Tiny RV620 and G98 die sizes leak out

Expreview has got their hands on die shots from NVIDIA and AMD's upcoming low-end chips, and while that isn't particularly interesting in itself, they managed to break tradition and leak the chips' dimensions at the same time.

[This piece is the second in a streak of experimental content which uses lists of facts and analysis rather than prose. We believe this should be faster to both read and write, while delivering a clearer message and being easier to follow. So let's see how this goes...]

Data

• AMD's RV620 is 7.0x9.5mm (~67mm²) versus RV610's 7.5x10.5mm (~79mm²). That's ~15% smaller.
• RV620 is expected to be a 55nm shrink of the RV610 (65nm), with added DisplayPort capabilities.
• NVIDIA's G98 is 9.0x9.5mm (~86mm²) versus G86's 10.0x10.5mm (~127mm²). That's ~30% smaller.
• G98 is expected to be a 65nm shrink of the G86 (80nm), with added DisplayPort capabilities but no 128-bit bus support.
• The RV620 is expected to be part of the HD3000 Series, while the G98 would remain branded as GeForce 8400GS.

Process Scaling Analysis

• In practice, Moore's Law doesn't result in twice the transistors per mm² for each full-node migration (i.e. 90nm to 65nm). Scaling for logic-only designs is likely to be 40% or less, and real chips with analogue and I/O will scale even worse.
• TSMC's 55nm process is a 10% linear shrink of 65nm in each dimension, or 19% overall. Interestingly, analogue and I/O are claimed to also scale by that amount.
• TSMC's half-nodes, such as 55nm and 80nm, tend not to really do half the scaling; going from 80nm to 65nm should thus result in slightly better scaling than 65nm to 55nm for digital logic.
• 80nm was also a 19% shrink, but it did not affect analogue and I/O. This means the scaling wasn't as good as 55nm's in practice. Similarly, 80nm to 65nm possibly didn't shrink analogue as much (if at all).

Chip Scaling Analysis

• Smaller chips, such as RV620 and G98, have a higher percentage of I/O and analogue, which makes 55nm an especially attractive half-node.
• The two are very hard to compare, however, as G98 loses 4 ROPs and half of its memory bus width compared to G86. On the other hand, analogue and I/O scaling is likely better for RV620.
• Overall, digital logic still dominates, so NVIDIA would likely have obtained slightly better scaling even if G98 was a straight shrink of G86. At worst, it would have been similar.

Cost Analysis

• Just like RV610, it is unlikely for RV620 to have a cut-down SKU. Shader core redundancy is probably similar to that in R600's (17 ALUs for every group of 16). It is unknown whether there is fine redundancy for anything else.
• It has also been rumoured that there would be a G98 SKU with only 4 TMUs, 8 SPs and 2 ROPs. It is unknown whether that will be the case, but a SKU with at least some coarse redundancy does seem likely.
• Both IHVs are benefiting from the extremely low DDR2 prices and their indirect effect on GDDR3/GDDR4 pricing.
• Assuming no IPC improvements from either side, performance with 800MHz DDR2 should roughly be a wash.

Overall, AMD has a clear cost advantage with the RV620: wafer prices for 55nm should be very close or even identical to those for 65nm, and yields aren't as much of a concern for such an incredibly small chip. The gap isn't as big as it was with RV610 and G86, both in relative and absolute terms, but it remains significant.

Given RV620's support for DX10.1 and its lower cost, it does seem like a splendid chip for the OEM market. We look forward to comparing G98 and RV620's success and capabilities in the coming months.