4 ms·

Well, Nvidia seems to be claiming in the article that this is everything, not just graphics drivers: "NVIDIA GPUs share a common driver architecture and capabil

Well, Nvidia seems to be claiming in the article that this is everything, not just graphics drivers: "NVIDIA GPUs share a common driver architecture and capability set. The same driver for your desktop or laptop runs the world’s most advanced AI workloads in the cloud. It’s been incredibly important to us that we get it just right." And For cutting-edge platforms such as NVIDIA Grace Hopper or NVIDIA Blackwell, you must use the open-source GPU kernel modules. The proprietary drivers are unsupported on these platforms. (These are two most advanced NVIDIA architectures currently)

Animats 2y ago
That's interesting. I've been expecting the AI cards to diverge more from the graphics cards. AI doesn't need triangle fill, Z-buffering, HDMI out, etc. 16 bit 4x4 multiply/add units are probably enough. What's going on in that area?
- p_l 2y ago
  TL;DR - there seems to be not that much improvement from dropping the "graphics-only" parts of the chip if you already have a GPU instead of breaking into AI market as your first product. 1. nVidia compute dominance is not due to hyperfocus on AI (that's Google's TPU for you, or things like intel's NPU in Meteor Lake), but because CUDA offers considerable general purpose compute. In fact, considerable revenue came and still comes from non-AI compute. This also means that if you figure out a novel mechanism for AI that isn't based around 4x4 matrix addition, or which mixes it with various other operations, you can do them inline. This also includes any pre and post processing you might want to do on the data. 2. The whole advantage they have in software ecosystem builds upon their PTX assembly. Having it compile to CPU and only implement the specific variant of one or two instructions that map to "tensor cores" would be pretty much nonsensical (especially given that AI is not the only market they target with tensor cores - DSP for example is another). Additionally, a huge part of why nvidia built such a strong ecosystem is that you could take cheapest G80-based card and just start learning CUDA. Only some highest-end features are limited to most expensive cards, like RDMA and NVMe integration. Compare this with AMD, where for many purposes only the most expensive compute-only cards are really supported. Or specialized AI only chips that are often programmable either in very low-level way or essentially as "set a graph of large-scale matrix operations that are limited subset of operations exposed by Torch/Tensorflow" (Google TPU, Intel Meteor Lake NPU, etc). 3. CUDA literally began with how evolution of shader model led to general purpose "shader processor" instead of specialized vector and pixel processors. The space taken by specialized hardware for graphics that isn't also usable for general purpose compute is pretty minimal, although some of it is omitted, AFAIK, in compute only cards. In fact, some of the "graphics only" things like Z-buffering are done by the same logic that is used for compute (with limited amount of operations done by fixed-function ROP block), and certain fixed-function graphical components like texture mapping units are also used for high-performance array access. 4. Simplified manufacturing and logistics - nVidia uses essentially the same chips in most compute and graphics cards, possibly with minor changes achieved by changing chicken bits to route pins to different functions (as you mentioned, you don't need DP-outs of RTX4090 on an L40 card, but you can probably reuse the SERDES units to run NVLink on the same pins).

Animats 2y ago

That's interesting. I've been expecting the AI cards to diverge more from the graphics cards. AI doesn't need triangle fill, Z-buffering, HDMI out, etc. 16 bit 4x4 multiply/add units are probably enough. What's going on in that area?

p_l 2y ago
TL;DR - there seems to be not that much improvement from dropping the "graphics-only" parts of the chip if you already have a GPU instead of breaking into AI market as your first product. 1. nVidia compute dominance is not due to hyperfocus on AI (that's Google's TPU for you, or things like intel's NPU in Meteor Lake), but because CUDA offers considerable general purpose compute. In fact, considerable revenue came and still comes from non-AI compute. This also means that if you figure out a novel mechanism for AI that isn't based around 4x4 matrix addition, or which mixes it with various other operations, you can do them inline. This also includes any pre and post processing you might want to do on the data. 2. The whole advantage they have in software ecosystem builds upon their PTX assembly. Having it compile to CPU and only implement the specific variant of one or two instructions that map to "tensor cores" would be pretty much nonsensical (especially given that AI is not the only market they target with tensor cores - DSP for example is another). Additionally, a huge part of why nvidia built such a strong ecosystem is that you could take cheapest G80-based card and just start learning CUDA. Only some highest-end features are limited to most expensive cards, like RDMA and NVMe integration. Compare this with AMD, where for many purposes only the most expensive compute-only cards are really supported. Or specialized AI only chips that are often programmable either in very low-level way or essentially as "set a graph of large-scale matrix operations that are limited subset of operations exposed by Torch/Tensorflow" (Google TPU, Intel Meteor Lake NPU, etc). 3. CUDA literally began with how evolution of shader model led to general purpose "shader processor" instead of specialized vector and pixel processors. The space taken by specialized hardware for graphics that isn't also usable for general purpose compute is pretty minimal, although some of it is omitted, AFAIK, in compute only cards. In fact, some of the "graphics only" things like Z-buffering are done by the same logic that is used for compute (with limited amount of operations done by fixed-function ROP block), and certain fixed-function graphical components like texture mapping units are also used for high-performance array access. 4. Simplified manufacturing and logistics - nVidia uses essentially the same chips in most compute and graphics cards, possibly with minor changes achieved by changing chicken bits to route pins to different functions (as you mentioned, you don't need DP-outs of RTX4090 on an L40 card, but you can probably reuse the SERDES units to run NVLink on the same pins).