SGI skips future Xeon E7s, lobs E5-4600s into UV2 supers
Adds Xeon E5s, Tesla K10 GPU coprocessors to rackable boxes
Supercomputer and cluster maker Silicon Graphics has fallen hard for Intel's new "Sandy Bridge-EP" Xeon E5-4600 processors for four-socket servers.
So hard in fact that with its next generation UV2 shared memory supercomputers, due later this year, SGI is not supporting the future Xeon E7 processors that are the obvious kickers to the Xeon 7500s and E7s that are at the heart of the current UV generation of machines.
"We are moving boldly into the Xeon E5, and this is the new way of doing it," Bill Mannel, vice president of product marketing at SGI, tells El Reg. "While the Xeon E7 offers memory addressability and other RAS features that the E5 doesn't, the price/performance of the E5 is compelling for our HPC market. The good news is that with the UV2, everything will be significantly bigger, better, and faster."
SGI could have already supported the 10-core "Westmere-EX" Xeon E7 processors in the existing UV 1000 servers, which were previewed in November 2009 and which started shipping them the following spring using the eight-core "Nehalem-EX" Xeon 7500 processors. SGI has dropped the Altix part of the name, due to its association with the Itanium processor used in prior generations of shared memory systems, so now they are just the UV 1000s.
The original UV machines were developed under the code-name "UltraViolet" and they marry Intel's Xeon 7500 processors and related "Boxboro" chipset with SGI's own NUMAlink 5 high-speed interconnect router. This mashup creates a shared memory system that has 128 blades with a total of 256 sockets that can support up to 2,048 Xeon 7500 cores and an aggregate of 16TB of global shared memory. The nodes are hooked together in an 8x8 (paired node) 2D torus. The prior generations of Xeon 5500 and 5600 processors did not have enough QuickPath Interconnect (QPI) links or the main memory capacity to be used in a shared memory system, but the Xeon E5-4600s, which have two QPI links per chip and support up to a dozen memory slots per socket, can do the task. And they have the benefit of a simpler memory architecture compared to the Xeon 7500s and E7s, which have buffered memory controllers and which are really designed for much fatter configurations (in terms of CPU count and main memory capacity) than the E5-4600s.
Mannel says that SGI will be doubling up the core count on the future UV2 supers from 2,048 in a single system image using a global address space to 4,096 cores. While SGI is not saying how it is going to do this and is not providing any details on the NUMAlink interconnect inside the UV2s, it seems likely that the machines will have a NUMAlink 5+ interconnect that can support up to 256 blades, each with two E5-4600 processor sockets. And because the memory controllers on the E5-4600 has been goosed, SGI can quadruple the global address space across those 4,096 cores to 64TB and, thanks to the AVX instructions in the processor as well as slightly higher clock speeds, will be able to double the performance in a rack from 5.5 teraflops with a top-end UV 1000 system to 11 teraflops with a future UV2 system (perhaps to be called the UV 2000).
That will be done by doubling up the blade count, and it looks like it SGI will be able to pack twice as many blades in a rack. A current 128-blade UV 1000 machine takes up four racks.
SGI is not giving any more specifics on the future UV2 machines, but has said that it will offering them in a number of "different packaging options" to address a wider number of customer needs and markets. No word on when the UV2 will ship, but SGI is expected to provider more details at the International Supercomputing Conference in Hamburg, Germany in mid-June.
SGI is also fired up about adding the E5-2400 processors for two-socket boxes to its preconfigured Hadoop clusters, which are based on its Rackable C2005 2U, half-depth, sometime in the "early summer." Compared to machines using the Xeon 5600 processors, Mannel says the updated Hadoop clusters will deliver 22 per cent better bang for the buck and 27 per cent better performance per watt running the Cloudera distribution of the Hadoop stack.
SGI has also announced that later this summer it will pack the new Tesla K10 GPU coprocessors, based on Nvidia's "Kepler" GK104 GPUs, inside its Rackable C1104 servers, which will allow for two Tesla K10s to be attached to the server node. This setup will cram around 75 teraflops into a rack for less than $200,000 in a starter kit configuration. ®
Sponsored: Hyper-scale data management