3 ms·
Why not. If you have an accelerator you have specific hardware for the feature. The is really no fundamental reason it can't be as good as a GPU, if GPU is the
by kevincox 3y ago
Why not. If you have an accelerator you have specific hardware for the feature. The is really no fundamental reason it can't be as good as a GPU, if GPU is the best architecture it would just be a GPU.
The main question is did they allocate enough space and power budget to the accelerator. Or is it anemic even for basic tasks.
- VHRanger 3y ago> The is really no fundamental reason it can't be as good as a GPU, if GPU is the best architecture it would just be a GPU. There is a reason, and that reason is called memory bandwidth.
- smolder 3y agoFor one thing, the memory on GPUs is generally much faster, with a wider bus than what typical CPUs have. The die sizes can be quite large, too. Putting a full sized CPU in the same package as the accelerator means making some sacrifices.
- hot_gril 3y agoGPUs are huge, TPUs etc are also quite large, and CPUs are tiny. I'm no expert on any of these, but intuitively you're losing something cramming that functionality into a way smaller chip. Probably something to do with bandwidth (where big helps) vs latency (where small helps).
- coffeebeqn 3y agoThe GPU chip itself isn’t that big if you take the heat sink off. But it has power delivery and memory etc on the board
- hot_gril 3y agoEven the GPU chip itself is still much larger than the CPU chip.
- cyanydeez 3y agosoftware is the fundamental component of AI. ITS basically this month the And accelerators are supported by llm
- coffeebeqn 3y agoHow useful is an integrated GPU at any AI workloads? People seem to generally use things like 3090 or 4090