A visit to NCSA
A well-connected friend sent along these pictures of his visit to NCSA, the National Center for Supercomputing Applications. He got to take a look at some amazing supercompting clusters based on some bleeding-edge tech, including Intel's new 64-bit processors. Some highlights:
- The PIII cluster is at 350+ processors in 1U boxes from IBM, as the pictures show. Each box has two 1GHz PIII processors. That cluster will be one teraflop capable.
- The Itanium cluster is comprised of two-processor boxes. There are some four-processor boxes there, too. The processors are 733 and 800MHz Itaniums. The Itanium cluster will also be a one teraflop cluster shortly.
- On Atlas, the Itanium achieves 3-4 instructions per clock (IPC), but that's simple code, not unpredictable code with lots of conditional branches. Still, it's impressive. As compilers improve, the Itanium's performance on in certain areas could potentially double, especially when using SSE2, so long as they work out the kinks.
- The Itanium cluster is using a Myrinet interconnect, which provides ~2+ gigabits of bandwidth. Yes, the interconnect is messy at the moment, but they are constantly changing configurations, and didn't expect someone to be there to marvel at their work.
- There is a picture of the NT supercluster there. On that machine, the latency of Myrinet is about 20 microseconds.
- Note the Origin Supercomputer. Each Origin has an LCD readout on it, which displays the status of each processor on the machine.
- We believe the Sun box is used for storage.
Mighty impressive, huh?