3 ms·
For that to work, they need a software stack to compete with CUDA. Nvidia is so far ahead that a single manufacturer won’t be able to compete for developers.
by hedora 2y ago
For that to work, they need a software stack to compete with CUDA.
Nvidia is so far ahead that a single manufacturer won’t be able to compete for developers. Instead, the only chance the rest of the AI GPU market has is to build some portable open source thing and use it to gang up on Nvidia.
That means bigger GPU players (AMD, hyperscalers) will need to be involved.
- Wytwwww 2y ago> For that to work, they need a software stack to compete with CUDA. Doesn't Intel have pretty decent support in PyTorch? It's not like most people working on/with AI use CUDA directly. And especially for stuff like LLMs software is hardly the main or even a significant barrier.
- shrubble 2y agoI am not a PyTorch user, but there is already Intel-supplied ARC acceleration for PyTorch: https://www.intel.com/content/www/us/en/developer/articles/technical/stable-diffusion-with-intel-arc-gpus.html https://www.intel.com/content/www/us/en/developer/articles/t... Having half the number of GPUs in a workstation/local server setup to have same amount of VRAM might make up for whatever slowdown there would be if you had to use less-optimized code. For instance running or training a model that required 192GB of VRAM would take 4x48GB VRAM but 8x24GB VRAM GPUs.