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.

Technology Stocks : Advanced Micro Devices - Moderated (AMD) -- Ignore unavailable to you. Want to Upgrade?

To: Daniel Schuh who wrote (4141)	8/9/2000 4:07:56 PM
From: pgerassi	Read Replies (1) \| Respond to of 275872

Dear Daniel: Re: Integrated Northbridge In a NUMA based server, the memory access is non uniform but shared. Instead of a number of processors sharing a pool of main memory (every processor has uniform access to any memory), there is local memory and the sharing of local memory of other processors connected by a switch matrix (See LDT discussion from AMD white paper on high scalable MP systems). Each CMP (Sledgehammer) is connected to another 4 CMPs through point to point LDTs. Each CMP has a memory bus to its local memory (Probably DDR or even QDR SDRAM). It gets to the non local memory by requesting it through the matrix of LDTs. Since each CMP could connect to say 4 DIMMs of 1GB each, a n way NUMA SMP system would have a total of n*4GB memory (say 1TB for a 256 way). Each CMP would have a fifth LDT to connect to a southbridge thus, allowing for massive connections to I/O. This is sort of a consistent Beowulf Supercomputer and as such can lead to absolutely huge massively multiprocessor systems like ASCI Red, ASCI Blue, and Compaq Superwildfire boxes for the government. This also reduces the scability problem of traditional uniform access SMP servers that run into a traffic jam on the main memory bus. These traditional servers start having less performance added when the n gets to somewhere between 6 and 12. NUMA SMP servers start having the same degradation with n somewhere around 300 to 1500 depending on the bandwidth ratio of the connecting LDTs versus local memory bandwidth. Pete

To: Daniel Schuh who wrote (4141)	8/9/2000 5:08:59 PM
From: Charles R	Respond to of 275872

Dan, <Integrated northbridge on Sledgehammer? It makes less sense on a server CPU than it does on a desktop, multiprocessors are shared memory, so putting the memory controller on the CPU is questionable. Though the AMD multiprocessor map seems to not be bus-based, so it's possible, I guess.> I am not saying it is. I have a feeling it is based on putting a few pieces of data that I have come across and the rumors that have been floating on the web. So, don't take it as certain. AMD guys I have known have been extremely tight lipped about Mustang and Sledge. Chuck