NEC Strikes Back With SX-8 Supercomputer

News for nerds writes "It was just 3 weeks ago that we learned IBM's BlueGene/L with 36.01 TFlops edged out NEC's Earth Simulator, but today NEC announces a new SX-8 supercomputer with a peak processing performance of 65 TFlops (press release). It may be available in the U.S. as Cray's OEM like SX-6."

Re:Runs Linux?

[mukund]

In reply to my own post..

The Cray SX-6 System runs the UNIX-based SUPER-UX operating system. [cray.com]

Sorry about that. Maybe they ported GFS.. dunno.

BGL will be back on top in short order

[tubbtubb]

Remember that the 36.01 TF figure for BlueGene/L was only using 8 racks.
The final BG/L will use 64 Racks.
Also, the SX-8 figure is only an estimation.

Re:65 TFlop is only an estimate

[foobar3149]

I just wanted to point out that is not an apples to apples comparison, and that such a comparison is not fair to either machine. For that matter, the number for the IBM machine is only a subset of BlueGene/L so then why not use the 360 TFlop projected performance of BlueGene/L as the comparison point?

Re:In the interest of perpetuating rumors

[Anonymous Coward]

Which rumors? System X is back up and running with 2.3 GHz chips. And bring any new surprise contenders/drop outs should place 6th, next go around.

Re:65 TFlop is only an estimate

[Troy Baer]

You're trying to say they could have overestimated the performance by almost 100%?

No, the original poster was saying that the 65TF number from NEC is theoretical peak performance based on the maximum possible number of FP operations per clock cycle (which can never happen in real code, due to pesky little things like having to access memory), while the ~35TF number for the Blue Gene/L prototype is measured performance on an actual piece of code called the Parallel Linpack benchmark. It's not unusual for systems to perform as low as 50% of peak on Parallel Linpack, although 70-90% is more typical on systems with decent memory bandwidth (which the SX8 presumably has).

(Note that I'm deliberately sidestepping the debate over whether Parallel Linpack bears any resemblence to reality.)

Re:Yes, but...

[Anonymous Coward]

Uh...
What is 43-1?

Re:65 TFlop is only an estimate

[enrique66]

Vector systems like SX-8 reach typically 50% of the peak performance in well tuned applications while massively parallel scalar systems crawl at less than 10% (no matter how much man power one invests in tuning). The effective performance of the SX-8 is much higher than that of BlueGene. The peak Linpack performance is meaningless, anyway. HPC Challenge benchmark makes a lot more sense for understanding what performance a machine can reach.