Google lets you put TPUs in your own machine

[mjlef]

https://www.cnet.com/news/google-to-let ... f-october/

[Milos]

https://www.cnet.com/news/google-to-let ... f-october/

Considering how uncompetitive they are in price/performance terms to even V100, not to mention 1080Ti's, this seems to be a desperate measure.
And in HPC domain the new DGX-2 is so much ahead of anything that Google has to offer that it is clear they've already lost any hardware race.

[jhellis3]

Yeah, I can't imagine why anyone would buy into that platform when Nvidia cards are there, which handle a much, much wider array of compute tasks with an already robust ecosystem.

Or even next gen AMD cards will likely be far more generally useful. But whatever... people are free to sell their wares, and people are free not to buy them.

[melajara]

For mobile development, the Kirin 980 from Huawei should be out soon offering 5 Teraflops per Watt (yes, it would mean 1 Petaflop for a 200 Watt typical NVIDIA card

[Milos]

For mobile development, the Kirin 980 from Huawei should be out soon offering 5 Teraflops per Watt (yes, it would mean 1 Petaflop for a 200 Watt typical NVIDIA card

5TFLOPS/W what a joke. Are you aware that cutting-edge research chips are at around 2TFLOPS/W?
There is nothing so far in production that is above 1.5TFLOPS/W.
What kind of marketing BS can one read here...

[melajara]

This is because it is achieved on a TSMC 7nm node process, when NVIDIA Titan Xp is still on TSMC 16nm.
More and more, the leading edge will be on the mobile segment where future smartphones will be differentiated mostly by the power of their embedded AI (NPU in SoC)

[Milos]

This is because it is achieved on a TSMC 7nm node process, when NVIDIA Titan Xp is still on TSMC 16nm.
More and more, the leading edge will be on the mobile segment where future smartphones will be differentiated mostly by the power of their embedded AI (NPU in SoC)

Titan Xp is at 0.05 TFLOPS/W. That is like 2 orders of magnitude lower of what you wrote up there. I really don't think you have a clue how scaling or for the matter cutting-edge AI chips work.
For getting above 1TFLOP/W you need a totally different approach, making single-precision or integer multipliers is not gonna cut it even at 4nm node.