Well nVidia has to bring that power consumption waaaaaaaaaay down before they ever become a serious contender compared to DSP, FPGAs that go straight into the end application hardware.
In the data center we will see they might have a lead now but that can change quickly. All hyper scalers are seriously on the move for alternatives and again here it is their power consumption which probably mainly the problem.
]]>At this Point Nvidia is totally dominatiing the market. They have more powerful Products than any intel offering by a factor 3-4x for low precision training (21 TFLOPS FP16).
Meanwhile the inferencing hardware is able to achieve 44 TOPS, compare that to the Xeon Phi which peaks at 6 TOPS.
And this new hardware from Nervana is 3 quarters away and is totally unproven.
It must suck when the facts slap you in the face like this 🙂
]]>