3 ms·

To address the article, this will be very interesting because in the market Tensortorrent is targetting the real challenge is software support for PyTorch, Tens

by belval 3y ago

To address the article, this will be very interesting because in the market Tensortorrent is targetting the real challenge is software support for PyTorch, TensorFlow, distributed training, etc... and less about the hardware itself. It's not like AMD GPUs can't do matrix multiplications at competitive TFLOPS but instead that if I wanted to use them to run a model I'd have to work through several workarounds to get similar performance to an Nvidia GPU that is mostly plug-and-play. We will see if Jim Keller's many talents include managing software.

cma 3y ago
Does AMD have similar tensor core stuff?
- kimixa 3y ago
  They now have Xilinx's AI cores since they merged (acquired?). https://www.xilinx.com/products/technology/ai-engine.html https://www.xilinx.com/products/technology/ai-engine.html seems to be a high-level overview. I think their latest APUs for laptops have this IP integrated. And they're blurring the line by adding more AI-focused acceleration instructions to their GPUs (maybe focused on their CDNA line rather than the consumer RDNA, but they're expanding their capabilities on both). There seems to be a bit of a philosophy difference between NVidia and AMD acceleration - NVidia add new specialised "cores", while AMD adds those similar capabilities as accelerated instructions to their shader units. Though this might just be marketing - I don't think we've got an open reference for the NVidia shader ISA like we do with AMD so difficult to compare, but the capabilities and integration in shaders of many of these acceleration paths seems to imply they might also be implemented as shader instructions, which stretches the definition of a separate "core" to me, and so might just be a marketing difference rather than anything technical. Same with things like the BVH acceleration used in ray tracing - AMD added shader instructions, while NVidia talks loudly about "RT Cores".
- belval 3y ago
  Yes that's my point. AMD has the hardware, but using that hardware with the existing DL ecosystem is much more painful and they haven't been able to bridge that gap.

by belval 3y ago

cma 3y ago

Does AMD have similar tensor core stuff?

kimixa 3y ago
They now have Xilinx's AI cores since they merged (acquired?). https://www.xilinx.com/products/technology/ai-engine.html https://www.xilinx.com/products/technology/ai-engine.html seems to be a high-level overview. I think their latest APUs for laptops have this IP integrated. And they're blurring the line by adding more AI-focused acceleration instructions to their GPUs (maybe focused on their CDNA line rather than the consumer RDNA, but they're expanding their capabilities on both). There seems to be a bit of a philosophy difference between NVidia and AMD acceleration - NVidia add new specialised "cores", while AMD adds those similar capabilities as accelerated instructions to their shader units. Though this might just be marketing - I don't think we've got an open reference for the NVidia shader ISA like we do with AMD so difficult to compare, but the capabilities and integration in shaders of many of these acceleration paths seems to imply they might also be implemented as shader instructions, which stretches the definition of a separate "core" to me, and so might just be a marketing difference rather than anything technical. Same with things like the BVH acceleration used in ray tracing - AMD added shader instructions, while NVidia talks loudly about "RT Cores".
belval 3y ago
Yes that's my point. AMD has the hardware, but using that hardware with the existing DL ecosystem is much more painful and they haven't been able to bridge that gap.