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: Joe NYC who wrote (4186)	8/9/2000 6:24:33 PM
From: EricRR	Read Replies (1) \| Respond to of 275872

Anyway, if each northbridge has to be connected with every other NB, it seems that the number of connections would grow exponentially. On the other hand, if there is some kind of ring of the LDT connections, isn't it going to increase latency? Yes and yes. LDT is flexible. One can either design a symetric cube with many connections, or "daisey chain" the cpus. This is a standard supercomputer tradeoff.

To: Joe NYC who wrote (4186)	8/9/2000 8:21:54 PM
From: pgerassi	Respond to of 275872

Dear Jozef: Re: LDT transmissions Of course this will increase latency the further one needs to transmit. There are three ways to decrease this. The first is obvious, increase the number of LDT links for each CMP chip. The old hypercube arrangement would require 9 LDTs, a pair for each dimension (positive going and negative going projective (0 is connected to n-1 (for n units a side))), and one to the southbridge. The second would be to add additional LDT links to each southbridge. Two additional LDTs in each southbridge could allow a modified hypercube arrangement (the CMPs are connected as a 2-D matrix and the southbridges connect the other 2 dimensions). The third way is to add LDT hubs say with say 8 LDTs each. The page shown would have two connections to a hub above the page and two connections to a hub below the page (both hubs would be in the same plane as the CMPs). Each hub would do the same to another plane with the same CMP arrangement. Thus another modified hypercube is possible. To reduce latency even further, a greater hub to CMP ratio could be used. Most high performance simulations however, have little long range data passing required. Most non local accesses are to neighbors anyway. This is why Beowulf supercomputers are well suited to most problems requiring large computational loads as well as other "cell" based topologies (this means that the MP subunits are all identical except for the number of connections and the routing of those connections). Pete