SAN JOSE, Calif. The HyperTransport Consortium aims to disclose more details before the end of the year about its plans to drive its interconnect into large high performance computing clusters.
The group will reveal extensions to the HyperTransport physical layer that enable a new class of systems using HyperTransport to link a mixture of processors and systems using both coherent and non-coherent links.
The so-called High Node Count initiative, announced earlier this year, includes encapsulated versions of HyperTransport that ride on top of Gigabit Ethernet and Infiniband, popular cluster interconnects. The approach could reduce the latency of today's clusters, but it requires new software to support a mixture of coherent and non-coherent memory links in a cluster.
"Software is the big question mark for us," said Mario Cavalli, general manager of the HyperTransport Consortium, speaking at a data center event organized by The Linley Group.
Cavalli said several industry players are working on various implementations of HNC, including the University of Mannheim that is "putting the finishing touches on an HNC switch design," he said. "I cannot disclose who is deploying this kind of capability, but I think we have the critical mass to push the software industry to support this architecture," he added.
Before the end of the year, the group expects to make some announcement about physical layer implementations of HNC, he said. "We give the industry an ability to get away from message-passing protocols that are latency laden," said Cavalli
"We are not opening a direct war with established interconnect standards--that's not our intent--but there are opportunities here," he added.
The High Node Count spec aims to help HyperTransport scale beyond the limit of about 100 nodes that can support a coherent link. The spec is finished but currently only available to consortium members.
The group has rolled out a number of technologies based on HyperTransport, including a board-level socket called HTX which never gained traction with developers. Just as HNC aims to lower cluster latency, HNX provided a lower latency alternative to PCI Express adapter cards.
However, the group found the market for a lower latency board alternative was too small to support new products. The same could prove to be the case for the new cluster alternative.
"We are in the process of complementing HTX with other physical layer technologies we will be able to talk about soon," Cavalli said.
HyperTransport was originally developed by Advanced Micro Devices as a processor bus for its Operton server chips. In 2001 it decided to make that bus an open standard, hoping to get broad industry support for using it.
A number of network processors adopted HyperTransport, but few of them ever got much market traction.
Now Intel Corp. is taking a page from the AMD playbook by licensing the high performance, low latency interconnect on Nehalem, its latest server chips. However, so far Intel has opted not to create a broad industry standard around its Quick Path Interconnect.