Good move for AMD
Not so sure about an end run but AMD was wise to buy Seamicro and consumers will benefit.
As last week was winding down, Advanced Micro Devices took control of upstart server maker SeaMicro, and guess what? AMD is still not getting into the box building business, even if it does support SeaMicro's customers for the foreseeable future out of necessity. Further: Even if AMD doesn't have aspirations to build boxes, the …
I've always wondered why they didn't make Opterons with many Hyperchannel links many years ago, so that system-builders could take Hyperchannels through a little buffer chip between boards, or fruther using optical technology. Why did they stop at three, while they watched people building clusters using slower and less integrated technologies?
Only obvious with hindsight?
Cray did something like that at least twice already. The XD1 uses ASICs to extend HyperTransport into a switched fabric to link 144 processors together into a single system image. The XT3 and successors connect the CPUs to the SeaStar interconnect via HyperTransport.
The processors can connect to each other without any glue logic but that consumes HT links quickly. One can only add so much to the die and packaging before it gets too expensive.
SeaMicro was in trouble. They had a hard time selling their boxes for $250K a pop. No IT manager wants to make such a large commitment in one shot - they are used to an incremental scaleable approach. This is a good outcome for them, the alternative was a fire sale.
AMD did acquire them for their technology which is interesting - however this only works in the scale-out big-data space which is small considering the rest of the market where Intel servers play. AMD might recover their investment in this acquisition but this hardly helps them compete against Intel in the other segments.
Sounds to me more a story for Rory Read and Lisa Su to show their board that they are doing something.
The only sensible thing to do with that seamicro IP is to glue it to ARM tech - OTOH i'm not sure their little torus is anything interesting compared to existing implementations - the whole managed box thingy sounds good but I'm afraid to find out the real issues / limitations.