Because they don't have a chip, their team is all over the place, they claim 1 of their chip is equal to 3 RTX 5090 in path tracing and FP64, has more than 2TB of VRAM, etc.
If a team is claiming these things, at least have employees who worked at Nvidia or AMD or Apple.
That's what I am thinking as well. They might have something great in the works, but, unless they can make some good on those promises, it's still vaporware.
The focus on HPC and gaming is also odd - HPC is close enough to AI training in that you can pass a single design as good on both - as it makes no sense focusing on one exclusively. Gaming is the odd one - unless they expect to generate large scale demand for their stuff. For that, however, they don't seem to have done the homework with reaching out to gaming studios.
So, we have this:
- Their hardware is good for HPC, but not great for AI. Maybe they got rid of everything that can process anything less than float32.
- There is a lot of connectivity, which can put a lot of memory ("2TB") within easy reach of a lot of compute units. This is good for HPC, and would be good for AI as well.
- Gaming is a completely different animal and shedding all the gaming-specific silicon would be beneficial for HPC (and AI)
These three put together kind of imply either more than one design, or that all the GPU part is a software implementation on top of the rest, and that there is zero hardware optimization for those tasks.
When nothing makes sense, it's usually because we are missing something.
If a team is claiming these things, at least have employees who worked at Nvidia or AMD or Apple.
Just outrageous claims.