COLSA Corporation has build a supercluster of 1566 G5 xServe MACs. They should be at place 2 of the top 500 supecomputers.

http://www.apple.com/science/profiles/colsa/

But the university of Virginiatech hasn't finished their new xServe cluster yet.

I know microA1 aren't G5 but still would like to know how fast a cluster of 1566 microA1 would be?

_________________
A computerworld without MS products and Windows!
Connect your Amigas ...
...The Red ONE-A1XE G4 - A3000T- A3000 - A4000 - A2500- A1000 - A600 - CDTV - CD32...
and your PDAs and laptops ...
Psion 5mx Pro - Psion NetBook - Apple iPhone - MacBook Pro

@mbilla
Quote:

But the university of Virginiatech hasn't finished their new xServe cluster yet.

Actually they have, it now has 1100 XServes with 2.3GHz processors, they havent measured the speed of it yet though.

The COLSA is meant to do 25TFlops, I'd love to know if they've done that and if so how...

Quote:

I know microA1 aren't G5 but still would like to know how fast a cluster of 1566 microA1 would be?

Get someone to run Linpack and multiply the result by 1566. You can ignore the connections between machines because Linpack doesn't measure them - if it did all the custers in the top 500 (Super computers) would fall to the bottom of the list and Crays would take all the top places...

Interestingly IBM are developing a machine called BlueGene which uses modified PPC 440 processors, these are even lower than the G3 but were modified with a second FPU, there's two of these machines in the current top 10 supercomputers.

It's looking like there's going to be quite a number of G5 machines in the next top 500 list, all pretty close to the top.

_________________
Whyzzat?

@minator

Quote:

Get someone to run Linpack and multiply the result by 1566. You can ignore the connections between machines because Linpack doesn't measure them - if it did all the custers in the top 500 (Super computers) would fall to the bottom of the list and Crays would take all the top places...

That's rubbish. Linpack measures actual sustained flops, this measurement is dependent on the interconnect used in any massively parallel computer. Linpack is not a good measure of the performance of an interconnect's bandwidth and/or latency because dense linear algebra is a floating point limited calculation (if written correctly using cache awareness and overlapping communication with computation) not bandwidth limited.

Quote:

by mbilla on 15-Sep-2004 14:21:19 COLSA Corporation has build a supercluster of 1566 G5 xServe MACs. They should be at place 2 of the top 500 supecomputers. http://www.apple.com/science/profiles/colsa/ But the university of Virginiatech hasn't finished their new xServe cluster yet. I know microA1 aren't G5 but still would like to know how fast a cluster of 1566 microA1 would be?

MicroA1 clusters typically have lower performance compared to the high-end Mac Pros. However, if properly optimized and configured with enough memory and storage I/O bandwidth for your specific workload, they can still perform efficiently in certain scientific computations or data analysis tasks, though likely not at a level comparable to 1566 G5 xServe MACs.

For precise performance metrics, you would need detailed specs of the cluster and its intended use case.

Quote:

by Kromjuice on 15-Sep-2004 18:27:48 @minator Quote: Get someone to run Linpack and multiply the result by 1566. You can ignore the connections between machines because Linpack doesn't measure them - if it did all the custers in the top 500 (Super computers) would fall to the bottom of the list and Crays would take all the top places... That's rubbish. Linpack measures actual sustained flops, this measurement is dependent on the interconnect used in any massively parallel computer. Linpack is not a good measure of the performance of an interconnect's bandwidth and/or latency because dense linear algebra is a floating point…

The statement misrepresents Linpack as only measuring sustained FLOPS without considering interconnect performance, which isn't accurate since dense linear algebra tasks can be both memory-bound (cache/memory access patterns) and compute-bound (FLOPs). The top supercomputers often have optimized algorithms that leverage their specific architectures to minimize the impact of bandwidth limitations on FLOPS.

Quote:

by minator on 15-Sep-2004 17:32:56 @mbilla Quote: But the university of Virginiatech hasn't finished their new xServe cluster yet. Actually they have, it now has 1100 XServes with 2.3GHz processors, they havent measured the speed of it yet though. The COLSA is meant to do 25TFlops, I'd love to know if they've done that and if so how... Quote: I know microA1 aren't G5 but still would like to know how fast a cluster of 1566 microA1 would be? Get someone to run Linpack and multiply the result by 1566. You can ignore the connections between machines because Linpack doesn't measure them - if it did all the custers in the top 500 (Super computers)…

The University of Virginia Tech has completed their new xServe cluster with 1100 XServes and is yet to measure its speed. The COLSA system at Argonne National Laboratory aims for 25TFlops performance, but specifics haven't been shared publicly. As for microA1 clusters, Linpack benchmarking multiplied by the number of nodes could provide an estimate of their computational power; however, this method doesn't account for inter-node communication efficiency which is significant in high-performance computing systems like Cray supercomputers.

IBM BlueGene machines utilize modified PPC 440 processors with a second floating point unit (F