Sun Microsystems is researching ways to make massive supercomputers even faster, including wireless connections between CPU and memory.
The shortcomings of multi-core were discussed in the November 2008 issue of IEEE Spectrum. The publication of the Institute for Electrical and Electronics Engineers, quoted scientists from Sandia National Labs, who felt that multi-core processors were detrimental to high performance computing in certain usage cases.
Sun (NASDAQ: JAVA) has several solutions up its sleeve, not the least of which is a project it’s been working on for the last four year years it calls Proximity Communications. The work is part of a Defense Advanced Research Projects Agency (DARPA) project, the same Defense Department group that gave us the Internet.
Proximity Communication will allow a CPU to communicate with memory without using the multiple layers of copper wires in the motherboard that run between the CPU and memory sockets. Instead, a pair of metal plates that don’t quite touch will send and receive the signal back and forth.
This would reduce significant latency between memory and the CPU, a problem made worse in the era of multi-core processors because more cores are trying to access the memory through the same bus. It also eliminates latency because there is no need to charge up the data and copper wires for transmission. Instead, it’s transmitted at light speed between the two targets.
Sun expects to make a major presentation on its Proximity Communications milestones at the Electronic Components and Technology Conference (ECTC) in May 2009 in San Diego. Bjorn Andersson, marketing director of high performance computing products at Sun, said it will still be a few years before Proximity Communication is viable, but it will shave off quite a bit of latency.
“We can’t eliminate latency completely but it’s one or two orders of magnitude,” he told InternetNews.com. “Latency is the charge to send data over the wire to communicate. Proximity Communication avoids that completely. It comes down to light speed if you do it that way.”
The communications research is part of a DARPA project, so Andersson said he could not say if it would be licensed to other vendors or used elsewhere, such as on network controllers, which are also prone to latency.
A potential game changer
If Sun can get this to work, it could be very game-changing, said semiconductor analyst Nathan Brookwood. “The reason I say that is if you look at a processor today, typically those chips have half of their area devoted to logic and the other half devoted to caches,” he said.
“If someone had a way of being able to make an off-chip cache as fast as an on-chip cache, suddenly you’d be able to make systems with very large caches to alleviate main memory access time problems and do it at a much lower cost,” added Brookwood, research fellow with Insight 64.
Next page: Why the technology has a lot of promise
Page 2 of 2
Why the technology has a lot of promise
Gartner analyst Martin Reynolds also thinks the technology has a lot of promise. “The key to proximity is these connections are short and dense. That takes away a lot of support stuff, so the size of a server can come down. This can make servers significantly smaller and allow for an increase in density in the datacenter,” added Martin Reynolds, vice president and fellow with Gartner.
“This is the elimination of all that mechanical stuff that allows for a much higher density of space,” he said.
To some degree, computing is advancing in that direction already. Intel (NASDAQ: INTC) and AMD (NYSE: AMD) have in recent years switched from pin-based CPUs that have to be carefully socketed in the motherboard to the Land Grid Array (LGA), where metal contacts on the bottom of the CPU simply connect against metal on the motherboard. There is no need for longer pins, which can easily get bent.
Proximity Communications is the next step, where the metal is close but does not need to touch.
Sun is taking other steps to address the shortcomings raised over multi-core supercomputing. Andersson thinks the main problem is not multiple cores, slow cores or the lag between CPU and memory as much as it is that old ways of doing calculations are still being used.
Thirty year-old algorithms
“If you look at the core problem, it really is on the software side,” he said. “There are many ways you improve different algorithms to solve the problems. To really take advantage of the performance you can get out of a multi-core system, you need to change the algorithms. Some might be 30 years old.”
To address this, Sun is coming up with tools to help customers parallelize their apps, including a new high level language being developed with DARPA that allows a developer to write their app without having to do any of the parallelism, and when the application is compiled, it is automatically parallelized.
One of the main points of the IEEE Spectrum article was that in some cases, multiple cores aren’t the answer. The application needs faster cores, not more of them. The semiconductor industry hit the clock speed wall several years ago and shifted to multiple cores as a way of getting more performance.
Andersson doesn’t see the gigahertz race coming back, even with recent advances in CPU technology.
“To some degrees, we are still guided by the laws of physics,” he said. “It would be real hard to go beyond certain speeds and create problems with power and cooling and other things when you get up to those levels of speed. From a physics standpoint, it’s much harder to push much further than where we are today.”