIs and will be shifting an increasing amount of effort towards developer tooling.
In this context AdaptiveCpp should also be mentioned. Started as a SYCL implementation, but recently-ish added a compiler for compiling a CUDA dialect to GPUs and CPUs from basically all vendors
SYCL is probably the most up-to-date CUDA alternative for all intents and purposes, at least if one likes modern C++ style (and lambdas inside lambdas). Expose it as C and get bindings to any other language for relatively little effort as well since it’s just C++. With AdaptiveCpp you can also compile SYCL to CUDA so both ways work with the CUDA dialect (PCUDA).
SYCL, as well as AdaptiveCpp, is a relatively active project though and has been for several years, feeding into the C++ standards committee work and is supported by several large organisations, including US national labs and several European universities. I suppose it’s worth keeping track of for people in related fields.
I suppose it’s just really hard to beat the head start and ecosystem integration NVIDIA has with CUDA.
Alternatives exist, but little demand outside hyperscalers and special uses.
Neocloud customers just want plug-and-play CUDA. It works, it's tested, it adapts faster, and has known performance. Alternatives give no significant benefits.
every CUDA alternative follows the same arc: bold launch, works for 3 operations, then a Discord server where the last message is 'any updates?' from 2024
i'm also interested in tenstorrent. they're building GPUs with cheap GDDR6 using a fast SRAM cache, and writing their own compiler stack (used instead of CUDA) that pipelines data to the SRAM ahead-of-time so you (in theory) never need to suffer the slow speed of GDDR6 for AI workloads. also they've got built-in SFP cages where the video ports would normally be.
Isn't the future of the industry specialized chips like those that Broadcom and Cerebras are making? I don't know how much longer I can tolerate 50 tokens per second. It feels like the dial-up era.
Weird, since the most used open source inference engine is faster on Vulkan on platforms that offer multiple options, with the sole exception being Nvidia, due to poor Nvidia driver quality (which I am forced to assume is intentional, Nvidia wishes to maintain their moat after all).
A couple of years ago I evaluated both Vulkan and Cuda as a choice for future projects. I couldnt get anything done after a week in Vulkan, but had the test prototype project working after just a day in Cuda.
Needless to say, I'd never ever pick Vulkan for any project after that experience. It's just way to needlessly overengineered and bloated.
I used to be big into Khronos API camp, even did my project thesis in OpenGL, up to the famous Long Peaks fail.
Vulkan ended up being the same extension spaghetti as its predecessor, and Khronos was only able to come up with something thanks to AMD offering Mantle, C++ bindings and a GLSL successor only came to be thanks to NVidia (Vulkan-hpp and Slang started at NVidia).
The "we build the specification", and then "the community builds the tools", leads to very poor experiences, and if it wasn't for LunarG own interests, there wouldn't even exist any kind of Vulkan SDK.
What they have going is naturally the vendor independence, however we can achieve the same with middleware with the benefit of much better developer experience.
I love how people say things like "extension spaghetti", as if all other non-standard APIs have the same problem: hardware gets new features that people want to use from that API, API gains extension to use that hardware feature.
CUDA is no different, in fact, often worse. Nvidia is bad at documenting which hardware does what things, and CUDA users often have to use third party tables to figure out what hardware can't do what and disappoint customers who unwisely invested into it.
They really don't, no. Vulkan: 50 lines to allocate device memory. Cuda: One single line. What kind of extensive documentation stack do you want for functionality that is trivial in Cuda? And that exact issue continues through every little step of the way to your first usable application. I know there is VMA, it is a very poor solution to a problem that shouldn't even exist, and it only poorly addresses one of 100 parts of the API where Cuda is vastly simpler than Vulkan. Cuda also doesnt force you to use queue families but you can optionally use streams. No ridiculous descriptor management and binding in cuda, just passing pointers and handles via launch arguments. No overengineered explicit syncing mechanis in cuda, everything is nicely implicitly synced until you explicitly opt in to parallel streams. etc.
These efforts to support CUDA on non-Nvidia hardware seem to me misguided. If all you want is to be able to easily use non-NVidia hardware then high level tools like PyTorch already let you do that (and torch.compile uses Triton for target-specific optimization). OTOH if you want to be programming close to the metal to achieve top performance then you are probably not using CUDA in the first place, and using some CUDA translation layer on non-NVidia hardware would be an even worse idea.
There's nothing wrong to run CUDA on non-Nvidia hardware. CUDA has an interface that is reasonably well-designed, well-documented/reverse-engineered, and battle-tested for decades. What we need is not to invent another interface just under the name of 'open standard', but to implement the same interface. ROCm is exactly doing this, and so are other hardware SDKs such as MooreThread and Alibaba T-Head.
Someone needs to stand up a benchmark suite for ROCM, this, and everyone else attempting it to really get the ball rolling here. SemiAnalysis could have a blast with this.
Most of these "alternatives" focus on CUDA C++, and overlook what actually makes CUDA interesting.
Already in 2020,
https://developer.nvidia.com/blog/cuda-refresher-the-gpu-com...
> Ease of programming and a giant leap in performance is one of the key reasons for the CUDA platform’s widespread adoption
This, so much. Other platforms continue to ignore developer UX, but it's one of the main things that get's new users onboard and keeps old users around.
We're actually targeting all of it, and not just CUDA C++.
Including stuff like Fortran, Haskell, Java, .NET via PTX, Python JIT, IDE tooling integration with major IDEs, graphical GPU debugging and profiling, libraries and co?
Then I guess all the best.
This post has some serious peanut-gallery vibes.
Peanut-gallery is happily using CUDA, and needs actual sound reasons to move.
Then the peanut gallery has nothing to complain when Nvidia jacks up prices.
Do you see me complaining?
Here is a tip, you don't always need to suffer from FOMO and get the very latest model card.
In fact, contrary to the competition, one can play with CUDA even on laptops, go figure.
Ambitious but neat, good luck if nothing else :)
If you were to guess, when do you think your Nsight Compute alternative might be ready with your own toolchain?
A guess would be some time next year — since our public launch our focus has generally been on API coverage and increasingly recently, on performance.
While performance improvements will always remain a target, we're soon at full coverage of the core CUDA APIs and will be shifting an increasing amount of effort towards developer tooling.
In this context AdaptiveCpp should also be mentioned. Started as a SYCL implementation, but recently-ish added a compiler for compiling a CUDA dialect to GPUs and CPUs from basically all vendors
SYCL is probably the most up-to-date CUDA alternative for all intents and purposes, at least if one likes modern C++ style (and lambdas inside lambdas). Expose it as C and get bindings to any other language for relatively little effort as well since it’s just C++. With AdaptiveCpp you can also compile SYCL to CUDA so both ways work with the CUDA dialect (PCUDA).
SYCL, as well as AdaptiveCpp, is a relatively active project though and has been for several years, feeding into the C++ standards committee work and is supported by several large organisations, including US national labs and several European universities. I suppose it’s worth keeping track of for people in related fields.
I suppose it’s just really hard to beat the head start and ecosystem integration NVIDIA has with CUDA.
Alternatives exist, but little demand outside hyperscalers and special uses.
Neocloud customers just want plug-and-play CUDA. It works, it's tested, it adapts faster, and has known performance. Alternatives give no significant benefits.
Things can change, but they are not changing now.
every CUDA alternative follows the same arc: bold launch, works for 3 operations, then a Discord server where the last message is 'any updates?' from 2024
Actually we launched in 2024 and the last message in our discord is definitely not that: https://discord.gg/KNpgGbTc38
There is also ZLUDA, which is open source and works on pre-compiled binaries.
https://github.com/vosen/ZLUDA
this is closest thing we have to "cuda on non-nvidia" hardware
We have a comparison page: https://docs.scale-lang.com/stable/manual/comparison/#zluda
i'm also interested in tenstorrent. they're building GPUs with cheap GDDR6 using a fast SRAM cache, and writing their own compiler stack (used instead of CUDA) that pipelines data to the SRAM ahead-of-time so you (in theory) never need to suffer the slow speed of GDDR6 for AI workloads. also they've got built-in SFP cages where the video ports would normally be.
Is tenstorrent building GPUs now, not just tensor processors?
I have been trying for cuda -> metal, to run it on mac, https://github.com/lulzx/cuda-metal
Why should I not just port my kernel to Triton? What's the appeal of Scale?
You can skip the porting part.
aren't llms smart enough to directly write custom kernels for custom hardware from cuda code?
Isn't the future of the industry specialized chips like those that Broadcom and Cerebras are making? I don't know how much longer I can tolerate 50 tokens per second. It feels like the dial-up era.
@claude add this to the graveyard of wannabees
Its easier to just get rid of your legacy code entirely and use Vulkan for compute, or have your compiler emit SPIR-V directly.
No reason to tie yourself to Nvidia's moat.
Unfortunately, Vulkan Compute doesn’t to all the things that OpenCL, SYCL, HIP or CUDA do.
Yep, there are inference stacks where it just does not work without cuda in any meaningful performance
Weird, since the most used open source inference engine is faster on Vulkan on platforms that offer multiple options, with the sole exception being Nvidia, due to poor Nvidia driver quality (which I am forced to assume is intentional, Nvidia wishes to maintain their moat after all).
There's nothing stopping any of us from writing a better Nvidia driver btw. LLMs are very helpful with reverse engineering.
A couple of years ago I evaluated both Vulkan and Cuda as a choice for future projects. I couldnt get anything done after a week in Vulkan, but had the test prototype project working after just a day in Cuda.
Needless to say, I'd never ever pick Vulkan for any project after that experience. It's just way to needlessly overengineered and bloated.
I used to be big into Khronos API camp, even did my project thesis in OpenGL, up to the famous Long Peaks fail.
Vulkan ended up being the same extension spaghetti as its predecessor, and Khronos was only able to come up with something thanks to AMD offering Mantle, C++ bindings and a GLSL successor only came to be thanks to NVidia (Vulkan-hpp and Slang started at NVidia).
The "we build the specification", and then "the community builds the tools", leads to very poor experiences, and if it wasn't for LunarG own interests, there wouldn't even exist any kind of Vulkan SDK.
What they have going is naturally the vendor independence, however we can achieve the same with middleware with the benefit of much better developer experience.
I love how people say things like "extension spaghetti", as if all other non-standard APIs have the same problem: hardware gets new features that people want to use from that API, API gains extension to use that hardware feature.
CUDA is no different, in fact, often worse. Nvidia is bad at documenting which hardware does what things, and CUDA users often have to use third party tables to figure out what hardware can't do what and disappoint customers who unwisely invested into it.
The other platforms have better ways to deal with progress instead of "here find entries on dynamic libraries by yourself", and good luck.
Profiles and API versions are much better approaches.
It is no accident than the ongoing efforts to make Vulkan more friendly are moving away from extension spaghetti into profiles.
Weird, most people have the exact opposite experience.
Having to deal with closed source opaque poorly documented stacks sucks.
They really don't, no. Vulkan: 50 lines to allocate device memory. Cuda: One single line. What kind of extensive documentation stack do you want for functionality that is trivial in Cuda? And that exact issue continues through every little step of the way to your first usable application. I know there is VMA, it is a very poor solution to a problem that shouldn't even exist, and it only poorly addresses one of 100 parts of the API where Cuda is vastly simpler than Vulkan. Cuda also doesnt force you to use queue families but you can optionally use streams. No ridiculous descriptor management and binding in cuda, just passing pointers and handles via launch arguments. No overengineered explicit syncing mechanis in cuda, everything is nicely implicitly synced until you explicitly opt in to parallel streams. etc.
Vulkan tooling is light years behind what CUDA offers in 2026, across programming languages, IDE tooling, graphical debuggers and libraries.
Ports are very often incredibly difficult and very time consuming.
One of the biggest complaints we hear from the industry is "we tried to port to X and we could never complete it".
An established codebase can have years of refinement. It will take time to achieve the same with the port.
And with our compiler, just using cuda is no longer putting urself inside the moat :)
Ironically, this is what people claim AI can do with a snap of the fingers.
Should be real simple if the HN AI echochamber is right, right?