- From: Felix Maier <xilefmai@gmail.com>
- Date: Fri, 15 May 2020 20:42:41 +0200
- To: Tianqi Chen <tqchen@cs.washington.edu>
- Cc: Corentin Wallez <cwallez@google.com>, public-gpu <public-gpu@w3.org>
- Message-ID: <CAL8VZWx1Gj6fP1wnhYqBNhne0S-rP=t8Vo9Lb0RHGOr_o7BO4g@mail.gmail.com>
As far as I know, only Vulkan offers Tensor Core acceleration for matrices (using VK_NV_cooperative_matrix <https://www.khronos.org/registry/vulkan/specs/1.2-extensions/man/html/VK_NV_cooperative_matrix.html>) and is limited to NVIDIA hardware. Felix On Fri, 15 May 2020 at 17:59, Tianqi Chen <tqchen@cs.washington.edu> wrote: > Thanks Corentin: > > subgroup(warp level semantics in CUDA) are useful to get the last mile on > certain GPUs, but not strictly necessary in many cases. > The most crucial part that WebGPU enabled(over WebGL) is the ability to > run compute shaders and effective use of shared memory(working-group > memory). > > To bring the compute to the next-level(e.g. making use of TensorCore in > nvidia GPUs when available), more extension would be needed, and we would > certainly be curious if that is under the scope of WebGPU. > > TQ > > On Fri, May 15, 2020 at 8:37 AM Corentin Wallez <cwallez@google.com> > wrote: > >> Hey Tianqi, >> >> Thanks for sharing! The blog post was very interesting and the results >> encouraging. I'm surprised it's this close even when WebGPU doesn't support >> float16 or subgroup operations yet. Impressive! >> >> The samples list in Implementation-Status is very out of date, there's >> been a lot of people starting to use WebGPU lately. Group, what do you >> think of having a "WebGPU users" wiki page to collect things beyond the >> most trivial samples? Apache TVM would fit in there. >> >> Cheers, >> >> Corentin >> >> On Thu, May 14, 2020 at 11:46 PM Tianqi Chen <tqchen@cs.washington.edu> >> wrote: >> >>> Hi WebGPU community: >>> >>> I am sending this along since I think it could be interesting to the >>> members who are also interested in machine learning. >>> >>> We recently introduced support for WASM and WebGPU to the Apache TVM >>> deep learning compiler. >>> Our pre-liminary experiments shows that TVM’s WebGPU backend can* get >>> close to native GPU performance* when deploying >>> deep learning models to the web. Please see the blog here. >>> >>> >>> https://tvm.apache.org/2020/05/14/compiling-machine-learning-to-webassembly-and-webgpu >>> >>> I am also wondering if it is possible to link to the blog as an example >>> of ML On WebGPU in the wiki page( >>> https://github.com/gpuweb/gpuweb/wiki/Implementation-Status). >>> As an open source community, we certainly love feedbacks and >>> collaborations >>> >>> Cheers >>> TQ >>> >>>
Received on Friday, 15 May 2020 18:43:47 UTC