I don't understand these articles that assume you understand what "two 2.40-gigahertz Intel Xeon E5645 6-core... and two NVIDIA M2050/C2050 448-core .... which reside on a Supermicro 2026GT0TRF motherboard" means, but didn't already know what "CPU" and "GPU" stand for.
From Feng: “The next frontier is to take high-performance computing, in particular supercomputers such as HokieSpeed, and personalize it for the masses.”
This is one of my favorite things about technology. As few as five years ago I was drooling about an Intel C2D chip and looking forward to quad-core chips. Nowadays that's standard technology in smartphones. I'm amazed by the graphics performance in the new Asus Transformer Prime with the Tegra 3 chip which has up to five cores.
The biggest problem with packing more power into smaller spaces is the battery consumption issue. Maybe there's no silver bullet for that only more efficient software and hardware usage.
I am not trying to be difficult but how is a $1.4M anything for the masses? Are they allowing the public to run jobs on the machine? Is there anything new about HS that brings SC500 power closer to my home office?
"The Virginia Tech team is figuring out how to best to farm out computing jobs so that the GPUs and CPUs do what they do best, without ever going idle, and without spending too much time communicating with one another.
It’s not easy, but they’re using HokieSpeed to build tools for designing and compiling software so that it can be tweaked to run fast on these systems. They’ve also built what they call an “automated runtime system,” which works with the supercomputer’s operating system to speed things up even further."
I think the 'supercomputing for the masses' part comes from the fact that the peak performance of a relatively low-cost GPU is large compared to a CPU, but actually taking advantage of that performance is still difficult. They're aiming to change that with improved software development tools. At least, that's what I gathered from the article.
Whilst your average person at home may not be able to fork out $1.4M, this is in the afordable range of many medium to large businesses, and opens up access to private compute clusters*
The highest cost for supercomputers isn't the raw hardware - it's the power consumption. This also includes the cooling needs of a large cluster. Over the lifespan of a cluster, these will add up to a significant multiple of their base cost.
* of course, this may not be the ideal way to provide access to these services
As a Hokie, this is good but SystemX was way more revolutionary. At the time, supercomputers weren't built on consumer hardware and the fact that it was in the top 5 (3?) with a bunch of G5s was like a shockwave. Basically, COTS supercomputers had finally arrived when SystemX came out.
As another VT alum (two times over, and the second time being last year), I can tell you that what they're probably most interested in is the energy efficiency. Feng's group does research with both GPUs and power efficiency. Wu Feng and Kirk Cameron (also at VT) were the ones who started the Green500 list.
I've noticed the last few years of supercomputer announcements has focused less on raw performance and more balanced performance/energy.
Small world, I'm also a '03 alum. I was just going through my address book and a Scott S. was in the list with a 540 area code. We might have been in a study group at one point.
ASCI Red was a supercomputer built for Sandia by Intel so not exactly COTS.
Beowulf was still in the experimental phase. Researchers were building 5-20 node clusters. IIRC, there weren't any Beowulf clusters you would consider supercomputer at that time. Keep in mind the fastest supercomputer at the time cost $400 million and this one cost just $5 million.
I left in 03 so AFAIK the researchers using SystemX were coming from ICAM and the Bioinformatics department. The professor I worked under (as a gopher) had a 4way G4 cluster doing protein synthesis. He was very excited to get processor time on this supercomputer.
But whatever, some publicity stunt. They were stupid enough to waste money upgrading and expanding it because no one was using it.
System was built in 2003 - 8 years ago.
Assume moore's law of doubles every 2 years - expectation of 16x more powerful
Actual increase in performance - 22x in 1/4th the size
Given that it's then a quarter of the size of system X, that's an amazing increase in peak performance.
There's only one problem - that speed increase appears to owe a lot to the use of GPGPU. As I understand it, whilst research into GPGPU for HPC* is a hot area at the moment, the scale of the actual benefits it offers is still a matter of debate (especially when considering costs and power consumption).
From my perspective, the biggest limitation in using GPUs for more general purpose computations is the communication latency. I published a paper that came to that conclusion: http://people.cs.vt.edu/~scschnei/papers/debs2010.pdf
In short, parallelism is not enough to get benefit from using GPUs. You need parallelism and data reuse.
17 comments
[ 2.2 ms ] story [ 46.7 ms ] threadThis is one of my favorite things about technology. As few as five years ago I was drooling about an Intel C2D chip and looking forward to quad-core chips. Nowadays that's standard technology in smartphones. I'm amazed by the graphics performance in the new Asus Transformer Prime with the Tegra 3 chip which has up to five cores.
The biggest problem with packing more power into smaller spaces is the battery consumption issue. Maybe there's no silver bullet for that only more efficient software and hardware usage.
"The Virginia Tech team is figuring out how to best to farm out computing jobs so that the GPUs and CPUs do what they do best, without ever going idle, and without spending too much time communicating with one another.
It’s not easy, but they’re using HokieSpeed to build tools for designing and compiling software so that it can be tweaked to run fast on these systems. They’ve also built what they call an “automated runtime system,” which works with the supercomputer’s operating system to speed things up even further."
I think the 'supercomputing for the masses' part comes from the fact that the peak performance of a relatively low-cost GPU is large compared to a CPU, but actually taking advantage of that performance is still difficult. They're aiming to change that with improved software development tools. At least, that's what I gathered from the article.
Here's a link to research they're doing on GPU computing: http://synergy.cs.vt.edu/
The highest cost for supercomputers isn't the raw hardware - it's the power consumption. This also includes the cooling needs of a large cluster. Over the lifespan of a cluster, these will add up to a significant multiple of their base cost.
* of course, this may not be the ideal way to provide access to these services
Small world, I'm also a '03 alum. I was just going through my address book and a Scott S. was in the list with a 540 area code. We might have been in a study group at one point.
Beowulf was still in the experimental phase. Researchers were building 5-20 node clusters. IIRC, there weren't any Beowulf clusters you would consider supercomputer at that time. Keep in mind the fastest supercomputer at the time cost $400 million and this one cost just $5 million.
But whatever, some publicity stunt. They were stupid enough to waste money upgrading and expanding it because no one was using it.
System was built in 2003 - 8 years ago. Assume moore's law of doubles every 2 years - expectation of 16x more powerful Actual increase in performance - 22x in 1/4th the size
Given that it's then a quarter of the size of system X, that's an amazing increase in peak performance.
There's only one problem - that speed increase appears to owe a lot to the use of GPGPU. As I understand it, whilst research into GPGPU for HPC* is a hot area at the moment, the scale of the actual benefits it offers is still a matter of debate (especially when considering costs and power consumption).
In short, parallelism is not enough to get benefit from using GPUs. You need parallelism and data reuse.