8 AMD FirePro S10000s (16 GPUs) achieve 8 TFLOPS real world double precision compute performance

Posted by Tony DeYoung on November 12, 2012

Consuming a total of 3.6 kilowatts, the folks down in the basement lab at AMD have built the fastest multi-GPU compute server using eight of their flagship compute powerhouse boards, the new AMD FirePro S10000!

This 16 GPU (eight S10000s) Exxact Computing Server provides more than 8 TFLOPS of real world double precision computing performance. While these are early drivers, this still means you are still seeing around 70% efficiency of the theoretical peak double precision floating point performance of 11.84 TFLOPS (47.28 TFLOPS peak single precision performance!).

Here’s a look at the DGEMM result and a glimpse at what’s under the hood of the compute server:






I wonder if the drivers are any good?!
Im even more courious about that wall o monitors on the right
That is work well done. We would definately be up for a beast like this! Some documentation of what was done during installation would be nice. Getting 16 GPUs up with BIOS (msot likely extended memory had to be used) would be nice.
so will it play crysis?
Hi! That thing IS just plain cool! But there are some questions I'm very interested in. Can you please write in details about all hardware, operating system and driver versions you use? On what matrix size did you achieve this result? Did you include transfer cost in these 8 TFlop/s or it's just a kernel-only performance? Thanks!
Hi, what is the brand of the chassis ? or motherboard ?
@remi, it's a TYAN FT77B7015. I would like to know what program you used to perform the benchmark. I'm looking for a console-based, multi-GPU OpenCL benchmark program to measure SGEMM/DGEMM performance.
We are considering buying similar machines at our institute, however we are concerned about the cooling of actively cooled cards when placed tightly one next to the other. We have had 3XHD5970s installed in a test machine with decent cooling, and the cards that had their vent sides covered by the neighbouring card simply fried (90-100 degrees Celsius under load). We had to seperate the cards by force to let a little air go in the middle. That reduced temperature to 80C, however this is not a clean solution. Our future chassis might not allow such hacks. Could you post temperatures under 30 minutes of LuxMark or something similarly long computations?
epixoip, check our multiGPU DGEMM research: http://devgurus.amd.com/message/1282591
+1 On how they got 8 dual GPU cards working... I'd like to know as well.. Also did the stock power supplies handle all 8 cards at max power? What OS is being used??
What is that, 30 grand worth of hardware? I don't even know what to say... Congratulations I guess?
They could probably scrub back and forth on an edit session of the Hobbit in its native 4k format, on that beast. I bet it generates a butt-load of heat too...
Page 1 of 1 pages

Add your comment

Note: All comments are moderated for spambots so there will be a posting delay.
Your email address will not be published.

Anti Spam: Please enter the word you see in the image below: