September 17, 2008
When I wrote about the new Cray CX1 and the brief and unfortunate history of personal supercomputing platforms in yesterday's blog, I intentionally didn't mention the SC072 deskside machine from SiCortex. This is because I knew I was about to get an update from the company about their product line and because the SC072 machine is not being positioned as an HPC production system, which the CX1 is. But that doesn't mean the gang at SiCortex isn't following the Cray announcement with a great deal of interest.
First, let's recap the highlights on the SiCortex refresh, which was announced on Wednesday. In a nutshell, the company says it made systematic improvements across the product line, increasing price-performance by a factor of two. They accomplished this by cranking up the clock speed on the SiCortex-designed MIPS chips and by improving compiler performance. At the same time, they reduced the cost of the hardware components, which means they were able to lower system prices across the board.
Clock speed on the SiCortex processor was bumped from 500MHz to 700MHz, a 40 percent increase. Meanwhile, power consumption increased only 25 percent. Since each processor CPU now delivers 1.4 gigaflops of performance (8.4 gigaflops per six-way node), the new SiCortex systems now achieve up to 400 megaflops/watt. "That actually takes us past Blue Gene/P," says John Goodhue, VP of Product Management and Applications Engineering. "And if you don't count [systems] with Cell processors, we are now the most power-efficient computer on the planet."
With the new hardware, their high-end SC5832 supercomputer can now achieve over 8 teraflops of peak performance for a mere 20KW of power. At the low end, the SC072 Catapult (now called the Personal Development System or PDS) puts out about 100 gigaflops while drawing just 300 watts.
The other piece of the performance improvement has to do with enhancements they've made to their in-house PathScale compiler. According to SiCortex, the compiler has been tweaked to generate code that, on average, is 20 percent more efficient across a range of workloads.
But their product line and target market space -- mid-range HPC systems -- will remain unchanged. SiCortex' interest in the Cray CX1 is based on its role as a turnkey deskside appliance for HPC applications. While the CX1 is not likely to be a development platform for Cray's high-end machines (architectures and software stacks seem too different), it could end up in such a role for a future lineup of Intel-based HPC appliances -- the same role the SC072 fills for the SiCortex product line. Currently, the SC072 is being bundled with the SC1458 and SC5872 production systems as a deskside development platform for the larger machines. The company will also make them available to academic institutions as HPC development machines for students. Starting next week, the SC072 appliances will be available on the company website.
A fully loaded SC072 will run about $25K, which just happens to be the entry point price for a CX1. But according to Kem Stewart, VP of Hardware Engineering at SiCortex, the SC072 is much more suitable as an office machine than the Cray CX1. Compared to the SC072, a fully populated CX1 using all 8 slots would consume up to 10 times more power. "That doesn't strike us as office material, since you would need two independent circuits," says Stewart.
But because of the lower raw performance of the SC072, that's not quite an apples-to-apples comparison. Stewart thinks a CX1 with just two compute blades, representing about 200 peak gigaflops, would deliver roughly the same application performance as a 100 gigaflop SC072, while requiring a lot more power. The tight integration on the SiCortex SoC and the internal Kautz graph topology interconnect fabric means that both memory bandwidth and node-to-node communication are generally superior to a system like the CX1, which is constructed from commodity parts. "The peak gigaflops performance really doesn't tell the story," explains Stewart. Ironically, Cray tells basically the same tale when comparing its high end supercomputers with scaled out commodity clusters.
Fully loaded, a CX1 could compete against the SiCortex middle-of-the road HPC appliance, the SC1458. The raw performance delivered by an 8-blade CX1 is attainable with a half-populated SC1458. But from Stewart's point of view, once you get to the high end of CX1, you're better off with a SiCortex box, since the SC1458 is more energy efficient, and gives you better price-performance, as well as expandability.
For those reasons, they feel like they have the CX1 bracketed between their development system and their production systems. At this point, SiCortex has no plans to go after the low end of the HPC market directly. Goodhue says that they'll continue to focus their efforts on the mid-market HPC segment, but admits they have considered repositioning the SC072 as a standalone production machine. "It's interesting, but it's not where we're going right now," he says.
Posted by Michael Feldman - September 16, 2008 @ 9:00 PM, Pacific Daylight Time
Michael Feldman is the editor of HPCwire.
No Recent Blog Comments
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?