Silicon Investor (SI) -- The First Internet Community

STOCKTALK

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.

Politics : Formerly About Advanced Micro Devices -- Ignore unavailable to you. Want to Upgrade?

To: Tenchusatsu who wrote (61241)	6/10/1999 3:12:00 PM
From: Rob Young	Read Replies (1) \| Respond to of 1572859

Tench, How goes it? It's true. Compaq/Digital will integrate four DRDRAM controllers into one 21364 processor die, for a total memory bandwidth of 6.4 GB/sec, almost two-and-a-half times more than that of the 21264 (which is 2.66 GB/sec). [And more than 2 and one half times more than Merced too, correct? ;-)] Microprocessor Report has an excellent article on it, along with a theoretical 16-way Alpha system which has all the processors interconnected in a 4x4 square mesh pattern. So for those who think DRDRAM is an Intel-only solution, better think again. ------ Nah, Digital, now Compaq has known about bandwidth for a long time. "It's the Bandwidth." The theoretical 16-way system? Wildfire ... late next year for EV7 based systems. Just in time to say hello to Merced systems. The more interesting thing for 21364 of course is much improved latencies. Single CPU load to use memory latency of 60 ns.. also they are routing CPU->CPU via onchip Network controller.. 15 ns CPU to CPU with 12 ns load to use.. that is a factor also. 1.5 MB L2... times 16 .. 24 MB routable L2 cache. 4 CPU config 100 ns (remote) latency (i.e. routed, memory hangs off CPUs) .. 40/20 GB/sec bandwidth (routed CPU->CPU and memory) 16 CPUs.. 60-300 ns latency .. average latency 197.5 ns load to use (my calcs, makes one slim assumption but should hold up to close scrutiny). Why the 21364 fixation with DRDRAM? It must have high bandwidth. The CPU -> memory read/write bandwidth is 6 GB/sec. Certainly wouldn't make sense to feed it any less would it? Which brings us to Merced. Quite content with SDRAM. Why? I'll bet Merced does 2.1-2.7 GB/sec read/write bandwidth (somewhere in that neighborhood). Two up and coming CPUs... 21364 and Merced. The 21364 will blow Merced out of the water... cheaper systems and much more powerful. Merced vapor will have to contend with the very cheap and sooner (year-end.. sampling now) EV68. EV68 is 150 mm2 .. 65 SpecInt95 , 100 SpecFp95... 1000+ MHz... think about it .. 150 mm2. 21364 is in a different league. Rob