Replies to post #71897 on Advanced Micro Devices Inc (AMD)

[chipguy]

chipguy - Do you agree with the general consensus that ZRAM is likely to be used in L3, and if so, how long do you think, realistically, it would be before it's put into a design for production?

I just can't see how ZRAM would have the latency needed for
L2 applications. It is most likely to be used within a high end
server device with a very big L3, AMD's big cache answer to
Xeon MP and IPF. IMO the most likely candidate to be pushing
for that would be Sun. Given how long Sun seems to take to
deploy new machines using regular Opterons I wouldn't even
try to guess when it could be in a position to deploy a high
end Opteron device.

Would we be right in assuming that a good bit of work has been done checking it out and getting the ball rolling if they already have the agreement?

Hard to say. It would take a lot of embedded DRAM know-how
to be able to architect a ZRAM based L3 with suitable SRAM
like characteristics. In addition AMD would need to do a lot
of due diligence on the technology with a fleet of special test
chips. DRAM hates electrical noise and high temperatures,
something prevalent on performance MPUs. You really don't
want to ship a high end server MPU that you weren't entirely
convinced was utterly bullet proof to OEMs.

[chipguy]

Do you agree with the general consensus that ZRAM is likely to be used in L3

If it is going to be used *at all* in AMD high end processors
IMO it will be for constructing huge L3s.

if so, how long do you think, realistically, it would be before it's put into a design for production? Would we be right in assuming that a good bit of work has been done checking it out and getting the ball rolling if they already have the agreement?

It probably would be used for a high end, big cache variant
of a 65 nm quad CMP MPU at the earliest. That is assuming
AMD thinks there is a big enough market for such a device
and doesn't run into any snags with the technology. Keep in
mind that 1) it would be a separate product line/mask set
than the rest of the Opteron/A64 family so it would operate
on its own cost/yield curve dictated by much smaller volumes,
and 2) OEMs get more conservative for bigger systems so
they would probably want to kick the tires of a ZRAM based
device for quite a while.