How much die area should a 512 KB L2 cache occupy at 90 nm?

[Eug]

The reason I ask is the PowerPC 970FX is 66 mm², but the PowerPC 970MP is supposedly 154 mm².

That means half a dual-core 970MP is only 11 mm² larger than the 970FX, despite the fact that it has twice the L2 cache (1 MB L2 per core) and presumably extra circuitry to accomodate the two cores on the single shared bus.

This has some people windering if the cache is actually on a somewhat smaller than 90 nm process. What do you think?

 

[Matthias99]

I don't know all that much about the newer IBM PPC CPUs, but it is essentially impossible to have a 90nm process and produce any part of the chip that is smaller than 90nm (bigger yes, by putting bigger gaps between the wires -- smaller, no).

I would think it is MUCH more likely that the 970FX actually has 1MB of cache with 512KB disabled (as is done on the Pentium 4 for lower-end models; except for the EE chips, P4s and Celerons all use the same chips with different things turned on and off). Or that they tweaked the layout to get a higher overall transistor density for the DC variant. For instance, if the layout of the single-core CPU looked like this:

(C = core, M = memory, E = empty)

CCCMMM
CCCMMM
EEEMMM
EEEMMM

And the dual-core looked more like this:

CCCMMM
CCCMMM
MMMMMM
MMMMMM
MMMCCC
MMMCCC

The second layout is only 50% "bigger" (in mm^2), but has 100% more transistors because there is less empty space. So it looks denser if you compare the # of transistors and die size. They also might have used a denser layout for the cache, although this is less likely (since it would be easier to just stick in two copies of the old cache).

 

[Eug]

OK thx. So, if I understand it you can't have a 65 and 90 nm process, unless you're at 65 already. In retrospect, that seems fairly obvious.

I'm pretty sure the 970FX cache is not 1 MB hacked down to 512 MB. That wouldn't make sense since the 970FX is the only shipping G5 in volume. It would just be a waste of die space.

The 970MP has increased cache though, and your explanation does make sense. I do wonder though if the initial 90 nm 970FX was more a combo of 90 nm plus bigger.