Can we have a NPU thread
Depends, can we call them nipples/nipps?
>>109678247only if you call GPUs gipples and CPUs cipples
>>109678252What about ram, diddles?Trump ddr's little kids with the Jews.
>>109678235What good does it do? There;s only so many deep-UV chip fab lines, the big players can afford to buy all their capacity for years to come. Transformers and diffusers need matrix multiplies, can you build a better, faster large-scale matrix multiply unit than nvidia, amd, or intel? Can you do it cheaper than them? More reliably? Can you actually deliver on your solution, or will it be stillborn due to supply or yield issues?Etc etc etc... it's been tried before.
Did someone say nipples?
It's great for running recall
>>109678235They are a bit pointless on a desktop workstation or server because they're made for power efficiency only. In theory they would be good for small AI tasks on laptops/handhelds but are still being beat by the gpu in tokens per watt if you can actually find software that will use the npu at all.Stick with gpus, npus aren't there yet.
What's a NPU? A non processing unit?
>>109678235>NPUSure.
>>109681034>Killer NICnow that's a real NPU
How come the chinks haven't copied Google's tensor units already? They're way better than GPUs.
>>109681060don't you think all the big chip companies would if they could? Google keeps it's secrets but sells yours
>>109681034>>109681040https://en.wikipedia.org/wiki/Neural_processing_unit
