REGISTER | LOGIN
Breaking News
News & Analysis

Google Revs Faster TPU

Machine-learning ASIC doubles performance
5/17/2017 01:20 PM EDT
4 comments
NO RATINGS
Page 1 / 3 Next >
More Related Links
View Comments: Newest First | Oldest First | Threaded View
dstow
User Rank
Rookie
Re: No the new ASIC DOES NOT handily beats GPUs on training
dstow   5/28/2017 10:46:00 PM
NO RATINGS
As jamestate mentions, if the TPU2 board (with 4 chips) has lower power than the top end GPU, then the TPU could indeed beat the GPU on training given the power limitations of the warehouse-scale server. 

The Volta has a TDP of 300 W, while a single old TPU1 in 28nm had a 40W TDP. In a power constrained system, it seems likely that more TPU chips could be used per CPU host and that they could run at peak/boost speeds (if applicable to the new TPU2), while the GPU may be power/thermally constrained. The TPU1 paper even specifically mentions that the GPU boost clock could not be used in the datacenter because of power limits and slow boost timing resolution.

Of course, Votla was also just released and Jensen admitted that the product shown was not yet in manufacturing ramp-up, so Google's statements may have been made without respect to the not-yet-tested Volta. Maybe we should take both Google's and Nvidia's statements with a grain of salt until real benchmarks are released. Their are also bound to be plenty of confounding factors related to framework optimization; it's likely that hardware-software co-optimization between the TPU ISA and TensorFlow drivers could make it easier to achieve peak throughput for a wider variety of neural network models. 

SalemtheCat
User Rank
Author
Re: No the new ASIC DOES NOT handily beats GPUs on training
SalemtheCat   5/27/2017 11:50:47 AM
NO RATINGS
Article quote: Google said that the new ASIC handily beats GPUs on training.

 

So Google outright lied (what a suprise) as a 45 TFLOP TPU does not beat a 100 TFLOP GPU.

 

As for the strawman argument about power there is no mention in the article about it and even if it did it doesn't change the lie about handily beating GPUs on training.

jamestate
User Rank
Freelancer
Re: No the new ASIC DOES NOT handily beats GPUs on training
jamestate   5/21/2017 6:47:18 AM
NO RATINGS
The ASIC could beat it if it uses less power per FLOP and costs less.

 

SalemtheCat
User Rank
Author
No the new ASIC DOES NOT handily beats GPUs on training
SalemtheCat   5/18/2017 12:02:52 PM
NO RATINGS
How does a 45 TFLOP ASIC beat a 100 TFLOP GPU.

 

Answer it does not.

 

It looks like GOOGLE is going the mis-information route Sad.

 

https://www.nvidia.com/en-us/data-center/volta-gpu-architecture

 

 

 

 

Like Us on Facebook
EE Times on Twitter
EE Times Twitter Feed