scott schrieb:
>> Then again, there's that heat transfer problem...
> 
> Hmm yes, given that fast RAM chips on some GPUs need heatsinks, having a 
> whole GB at L1 speeds is probably going to cook itself instantly.

Then again, maybe not. SRAMs - which AFAIK are the only type of memory 
that would do that job - are freakishly huge compared to DRAMs, so for 
the same amount of memory you'd need about 30 times the die area - at 
the benefit of also having 30 times better heat transfer.

Furthermore, SRAM cells consume virtually no power as long as they are 
idle, and read access costs only the power to drive the address select 
and output lines. DRAM cells, on the other hand, need additional power 
to perform a full rewrite after each read access - and they need to be 
refreshed frequently, which costs power as well.

I'm not sure how much of a difference it makes, but I imagine SRAM to 
consume quite a good deal less power on the whole.


Still - given that caches in modern CPUs appear to occupy something like 
50% or more of the die area, and that's just something like 10 MB, you'd 
have to stack something like 50 chips on top of each other for 1 GB... I 
don't think there's any way to bring 1 GB of memory close enough to the 
CPU to access it at L1 or even L2 speed.

Post a reply to this message