Thursday, August 02, 2007

Next up for 10GbE: The Squeeze

I got a few minutes with one of the Valley's leading computer designers after a panel discussion today. Since it was just one of casual encounters and he was speaking pretty frankly, I'll let him remain anonymous.

He estimated more than $300 million in venture capital has gone into funding companies designing 10G Ethernet media access controllers. "That's bigger than the whole market for 10G MACs," he quipped.

For all the clever design work, the winners will come down to the two or three companies that have superior price/performance while staying under 5W. And success may depend more on software given the complex iSCSI, RDMA, TOE and etc. stacks these chips may need to run. When it's all done there may only be enough market for three or four large companies who sell in enough volumes to make profit in this cost-squeezed sector, he said.

So, now that the 10GbE controllers are about to ship from Broadcom, Intel and Marvell—as well as many startups, it won't be long before the numbers are out and the shakeout begins. Bob Wheeler over at The Linley Group thinks the big OEM design wins may not come down until next year at this time given the software complexity and the relative lackluster demand for 10GbE.


Got some independent test results on one or more of the new 10GbE controllers? Drop a post here or send me a heads up at rbmerrit@cmp.com

3 comments:

Anonymous said...

RDMA is almost already dead for 10GbE, nobody's going this way anymore. TOE tries to survive but most people are not following it anymore apart from Chelsio (and a couple announces from a few vendors without actual shipping later).

TSO/LRO (and IO A/T) won the fight, lots of vendor are going there on the hardware side (Neterion, Myricom, NetXen, IBM, Intel, ...). More important, the Linux kernel is going there too: TSO already available, generic LRO is getting merged soon, while RDMA and TOE implementations got rejected multiple times in the past.

Anonymous said...

I completely agree with the previous comment. Stateless offload (TSO/LRO) has won the battle against TOE, in part because the Linux folks are radically against it for a lot of good reasons (http://linux-net.osdl.org/index.php/TOE), and because the performance improvement, if any, is already too small on today's machines.

For RDMA over Ethernet, I would not be definitive at this point. I would not put my money on iWarp, as it sits on top of TCP so you need a TOE to do it efficiently, and because of many protocol incompatibilities with IB (last byte written last, connection active/passive, etc).

There is a place for zero-copy protocols on Ethernet, but definitively not on top of TCP. The old PM/Ethernet from SCore in Japan or the more recent MXoE from Myricom in the US are good examples.

Anonymous said...

And by the way, iWarp survival in the kernel isn't so clear, the network stack David Miller is kind of bored by it, see http://lkml.org/lkml/2007/8/9/491

 
interconnects