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: fyodor_ who wrote (79882)	5/10/2002 8:55:06 PM
From: pgerassi	Respond to of 275872

Dear Fyo: There is only one HT link in the host CPU/GPU/DSP. Slot one is always the primary module. It can be any size. The second two slots are driven from the output side of the first module. You also forget that the first module may contain more than one chip. A three die first module still has two output ports (actually four, but two are unused). All modules could be the same size, or they could vary in size from 1 chip to 32 chips. The only thing required of the slot is that it have three HT ports, one master and two slave, if that nomenclature makes it easier, plus standby (maintain memory with internal refresh) and main power. A two slot MB or peripheral card just has a chip integrated making slot one always present. Thus default memory would be enough for test, repair, etc. and the two expansion slots could just extend memory beyond the default. Remember, each die cotains all of the glue required and the best module and overall layout looks like a symmetrical binary tree with the external HT links coming from the root and the closest to the root unused slave HT links. Also each die is exactly the same except from DRAM array size. Thus the modules and the MB control the layout and where the die is placed in that layout determines its start address and latency penalty. The initial setup "walks the tree" and sets the two neeeded parameters, the address of the die and the threshold address between a command being sent down the first slave HT link or being sent down the second (any matching the die's address are acted upon and not sent to the slave links). Any packet sent up from the slave HT links is forwarded to the master HT link. So a 128MB of local memory for a GPU is built onto the AGP card with two expansion slots. The maximum local memory is over 31GB, if each expansion modules contains 127 128MB dies. Given the normal limit of either 15 or 31, the maximum GPU local memory could be as high as 3-7/8ths GBs or 7-7/8ths GBs. An amount far larger than any current GPU card even for high end workstations. Thus any GPU could have anywhere from 128MB and up in 128MB increments without any changes to GPU, AGP card or memory dies making all three able to be commodities. Pete