RV630

RV630 is what I've been most excited about on a personal level, and it's the chip in the new family that we've been most keen on internally at B3D, as the picture of what was coming before R600 launched became clear. Its balance of perf/area seems to be keenly judged on paper.

You get three clusters of shading in RV630, but each is half that in R600. So that's 8 shader ALUs, each 5-way scalar, giving 40 per cluster and a chip total of 120. That's the family's two-way shader core scaling in evidence again, the three chips in the family offering different cluster and ALUs-per-cluster counts to let them scale the architecture in terms of performance.

Just one quad ROP remains, a quarter of what's in R600 as mentioned, but texture sampler unit count is just half. With unit ratios scaling in two directions now, ALU:TEX is variable across all three chips in the family. The tesselator is present and correct, as is UVD, and the output options are exactly the same as they are on R600. That's dual dual-link DVI, HDMI with audio support, good analogue output support, Xilleon and friends.

The external memory bus is 128-bit and the controller supports DDR2, GDDR3 and GDDR4, although there's some wiggle room in terms of how board partners build single SKUs (memory choice, clocks and amount can all vary for HD 2600 XT it seems, for example). Setup rate is 1tri/clk, like R600, and with the tesselator the same as R600 too, triangle rate on Radeon HD 2600 XT (the top RV630 SKU) is the highest it has ever been on a Radeon graphics product.

The 13.3x11.5mm chip is 390M transistors big and it's built using TSMC's 65G+ main node variant as mentioned. Taking a look at the chip in terms of raw numbers, where you can look at things like flops/area/watt/pixel/fetch/filter/whatever, you get a peek at the scaling we talked about with Eric Demers in our recent interview, where each logical processing block on chip is separate from the others in key ways. So looking at R600, RV610 and RV630, we see ratio differences that span the entire family, where no chip is entirely a division of the other in all ways.

For more than half the transistors of R600 for RV630, on paper, it doesn't look like you'll get half the performance. That lets you think about where logic's used in RV630 (and RV610) if you're keen, but it's something we can't spend too much time on here for reasons of getting the article out of the door. We're architecturally rushed, but we'll fill in the gaps on the forums.

So for RV630's area and power consumption, and theoretical performance, you have a chip that's also making a big deal of its motion video abilities, just like RV610, on top of everything else it delivers. Let's take a look at the reference hardware for the first two RV610 and RV630 SKUs out of the door, then.