RV670
Rather than stick to tried and tested manufacturing nodes, AMD have gone straight to production with RV670 on TMSC's most cost-effective 55nm node: 55GT. It's essentially an optical shrink of their 65nm production technology and applies to most of their wafer variants, including the GT one that RV670 gets built with today.
The smallest process node ever used for mainstream consumer graphics means that RV670 is a frankly tiny chip, a key thing if you remember the prior page's diatribe, at 192mm2. 666M transistors -- insert your own joke about the number here -- make up that area, which means it's a few tens of millions of transistors less than R600, for less than half the size.
That size and tiny-thing-count only outwards affects I/O, though, since RV670 has all of R600's compute and data processing goodies. The chip size dictates a DRAM bus half the size of R600 at 256-bit (512-bit bi-directional interal ring bus of doom, one presumes), but that's pretty much all that's changed. Major design goal achieved, then? Yeah.
RV670 gets some crucial extras that R600 didn't have, too, to factor into those calculations about where the missing transistors have gone that we know you'll have on the forums. Chip layout changes, the production process and the usual rejigging just won't do for the savvy technologist to bat around these parts.
So why not think about UVD, AMD's fabled video decoder block that can decode hilariously complex video streams so big and so bitty that you wonder how it's only 4.7mm² on 65nm and not half the transistor count of the GPU on its own. RV670 sports it. Why not think about its DX10.1 support, too. That's got to cost you some transistors, right? Well, no, since technically already-released RV6xx parts could get certified as supports DX10.1 and SM4.1, we reckon, but it might be _some_ cost in your book versus R600.
Ah, and then there's the double precision support. Out of the full 320 scalar shader units in RV670 -- the same number as R600 remember - 4/5 of those, the thinner ones, can do double precision IEEE754-compliant math. The rates might not be massive compared to the SP performance (around half speed for ADD and a quarter for MUL, depending on other factors), but it's hardware accelerated support for one of the biggest selling points for RV670 in non-consumer 3D markets. Think about that for a second. When was the last time you saw an IHV build in a significant hardware-level feature to a GPU, that wouldn't in some way benefit the chip executing in a 3D-led context?
The sampler hardware is intact, too, with RV670 making use of the same number of filtered and unfiltered address and fetch units as R600, feeding into and out of the same size caches. That's with all the rest of the main on-chip memories the same size as R600 too, at least that we can tell so far anyway. The latest PCI Express buzznumber, 2.0, gets support as well, with AMD seeing good performance improvements on a compatible host.
With the memory controller sporting GDDR4 support, it's not hard to imagine high levels of performance when equipped with high-speed memory, despite the half-size external memory bus compared to a full R600. 55nm means high clocks, too, with the top-end RV670 product at launch posting bigger numbers than Radeon HD 2900 XT in every area bar external memory bandwidth peak.
That's right, the top-end RV670 product has some serious processing balls. Bigger theoretical processing cojones than an 8800 Ultra, even, when it comes to instruction issue and execution, if the shader core's allowed to do its thing. And all for less than 250 Euros as you're about to see.
If you compare NVIDIA and AMD's latest GPU architectures and implementations to each other, at any given price point, you'll see that AMD have never been short of execution rate. Therefore it's pretty much another fight for efficiency and the hope that the comparative bilerp rate won't bite, given RV670's competition has the same size external bus. Take a look at our G80 and R600 architecture discussions to find out why that's the case.
We'd make a table with columns, filled with numbers, but you're smart enough to know from the following pages alone, where we divulge the initial configurations and speeds, that RV670 at high clocks is to be feared by the opposition. The basic architecture is therefore sound, well known, and considerably well debated.
