4 ms·
What it does show is that CUDA leaves serious performance optimization on the table despite its gigantic code base. Using compression to reduce memory bandwidth
by nialse 2y ago
What it does show is that CUDA leaves serious performance optimization on the table despite its gigantic code base. Using compression to reduce memory bandwidth is a well known trick in quantization, and in other scenarios since forever. There has been little competitive pressure on Nvidia to go further since their software stack leaves the competition in the dust. This time, they may actually need to step up their efforts, due to customer pressure. Good times!
- mrbungie 2y agoI see almost no incentives for NVDA to do so as long as they have no real competitors in that space.
- frontalier 2y ago> I see almost no incentives for NVDA to do so > as long as they have no real competitors in that space. ?! isn't _avoiding competition_ a good enough incentive for them to improve their offering?
- kristjansson 2y agoA careful implementation adapted to a specific workload is always going to beat a more generic implementation.