Hot Chips 2017: A Closer Look At Google's TPU v2

[Guest]

Google's Cliff Young shared details about its TPU (Tensor Processor Unit) at Hot Chips 2017, but most importantly, the company also revealed more details about the Cloud TPU, otherwise known as TPU v2, during its keynote.

Hot Chips 2017: A Closer Look At Google's TPU v2 : Read more

 

[bit_user]

If you wonder why HBM is in short supply, the Cloud TPUs might be part of the answer.

I had this thought, as it seems AMD was caught off-guard by higher-than-expected HBM2 prices. So, do we know it's definitely HBM and not HBM2?

 

[bit_user]

It's also rather disingenuous to compare TPU2 to Kepler. Based on when it came online, they should've compared it to Pascal.

However, the more interesting comparison is between the TPU and GP102. The latter claims to achieve 45.5 8-bit TOPS, with a nominal TDP of 250 W. So, the original TPU still has 6.7x better efficiency.

Anyway, I wonder why the TPU2 with 32-bit, while everyone else is trying to do inferencing with 16-bit floats.

 

[TadashiTG]

Eh, I think by now when someone says "HBM" people should know they are referring to HBM2. The first generation is probably not very useful anymore, and could be out of production altogether.
You can also tell by the capacity being larger than 4GB.