I'm still searching for a Stabuck's in Seattle. :) If anyone knows where one is, please let me know. Just in case - this is a joke. You can't walk two feet without finding a Starbucks or at least some sort of coffee place.
I got to walk the floor a bit more on Wednesday and talked to some interesting companies. I was focusing on interconnects because I want to update my interconnect survey article in the near future.
I started my walkabout talking to the major cluster interconnect companies and what new products they have coming down the pipe.
Myricom has their new 10G product coming any day now. It is interesting because the NICs can be plugged into a 10 GigE switch and they will behave like normal 10 GigE TCP NICs. They can also be plugged into Myricom's switches and they behave like Myrinet NICs (running mx I think). Pretty interesting idea. I'll visit more with them on the last day.
Speaking of 10 GigE Ethernet, I spent a great deal of time wandering the floor to discuss 10 GigE. I stopped by the Chelsio and Neterion booths (makers of 10 GigE NICs) and will stop back today. I also spoke with a new company, Fulcrum Micro and talked to them about a new 10 GigE switch ASIC they are developing. It has great performance (about 200 ns latency) and should be out in switches in Q1 2006. A number of vendors are looking at them for making HPC centric 10 GigE Ethernet switches. We also talked about constructing large fat tree topologies using 24-port 10 GigE switches. Using 24-port switches from their ASIC and two levels of switches, it's easy to construct a 288 port fat tree network that has only 400 ns of latency. They say that the topology runs every port at line speed (sweeeet!!!). Keep an eye on this company.
I also stopped by the Quadrics booth to talk about 10 GigE. they have a new 12 slot 10 GigE switch chassis that uses line cards with up to eight 10 GigE ports on them for up to 96 ports. They are using a 12-port Fujitsu ASIC on the line cards. Each card uses 8 ports for external connections and uses 4 ports for linking to ASICs in a fat tree topology. It is 2:1 oversubscribed, but looks to have very good performance. This is the one of the largest 10 GigE switches I know of except for the huge switches from Foundry and Force10. Pricing should be forth coming, but it looks very competitive.
I stopped by the Clearspeed booth and talked to them about their accelerator card. They have an array processor chip with 96 processing units in each chip. They put two of these chips on a card with some local memory for doing large floating-point computations. They are working on a BLAS and FFTW library for the card so all you have to do is link to their library and can run on the card. They also have an API for writing your own code. They were demoing the card running a simple DGEMM (double precision matrix multiply) computation. The card was getting about 30 GLFOPS sustained performance! (a fast Opteron gets about 8). Also, the card only uses about 25 watts of power!
I've got to run to get to the floor for the last day. Traditionally the last day is kind of slow where people try to snag the last of the swag (the toys) from the vendors. I don't know about the toys, but I want to talk to the various vendors while I have the chance.