News Focus
News Focus
icon url

wbmw

09/18/07 2:12 PM

#48587 RE: ChipGeek #48582

Re: Keep staring at that die, though; there are clues there for the observant...

Maybe we could get Hans....

No wait, he just does stuff to promote AMD.
icon url

Saturn V

09/18/07 4:05 PM

#48623 RE: ChipGeek #48582

Keep staring at that die, though; there are clues there for the observant...

I shall pull a deVries analysis by inspired guesswork.
Bottom: Common L3 cache(or maybe L2 ?)
Middle Centre: 4 vertical cores
Top: CSI ( Speed path interconnect)buffers - 20 wide output
Right & Left : Memory Buffers

Need a higher resolution photograph for anything better.

icon url

Saturn V

09/19/07 12:23 AM

#48714 RE: ChipGeek #48582

Keep staring at that die...

Nehalem speculations in lieu of HansDeVries:
It is not clear if the 4 core chip has 1 or 2 CSI channels[20 bit wide]. I think that the chip has two independent memory paths,one on the right and one on the left.

Performance Scalability: Eight core chip will be obtained by mirroring the chip vertically and doubling the CSI and memory channels.Two core chip will be obtained by cutting it in half, with half the memory buffers, and half the CSI resources. Obviously the control logic in the central vertical path cannot be neatly sliced, and most of it will be retained. So external memory and interconnect bandwidth scale with the number of cores. So performance scales linearly with number of cores.

Power Scalability follows by each core having its independent clock, which can be shut off independently,as in Barcelona.However if a thread is shut off, the power savings are small since the L1/L2/ALU still need to be powered, but half the register file can be shut off.

Unknowns: Die Size ? Power per core ? Performance ?