REGISTER | LOGIN
Breaking News
News & Analysis

China Tops U.S. in Supercomputers

Top ranked system uses China custom CPU
6/20/2016 03:00 AM EDT
5 comments
NO RATINGS
1 saves
Page 1 / 5 Next >
More Related Links
View Comments: Newest First | Oldest First | Threaded View
chipmonk0
User Rank
Author
who made the chips ?
chipmonk0   6/24/2016 1:49:59 PM
NO RATINGS
RICK M. :

 

You wrote :

 " It's not known who makes the current SW26010 or what process node it uses. "


Could they get the 2x power efficiency ( GF / watt ) improvement over Xeon Phi with just 28 nm which is the best that SMIC can manage so far ?


If not, could these chips been made at TXXX or UXX as condition for doing business in China  ?

 

Enquiring minds like to know !

LeonardTsai
User Rank
Rookie
Re: Sold out by Intel
LeonardTsai   6/21/2016 5:38:49 PM
NO RATINGS
No, TaiHuLight architecture is not based on TienHe 2; The interconnect and topology is different. Although HPC similarity is common. 

Inside China, there are multiple teams compete for the funding in HPC field; It is similar like how they bet on ISA by collecting them all one way or the other. I speculation is that Chinese believe their domestic demand is big enough (to support all variance) so let them compete and the winner will come out eventually (somewhat crude way of Darwinism although party does come in sometimes to direct the resource feed). 

Once China pick up one tech, they will replicate (although they have not yet master the art of 'copy exact'). This is almost by (market) nature. 

I think US complacency gave up on HPC field 15 years ago. I've seen many friends (briliant engineers that create standards used inside HPC and other field) that were out of jobs (from Cray, SGI, and many places) and had to find other works to support family. It was a sad (and long) story to tell. 

LeonardTsai
User Rank
Rookie
Re: looking forward
LeonardTsai   6/21/2016 1:08:23 AM
NO RATINGS
My view that FPGA integrated into CPU core will promote ISAaaS (Instruction set Architecture as a service) that enable to sell specific micro code accelerator as "App market".  The question remainds is whether the FPGA interconnect communcaiton over head will not overload the performance gain. 

NV GPU accelerator approach has its immediate advantage of hard-wired core, but in long term, software define ISA on FPGA will allow core case acceleration that probably more efficient than kernel on GPU. 

At least, HPC race is heating again and no longer the lonely field for aged engineers. :)

Ping you next time I route through bay area. 

 

 

 

 

rick merritt
User Rank
Author
Re: looking forward
rick merritt   6/21/2016 12:58:56 AM
NO RATINGS
Hey, Leonard, it's been too long. Drop me a line and let me know where you are these days and what you are up to.

Meanwhile, regarding your question: Intel will try to make FPGAs huge in the data center and HPC, driving a hole Xilinx and partners will try to drive through with chips riding their CCIX interconnect.

But Nvidia is all over this already as the most used accelerator in the Top500 and with its NVLink on the IBM Power8.

I think there will be all sorts of questions about tools and programming models and thermals and interconnects to work out for years.

Note that Google is already using its own custom ASIC accelerator and others are rumored to be developing ASIC accelerators, too. And Giogle and Facebook are already using lots of GPU accelerators for machine learning.

This area will be hot for years.

LeonardTsai
User Rank
Rookie
looking forward
LeonardTsai   6/21/2016 12:33:45 AM
NO RATINGS
Nice article, Rick,

With FPGAs getting into data center mix, how do you view the impact on HPC designer and future architecture?

Leonard 

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