Sign up for our newsletter and get the latest HPC news and analysis.
Send me information from insideHPC:


Introducing the SGI UV 300 Big Memory Supercomputer

In this video from SC14, Brian Freed from describes the company’s new SGI UV 300 supercomputer.

SGI has been delivering scaleable shared memory systems for over 2 decades. The 7th generation SGI NUMAlink7 technology is being delivered in the SGI UV 300, launched at SC14. A single UV 300 chassis (5U) provides 4 Xeon E7 processors, up to 96 DIMMs, and 12 PCIe slots. Connecting 8 of these units together in a single rack, using NUMAlink7 interconnect, then creates a huge pool of resources, up to 480 cores, 24TB DRAM, which can be extended with additional PCIe devices for highest performance FLASH in the industry, as well as compute acceleration with Intel Xeon Phi and visualization capabilities – all in a single system!

Full Transcript:

insideHPC: I thought we should get a little acquainted with this new device here, the SGI UV 300 and learn what’s up.

Brian Freed: Absolutely. So this is the UV300, it’s the follow on. It’s our third generation x86 Scale-Up Solution. But it really is a departure from our historic x86 Scale-Up solutions, that it’s focused on I/O versus compute power. So its predecessor was designed to scale, compute to the max. So 256 sockets in a single Linux instance. This one is designed to give you much more I/O. So we’ve tripled the I/O per socket from 8 DIMMS per socket to 24. And we’ve also improved the point-to-point connectivity, so that it’s all connected from 4 to 32 sockets. To give the lowest possible latency between any two sockets. So, it’s built in four-socket chassis, so from any four-socket chassis to any other four socket chassis in this rack, is less than 500 nanoseconds of latency.

insideHPC: So what kind of jobs would shine in this kind of architecture?

One of the jobs that shines is SAP HANA. We won an award for best use of HPC in the commercial space with our SAP HANA box. It is built on this platform, the UV 300. But in addition, there’s a number of more traditional HPC codes. We’re really excited about opportunities around graph analytics, and some electromagnetic simulations look really promising on this solution. In addition, when we look at the big data space, one of the trends we see is a desire to combine applications. Instead of thinking of just one application running on this box, it’s scenarios where they combine multiple applications. So things like search, semantic analysis, text analytics, Geospatial and graph, all into a single application. So solving problems using multiple applications each accessing the same data.

insideHPC: It retain the goodness of the UV that I know, which is the single system image, right? 

Brian Freed: It does. It retains the single system image which is, it’s really core to UV. But it brings a tremendous amount of new flexibility and new application use cases given the I/O that it delivers. The other thing that we’re demoing today, and there’s a number of use cases we can think of, and we expect there’s going to be more evolved, is we’re demoing a solution that uses Intel PCIe NVM brand. So, this is the Intel P3700. And at 32 socket UV that’s in our lab in Chippewa Falls, we have 64 of these, plugged into the PCI slot. It delivers a 128 terabytes of PCI based flash storage, 30 million IOPS. So probably the most powerful machine in terms of IOPS in the world and about 200 gigabytes per second. So it’s a tremendously powerful solution from a storage perspective.

So, I guess the question you’d ask is, what would you use all that for?

insideHPC: Yes.

Brian Freed: There’s really two things. I think first is, you could use it for warm data. So not everything needs to be in the DRAM. So it still got 24 terabytes of DRAM. But now you can have that 24 terabytes of DRAM augmented by another 128 terabytes of flash, that’s very near to you, it’s a near line flash.

The second thing it does is, it allows you to cycle models quicker. So when you think of one of the challenges of modeling, often times the set up, getting the data into the system, into the memory takes more time than it actually process it. In this case, you have the data’s – you’re sitting there hot standby and flash and you can bring it in and cycle through the processes much more rapidly.

insideHPC: Well, it sounds like a big data monster to me. Would that be a good characterization here?

Brian Freed: Absolutely. We do believe it is a big data monster, and it’s also good for many applications in the HPC as well as the commercial space.

See our Full Coverage of SC14. * Sign up for our insideHPC Newsletter.

Resource Links: