back to article SiCortex sticks to Summer supercomputer ship

Tied to such a bizarre supercomputer design, SiCortex seemed certain to miss its product ship target. Not so, according to engineer Win Treese, who promised the boxes will arrive this Summer (beta systems out now) as planned, during a speech here at the Usenix conference. Some of you will remember that SiCortex took our TopFLOP …

COMMENTS

This topic is closed for new posts.
  1. Anonymous Coward
    Anonymous Coward

    milli-kilo-mega

    How does the math work? (10W/chip)*(27 chips)*(972slap-togethers) = 262440W = 262kW (kW = kilo-Watt = 1000W). Is there some option to slow the beast down and then consume 600mW (mW = milli-Watt = 0.001 W)? Comes close to a perpetuum mobile... Please be careful about the math and the units!

  2. Anonymous Coward
    Anonymous Coward

    someone got the numbers wrong

    According to SiCorex website, SC5832 "offers 5832 1GFlops 64-bit processors, each dissipating just 600 milliwatts of power". There are indeed 27 chips on a board, and it seems like one chassis contains 26 such boards, so we have: 6 (cores) * 27 (chips) * 36 (boards) = 5832 CPUs. It also seems that one can connect two or more of those machines. I/O controllers including two DDR-2, PCI Express and fabric (the one that interconnects all CPUs) are embedded into a chip. Each board seems to offer 3 PCI Express running at 20Gbps available to the outside world (or another SC5832).

  3. Ashlee Vance (Written by Reg staff)

    Holds Head in Shame

    Sincere apologies there. Mucked up those numbers. Have fixed story. Thanks for that, fellas.

    AV

    www.theduckrabbit.com

  4. Matt Reilly

    Filling in the numbers

    Ashlee -- sorry about any confusion we may have caused on our end.

    To repeat the numbers:

    We've done the engineering necessary to keep the power low. We started out with an engineering budget of about 15 per node (node chip + 2 DIMMs). The node chip aimed at 10W, the rest allocated to the DIMMs. Add to that the overhead for power supply and conversion losses and you get about 18W from the wall for each node.

    The 648 processor systems that we've built are measuring about 1500W from the wall running real workloads. That says that the actual node power is below our engineering budget.

    So, the math again is

    SC648:

    6 processors per node

    27 nodes per module

    4 modules in a system

    4 modules * 27 nodes/module = 108 nodes

    108 nodes * 18 W/node = 1944W

    SC5832:

    6 processors per node

    27 nodes per module

    36 modules in a system

    36 modules * 27 nodes/module = 972 nodes

    972 nodes * 18 W/node = 17496W

This topic is closed for new posts.