As you may know if you follow these things, AMD's Radeon HD 2900 XT graphics processor, also known as the R600, wasn't exactly a rousing success in all areas. The chip brought big gains in performance, features, and image quality, but it was late to market. When it arrived, it was fairly power-hungry and simply couldn't match up to the performance of Nvidia's top GeForce 8800 GPUs. Worse yet, the gaping hole between the $149 and $399 price points in the Radeon HD lineup left many enthusiasts wanting more.
Fortunately, graphics chips have relatively short lifespans, and the regular introductions of new chips bring ample opportunity for redemption. Today is one such opportunity, as AMD pulls the curtain back on its latest creation: a chip modeled on the R600 architecture that promises similar clock-for-clock performance, yet is under half the size and draws a fraction of the power. Better still, the new graphics cards based on this chip, the Radeon HD 3850 and 3870, look to be very affordable. That's progress on all the right fronts, and it sets up PC gamers beautifully at a time when we're seeing more good games released at once than, well, maybe ever.
Here's the question: is this progress sufficient to allow AMD to catch up with Nvidia's just-introduced and dauntingly formidable GeForce 8800 GT? Have the AMD guys engineered an even better stocking stuffer? Perhaps. Keep reading for some answers.
The RV670 GPU
If you'll indulge me, I'd like to start out with the chip geekery, and then we can move on to the graphics cards themselves.
The subject of our attention today is the RV670 graphics processor, which is basically a revised version of the R600 that's been converted to a smaller chip fabrication process and tweaked in numerous ways. The R600 itself was manufactured on a 80nm fab process, and it packed roughly 700 million transistors into a die area of 408 mm². That's one big chip, and it was part of a growing trend in GPUsor shall I say, a trend of growing GPUs. The things were adding transistors faster than Moore's Law allows, so physical chip sizes were rising like crude oil prices.
This latest generation of GPUs is throwing that trend into reverse. Nvidia's G92, which powers the GeForce 8800 GT, shoehorns an estimated 754 million transistors into a 324 mm² die (by my shaky measurements) via a 65nm process. The RV670 goes even further; its 666 million transistors occupy only 192 square millimeters, thanks to a 55nm fabrication process. The move to smaller chip sizes means several things, including cheaper chips, lower power consumption, and less heat production. In the case of the R670, AMD says the advantages of a 55nm process over a 65nm one are mainly in size and power. They can fit 30% more transistors into the same space with a 10% reduction in power use, but with no real increase in transistor switching speed over 65nm.
Here's a quick visual on the chips, to give you a sense of relative size.
You can see quite readily that the RV670 is a fairly small GPUmore of a wide receiver than a fullback like the G92. Yet the RV670 packs the same mix of 3D graphics processing units as the R600, including 320 stream processors, 16 texture units, and 16 render back-ends.
As you may have noticed, though, the RV670's transistor count is down from the R600, despite a handful of new features. The primary reason for the reduction in transistors is that AMD essentially halved the R600's memory subsystem for the RV670. Externally, that means the RV670 has a 256-bit path to memory. Internally, the RV670 uses the same ring bus-style memory architecture as the R600, but the ring bus is down from 1024 to 512 bits. Thus, the RV670 has half as many wires running around the perimeter of the chip and fewer ring stops along the way. Also, since the I/O portions of a chip like this one don't shrink linearly with fabrication process shrinks, removing half of them contributes greatly to the RV670's more modest footprint.
Of course, the obvious drawback to this move is a reduction in bandwidth, but we noted long ago that the R600 underperformed for a chip with its prodigious memory bandwidth. AMD says it has tweaked the RV670 to make better use of the bandwidth it does have by resizing on-chip buffers and caches and making other such provisions to help hide memory access latencies. On top of that, higher speed memories like GDDR4 should help offset some of the reduction in bus width.