Monday, April 23, 2007

High performance pointers

HPCwire has a good profile of Woven Systems Inc. and its new high performance 10G Ethernet switch that aims to supplant Infiniband in the data center. The company claims its EFX 1000 will deliver 10G (half what Infiniband can do now) at four microseconds latency (twice or more than IB) while consuming 16W per port.

Woven put dynamic congestion management features into its ASIC. That's a hot new feature, but given that an IEEE effort is still in the works in the area, it is only likely to appeal to users who crave the performance and can't wait for the standard.

Not surprising then, the company's first beta tester will be a national lab benchmarking its system against Infiniband. Woven hopes to get four beta users up and running to shake out the system before it is in production this fall.

Separately, Engadget has a much more spartan, but still interesting write up on the first FPGAs to plug directly into a Xeon processor front-side bus. Xilinx and Altera demonstrated their wares at last week's IDF Bejing. It marks a small start on the road to standard co-processors using the PCI Express Geneseo extensions and, ultimately, Intel's Common System Interconnect which I expect to hear more about this summer.

1 comment:

Anonymous said...

The company claims its EFX 1000 will deliver 10G (half what Infiniband can do now) at four microseconds latency (twice or more than IB) while consuming 16W per port.

Today's "20G" IB cards use PCI-E 8x which has a theoretical data rate of 16 Gb/s (20 Gb/s signal rate). After taking into account PCI-E header overhead, the actual throughput of IB DDR is around 11-12 Gb/s.

And I do not know where there are any academically published IB latency numbers below 3.5 us for IB of any variety.

 
interconnects