BAD HARDWARE WEEK

Cray's Rainier 250TFlops supercomputer, scheduled for 2007/8 is based on AMD's new enhanced core K8L, that should be up to 40% faster than the old K8 one.
System will be based on Opteron equivavalent of Altair Quad core processors named Deerhound. Deerhound chips will naturally use the same 1207-pin socket and will keep inside the same thermal envelope as the new Santa Rosa dual-core Rev F chips, says Vladimir Rozanovich, who is director of AMD's North American commercial business.
Black Widow should be 120TFlops supercomputer sytem with 40TB of RAM and 400 TB of disk space. However, later Rainier's scaling (heat pipe cooled!) should double its peak computing power up to 250 TFlops. The most simple way to do it is to replace dual core Opterons inside with quad core, socket compatible ones, isn't it? Yeah, 125W per processor will not be so easy to cool.
Cray says that the Red Storm design can scale to around 60,000 Opteron cores (11-15K quad cores, or 3K computing nodes?), which should mean Red Storm can scale up to 300 teraflops or more of peak performance, depending on how well the "SeaStar" 3D interconnect, created by Cray to lash Opteron processors and their HyperTransport interconnects together, works.

Some FP instruction references have been found reserved in Rainier performance optimization test tools:

FP_ADD_PIPE 0x40000000
Dispatched FPU ops - Revision B and later revisions - Speculative add pipe ops excluding junk ops

FP_MULT_PIPE 0x40000001
Dispatched FPU ops - Revision B and later revisions - Speculative multiply pipe ops excluding junk ops

Note that AMD K8L = Revision B.
Initially, AM3 socket was planned for the Altair quad cored processors in 3Q2007, with K8L up to 40% more efficient than the old K8, but alas after some problems, AM3 is postponed for 2H2008.
Obviosly, K8L is socket backward compatible, so the only change needed is eventually in BIOS. But, for say AM2+ to AM3 desktop transition, even that would not be required.
Interestingly to note is that if K8L would be announced in 3Q 2007, than that schedule fits nicely with 65nm shrink of K8 core in the first half of the next year. Just not to repeat Prescott failure with simulatenous introduction of new geometry and enhanced core.

And those who don't know history are condemned to repeat it !.
Yeah, K8L will support 48 bit addressing, that means 256TB of memory. Quite enough to keep 250 Tflops Rainier machine running smoothly.


Oooops, K8L clock will run up to 3,5Ghz ?, look at below.
Clock rise of 25% + some 15% boost due to architectural enhancements, that is what gives supposed K8L's 40% performance rise !!
K8L has dynamic I/O bus ! Click on the pictures for higher resolution.