NEC’s SX-8R Vector-CPU to do a 35.2 GFLOPS

Via NEC HPCE Europe:

SX-8R single node - picture (C) NECNEC introduces the SX-8R CPU, a makeover of the SX-8 CPU. The CPU now has two Add- and Mul-vector-pipelines each and NEC raised the Gigaherzen from 2.0 to 2.2 GHz. A single CPU is supposed to do a sweat 35.2 GFLOPS. Two entry-level single-node system will be introduced as well; The SX-8R/B will have 1-4 CPUs with a peak-power of 149.6 GFLOPS, where the SX-8R/A will have 4-8 CPUs with a peak-power of 281.6 GFLOPS. The single-node systems use up to 64/128 GB of shared-memory.

The multi-node version SX-8R/M scales up to 512 nodes with 8 CPUs each, resulting in a total peak-power of 153.2 TFLOPS. The bandwidth is supposed to be 288,358.4 GB/s. The system can utilize up to 256 GB per node, resulting in a total 64 TB of RAM for the maximum configuration.SX-8R multi node setup - picture (C) NEC

So, what to do with this beast? Same as we do with all vector-machines: Doing matrix-manipulations. Contrary to “conventional” CPUs they don’t have to iterate over individual “cells” of a matrix doing calculations, vector-CPUs suck in total rows and columns at once into their registers, doing the calculation in a single step. A matrix-multiplication A x B = C for 5 rows and 5 columns would mean that a traditional CPU without optimization would need 5²=25 steps of addition and multiplication for getting the result C. A vector-CPU sucks in a whole row and column at once, resulting in only 5 steps. To say it short: “Traditional” iterative approaches run O(n²), whereas vector-approaches run linear in O(n). This also scales over the number of available pipelines and CPUs: Since the SX-8R has two vector-pipes one could say (naively said) that they’d do a matrix-multiplication in O(n/2) steps. Generally speaking: O(n/C), where C is the number of CPUs, iterative SMP-approaches scale about O(n²/C).

Obviously that’s a naive approach to explain what they’re doing, but that’s how it’s basically working.

Since I work and live in Düsseldorf I really should look into attending a freakshow in the European Supercomputing Centre. By the way, I had a very splendid freakshow at the vintage supercomputing centre of in Munich two weeks ago. Will put some pictures online later.

And: Sorry for not updating the blog for quite some time, I’m currently busy with a new project (Deploying new clusters for a mobile-phone carrier in Germany). And I’ll miss the SC06. *sigh*
SX-8R press-release

Tech Tags:

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: