REGISTER | LOGIN
Breaking News
News & Analysis

Google Designing AI Processors

TensorFlow accelerators used in AlphaGo
5/18/2016 04:45 PM EDT
23 comments
NO RATINGS
1 saves
Page 1 / 2 Next >
More Related Links
View Comments: Newest First | Oldest First | Threaded View
Page 1 / 3   >   >>
rick merritt
User Rank
Author
Beyond cost
rick merritt   5/21/2016 9:32:32 PM
NO RATINGS
Whatever the costs, this was a chip all Silicon Valley has been buzzing about for years.

We knew Google acquired semi startups and hired top talent like Norm Jouppi, but this was the first peek we'e had of what they are up to.

It seems to point beyond the kind of microprocessors that put the Valley on the map, to news kinds of architectures for new kinds of tasks.

 

DeeJee0
User Rank
Author
Re: Cost of TPUs
DeeJee0   5/20/2016 8:04:31 AM
NO RATINGS
It seems to me that the cost per TPU is not Googles main concern as long as they can pay it. IMO they chose to keep an advantage over the competition by not help start up nor make viable a future neural network chip giant over chosing to get the cheapest parts that allows for a stronger rollout but helps competition to get access too.

It is a smart move: programming these things is the main botlleneck. Look at the battle between IBM and Google. IBM has Watson and its dedicated neural network chip but very few commercial applications. Google has the GO succes and plenty of potential in house applications like facial and object recognition in pictures. They do not need to find out how to market it and convince potential custommers.

Buying these chips form IBM or startups with architectural breakthroughs is unneeded. It allows these new architectures to gain market share and reduce their cost. Competition can make good use of it even if their programming skills of these neural nets is subpar. Google seems to have enough time to go and explore a simple gpu-to-neural net optimisation by adjusting precision and so first while gaining simultaniously more neural net programming experience. By the time there is a huge non-Google market for dedicated neural net hardware, they have way more advanced software for it and can still reconsider doing the hardware in house or buy it.

 

sw guy
User Rank
Author
Cost data
sw guy   5/20/2016 7:44:27 AM
NO RATINGS
No need to break any NDA to have cost indications well below 200M$.
Just have a look to another news of same day: "ARM Describes 10nm Test Chip"

 

GSMD
User Rank
Author
Re: Cost of TPUs
GSMD   5/20/2016 1:49:05 AM
NO RATINGS
Most assuredly ! GPUs are not proving to be a panacea and just about everyone is exploring alternate architectures. Power is one issue and the programming model is another. No idea which is the most promising approach. We have chosen to experiment with two core variants,  using the neuromorphic approach for one and a stochastic approach for another. We are also exploring techniques like oscillator based designs (instead of gates) where search dominates.

I expect  AI based vision systems, like yours,  will also take the non-GPU path.

GSMD
User Rank
Author
Re: Cost of TPUs
GSMD   5/20/2016 1:09:29 AM
NO RATINGS
Completely agree, there seems to be a campaign going around to discourage folks from doing their own designs ! To confirm your nos

tools cost - 0.5 to 1.5 M per part. A startup can easily do it at half a mil

IP cost - dropping like stone, cores are free these days (RISC-V). High speed PHY is te pain point but yiu can negotiate with the foundry. So budget 0.5 - 4 M

Engineer cost - depends on country. In my earlier startup I concerted my PCB designers to IC backend guys. So I got P&R guys at USD 600 per month ! Even these days a 5 member team can be had at 20K USD per month. Use tools like Bluespec to bring down design costs.

Die costs - 28 nm is going for around USD 100 per sq mm for sub 50 lots and USD 2-3 per sq mm for sub 500. So after spending a M or two on tapeout  larger chips (above 100 sqmm) typically cost  USD 250-500 in low lots. Show an NDA and I will share more details ! Add pacakging costs. Not a great increase to these nos. I got these nos from my cat, so it should be taken with a grain of salt !

 

SW dev costs are negligible. I can guys who can reqrite Linux at USD 5000 per month. In fact one is writing me a microkernel from scratch. Most devs cost a lot less. These are Indian prices of course.

To sum up, even 5M is not a bad budget. If yiu are talking microcontroller parts, then 1M will do assuming rented tools. Tapeout will be sub 50k.

 

MPW nos can be found in europractice-ic dot com

 

 

bobbytsai
User Rank
Author
Re: Cost of TPUs
bobbytsai   5/20/2016 12:40:53 AM
NO RATINGS
dev board / ate test / qual cost : @3dlabs we did a reticle limited chip for costs you mentioned.  maybe a little more on dev board but less on ate hw / qual.

for startups is there a rent an EDA tool model ?

back 20 years ago we did shuttles @ umc and tsmc for new processes.  there were 8 - 20 chips on a mask set.  some designs were linear shrinks from .35um (now that is being cheap). 

adapteva
User Rank
Author
Re: Cost of TPUs
adapteva   5/19/2016 11:22:58 PM
NO RATINGS
Kevin,

I am bound by NDAs in all directions (foundry, eda, ip, partners) so clearly I cannot defend my cost data properly.

Still...I ask you and anyone else who keeps propagating these inflated chip design cost myths to back it up with real data or stop bringing it up!

Propagating this myth is doing damage to an already frail US semiconductor eco-system.

 

Kevin Krewell
User Rank
Author
Re: Cost of TPUs
Kevin Krewell   5/19/2016 8:44:52 PM
NO RATINGS
I'm checking with my sources on costs and power. Considering the size of the heatsink on the TPU and the 6 power transistors feeding it, I think it could be over 25W. GPUs are allowed up to 75W and this is a closed system, so Google could bend the rules. Also, we haven't seen the backside of the card which could have additional power rails.

WRT to design costs, I will reference this EETimes article on ASIC vs. FPGA costs:

http://www.eetimes.com/author.asp?doc_id=1322021

Google could have also perfomed more simulations before tape out. I think Andreas cost for EDA tools is way low. The latest Synopsis and Candence tools, as well as other specialized tools are way more than $1M. Probably more like $1M/person/year. 

WRT to Nervana Systems, they have been in business since 2014 and already have a software product, so they could either be funding the chip from revenue or they could go out for another round of funding to support tape out and initial production.

rick merritt
User Rank
Author
Re: Cost of TPUs
rick merritt   5/19/2016 7:41:59 PM
NO RATINGS
@SamF787  Good point on the thermal limit given the PCIe card host!

 

alex_m1
User Rank
Author
Re: Cost of TPUs
alex_m1   5/19/2016 6:41:27 PM
NO RATINGS
Kevin: Nervana systems is a startup aiming to do the same thing: design a TPU chip, offer a cloud services with it, and i heard they will also have about 10x improvement in performance/watt. As for performance per chip, we don't know either on them or on Google.


And Nervana's total funding is $24 Million, and they say they chip will be ready in 2017. So i'm with Andreas on this one.


BTW, the research about this subject is talking about energy savings of 100x-3000x , so i wonder why Google's number is so low.

Page 1 / 3   >   >>

Datasheets.com Parts Search

185 million searchable parts
(please enter a part number or hit search to begin)
Like Us on Facebook
EE Times on Twitter
EE Times Twitter Feed