Since some folks have already built Llano APU powered servers with excellent success and very low power consumption, I would not be surprised to see SeaMicro develop a Trinity (and it's successors), powered server system for those who want impressive low power consumption with good performance.
I don't see how Trinity is ever appropriate for a server. The iGPU is practically useless - too slow for any kind of HPC use, and unnecessary for other applications. Piledriver Opterons should be more power efficient and support things like registered memory.
That's something I'd like to see as well, and it would have a place. Remember, what you're talking about in a server situation is many times a 'headless' system (ie: no monitor/keyboard/mouse) so there's no need for a graphical interface (command line only). As such, there's plenty of die space (the iGPU) just sitting idly by instead of being utilized - and it's all still fitting in the 15-30w TDP.
So instead of wasting the iGPU, put it to use. Putting Trinity into a quad-configuration would come in at 60-120w TDP (lower than Bulldozer) - provide 16 cores and 4 decent IGPU's (which Bulldozer lacks). With the dynamic power control, you could go as low as 1 core, 1iGPU and scale it as needed up into what I would consider a pretty powerful machine.
Just counting CPU/GPU TDPs, they are very similar, with the 7970 system actually slightly ahead: 1) 20x 35W = 700W 2) 2x 95W + 2x 230W = 650W
Keep in mind that to run 20x Trinity, you also need 20x motherboards, memory, disks, etc. all of which consume power. It's also much more difficult to keep 20 systems up and productive compared to one or two. Seamicro's cluster concept alleviates the problem somewhat regarding hardware and power, but it does not solve the software side at all.
GPUs are very useful in compute servers. There are entire supercomputers built with GPUs (http://www.bit-tech.net/news/hardware/2010/10/28/n... so if you are going to use software appropriately written to be accelerated by GPU, the Trinity's version of this "cluster" is going to kick Ivy Bridge where it hurts. It all depends on the applications: it is clear that for CPU-intensive applications Ivy Bridge has a significant advantage, but these are not machines that are intended to run MS Word ...
You are right in saying GPUs can be useful. However, your comparison is missing the point. Instead of using solely Trinity or Ivy Bridge, you could use an ordinary processor and a couple of Tesla cards (as the supercomputer you linked to does), which would use 5x the power per server but give you 50x the GPU performance.
Again if you look at the efficiency there's still a good case for using a cluster of APU's where you utilize both CPU and GPU's the fact that you can build something better will always be true, but efficiency wise a cluster of Trinity GPU's is not something to scoff at.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
9 Comments
Back to Article
Casper42 - Monday, September 10, 2012 - link
I can get with Chassis or Enclosure or even Cluster, but calling this a "Server" just sounds odd.Beenthere - Tuesday, September 11, 2012 - link
Since some folks have already built Llano APU powered servers with excellent success and very low power consumption, I would not be surprised to see SeaMicro develop a Trinity (and it's successors), powered server system for those who want impressive low power consumption with good performance.madmilk - Tuesday, September 11, 2012 - link
I don't see how Trinity is ever appropriate for a server. The iGPU is practically useless - too slow for any kind of HPC use, and unnecessary for other applications. Piledriver Opterons should be more power efficient and support things like registered memory.Medallish - Tuesday, September 11, 2012 - link
Well the Trinity iGPU is pretty efficient, if you throw a bunch of them in a low power server, even the iGPU's should have a place.bill.rookard - Tuesday, September 11, 2012 - link
That's something I'd like to see as well, and it would have a place. Remember, what you're talking about in a server situation is many times a 'headless' system (ie: no monitor/keyboard/mouse) so there's no need for a graphical interface (command line only). As such, there's plenty of die space (the iGPU) just sitting idly by instead of being utilized - and it's all still fitting in the 15-30w TDP.So instead of wasting the iGPU, put it to use. Putting Trinity into a quad-configuration would come in at 60-120w TDP (lower than Bulldozer) - provide 16 cores and 4 decent IGPU's (which Bulldozer lacks). With the dynamic power control, you could go as low as 1 core, 1iGPU and scale it as needed up into what I would consider a pretty powerful machine.
madmilk - Tuesday, September 11, 2012 - link
You need to consider performance AND low power in efficiency.I would consider these comparable on absolute CPU and GPU performance, assuming very good parallelization:
My examples use Xeons, but Piledriver Opterons may be a viable alternative:
1) 20x Trinity (80 Piledriver "cores" @ 2.3GHz + 7680 VLIW4 cores @ 500MHz)
2) 4x Xeon E5 8-core + 2x 7970 (64 SB-E threads @ 2.3GHz + 4096 GCN cores @ 900Mhz)
Just counting CPU/GPU TDPs, they are very similar, with the 7970 system actually slightly ahead:
1) 20x 35W = 700W
2) 2x 95W + 2x 230W = 650W
Keep in mind that to run 20x Trinity, you also need 20x motherboards, memory, disks, etc. all of which consume power. It's also much more difficult to keep 20 systems up and productive compared to one or two. Seamicro's cluster concept alleviates the problem somewhat regarding hardware and power, but it does not solve the software side at all.
yankeeDDL - Tuesday, September 11, 2012 - link
madmilk,GPUs are very useful in compute servers.
There are entire supercomputers built with GPUs (http://www.bit-tech.net/news/hardware/2010/10/28/n... so if you are going to use software appropriately written to be accelerated by GPU, the Trinity's version of this "cluster" is going to kick Ivy Bridge where it hurts.
It all depends on the applications: it is clear that for CPU-intensive applications Ivy Bridge has a significant advantage, but these are not machines that are intended to run MS Word ...
madmilk - Tuesday, September 11, 2012 - link
You are right in saying GPUs can be useful. However, your comparison is missing the point. Instead of using solely Trinity or Ivy Bridge, you could use an ordinary processor and a couple of Tesla cards (as the supercomputer you linked to does), which would use 5x the power per server but give you 50x the GPU performance.Medallish - Wednesday, September 12, 2012 - link
Again if you look at the efficiency there's still a good case for using a cluster of APU's where you utilize both CPU and GPU's the fact that you can build something better will always be true, but efficiency wise a cluster of Trinity GPU's is not something to scoff at.