c0t0d0s0.org

A really neat concept: Plugging a FPGA-Coprocessor right into an Opteron socket. The idea that came into my mind, was a Niagara with an opteron compatible interface as a web/transaction accelerator. Or the opteron for accelerating a Niagara on certain task. Mix and match as you need it. Running on multiple operating systems coupled with RDMAoverHypertransport for communcation. This would only the same idea like the one of DRC taken to an extreme.

What is the rationale behind a connector that only used in systems with a few processors? Surely, HTX is an interesting technology for systems where you really need an really-badass-ultralowlatency interconnect. But to be realistic: the only market for this seems to be HPC. This is a relativly small market , and even in theis market there there are only a small amount of customers who need more than the performance offered by PCI-Express.
Further problems that came in to my mind : Will there be a ecosystem of card manufacturers? Will be there more cards than a few Infiniband cards ? Are the cards really capable to use the advantages or do they use the same chips only with a different interface ? Will the market big enough to justify the development of adapters that take full advantage of HTX.
You must take all this factors into consideration when you sacrifice one of your slots. For a Tier-1-manufacturer its more important to put more FC- or networkcontroller with PCI-E or PCI-X into their systems than a special purpose connector for a special protocol. And with this arguments it makes sense, not to use HTX.
But i'm sure that IBM and HP will sacrify a port only to have a bullet more in their data sheet and i`m sure there will be enough customers who use this bullet to explain their decision against Sun with this bullet point. But at the end we will have a vast amount of unpopulated HTX-Ports. And at last: There has to be a niche for Tier-2 and Tier-3 companies.