Single page Print

AMD's Radeon HD 4770 graphics processor


..and first to 40nm
— 6:50 PM on April 29, 2009

The Radeon HD 4850, I think we can agree, is a pretty good graphics card. Since its arrival, the 4850 has set the tempo for much of the graphics card market by delivering strong performance for the price, along with a pretty good total package of image quality, features, and power efficiency.

So here's an interesting question. As you know, chip-making technology regularly advances. What would you do, if you were in the shoes of the folks at AMD, as a follow-up to the 4850?

Their answer, it turns out, is a pretty good one. They've created the RV740, a new graphics chip that's just over half the size of the GPU in the Radeon HD 4850 but packs nearly as much punch. And they've packaged up this new GPU into the Radeon HD 4770, a graphics card that sells for just north of a hundred bucks. Remarkably, this welterweight graphics card may be all you need to play the latest games at buttery smooth frame rates. Sounds pretty good, no? Let's have a closer look.


The RV740 graphics processor
The RV740 is the first graphics chip to hit the market that's manufactured using TSMC's 40-nanometer fabrication process. Most current GPUs are made by chip foundry TSMC, and almost all use that firm's 55nm process. The benefits of a smaller, more advanced fabrication method are fairly straightforward: smaller chips are cheaper to make and tend to consume less power. A new fab process may also enable higher clock speeds, either by allowing transistors to switch more quickly or, as is often the case these days, simply by freeing up additional thermal headroom.


A block diagram of the RV740 architecture. Source: AMD.

AMD's first vehicle for this 40nm process shares its architectural roots with the rest of the RV700-series chips, which power the Radeon HD 4000 series of graphics processors. In fact, the RV740 architecture looks to be only a slightly scaled-down version of the RV770 GPU from the Radeon HD 4850.

As you, erm, might be able to see if you squint really carefully at the block diagram above, the RV740 has eight SIMD shader arrays, each of which contains 16 superscalar execution units. Those units have five ALUs each, so the GPU has a grand total of 640 ALUs, or stream processors, as AMD likes to call them. Each SIMD array has a texture unit associated with it, and each of those can address and filter up to four texels per clock, so the RV740 as a whole can process 32 texels per clock. In terms of both shading and texturing, then, the RV740 has 80% of the RV770's capacity, clock for clock.

However, just like its older brother, the RV740 has four render back-ends. That gives it the ability to produce up to 16 pixels per clock, which it does with the same amount of antialiasing resolve power as the RV770, as well. Perhaps the biggest concession to its lower weight class is the RV740's dual 64-bit memory interfaces—only half as many as its elder sibling. The first implementation of the RV740 makes up for it by using 800MHz GDDR5 memory, which transfers data four times per clock cycle rather than twice, like GDDR3. As a result, the Radeon HD 4770's rated memory bandwidth isn't too far behind that of the Radeon HD 4850.

At 750MHz, the Radeon HD 4770's GPU clock frequency is a little higher than a stock Radeon HD 4850's, allowing it to make up much of the ground lost by the omission of those shader arrays. All told, the 4770 lands roughly in between the Radeon HD 4830, which it replaces, and the 4850 in terms of key specifications. Here's how the three compare.

Peak
pixel
fill rate
(Gpixels/s)
Peak bilinear
texel
filtering
rate
(Gtexels/s)
Peak bilinear
FP16 texel
filtering
rate
(Gtexels/s)
Peak
memory
bandwidth
(GB/s)
Peak shader
arithmetic (GFLOPS)
Radeon HD 4830 9.2 18.4 9.2 57.6 736
Radeon HD 4770 12.0 24.0 12.0 51.2 960
Radeon HD 4850 10.9 27.2 13.6 67.2 1088

Memory bandwidth is the only major category in which the Radeon HD 4830 is superior; the 4770 is otherwise faster across the board. Meanwhile, the 4850 trails the 4770 solely in terms of pixel fill rate, due to the 4770's higher clock speed.

But here's the kicker. At 40nm, the RV740 crams 826 million transistors into a 137 mm² die. Elder brother RV770 has 926 million transistors but occupies approximately 260 mm², nearly twice the size. Nvidia offerings in this price range are based on the G92 GPU, which we've measured at roughly 324 mm² in its 65nm form and 256 mm² at 55nm. (Both incarnations are still out in the wild, so you don't always know which version you'll get.)

To give you a sense of the scale involved, I've included pictures of each of these chips below, next to a quarter for reference. My image sizing isn't exact, but hopefully it's close enough, in combination with the size reference, not to mislead.


RV770


RV740


G92 (the "b" rev or 55nm version)


G92 (the 65nm original)

Teeny, innit? Once AMD and TSMC work out the almost inevitable kinks in this brand-new fab process, the RV740 ought to bring considerable graphics power to the market at very affordable prices. That trend begins, obviously, with the introduction of the Radeon HD 4770.