Bringing the best out of Fermi
Fermi refined
NVIDIA brought in a new high-end GPU architecture last month. Known as the GeForce GTX 580 1,536MB, it uses fundamentally the same Fermi core from the previous champ, GeForce GTX 480, but thorough re-engineering provides the GTX 580 with an extra 15 per cent performance across a wide range of games, achieved with no commensurate increase in power-draw. The GeForce GTX 580, then, is the fastest single-GPU card around, and putting two together in SLI results in a near-doubling of the already-high frame-rates.
The next logical step for NVIDIA is to use the GTX 580's revised architecture (GF110) to introduce a cheaper GPU to the fold. This is exactly what it is doing with the arrival of the GeForce GTX 570 1,280MB card, released today.
Putting the new launch into context, here is a four-GPU table that highlights the best graphics cards from NVIDIA. The left-hand duo take in the new GF110 design, while the right-hand pair are the performance leaders from March 2010.
| GPU | NVIDIA GeForce GTX 580 1,536MB | NVIDIA GeForce GTX 570 1,280MB | NVIDIA GeForce GTX 480 1,536MB | NVIDIA GeForce GTX 470 1,280MB |
|---|---|---|---|---|
| Transistors | 3.0bn | 3.0bn | 3.0bn | 3.0bn |
| Die size | 520mm² | 520mm² | 529mm² | 529mm² |
| Fermi revision | GF110 | GF110 | GF100 | GF100 |
| General clock | 772MHz | 732MHz | 700MHz | 607MHz |
| Shader clock | 1,544MHz | 1,464MHz | 1,401MHz | 1,215MHz |
| Memory clock | 4,008MHz | 3,800MHz | 3,696MHz | 3,206MHz |
| Memory interface | 384-bit, 1,536MB GDDR5 | 320-bit, 1,280MB GDDR5 | 384-bit, 1,536MB GDDR5 | 320-bit, 1,280MB GDDR5 |
| Memory bandwidth | 192.4GB/s | 152GB/s | 177.4GB/s | 133.9GB/s |
| Shaders | 512 | 480 | 480 | 448 |
| GFLOPS | 1,581 | 1,405 | 1,345 | 1,089 |
| L2 cache | 768KB | 640KB | 768KB | 640KB |
| Texturing | 64ppc
bilinear 64ppc FP16 |
60ppc
bilinear 60ppc FP16 |
60ppc
bilinear 30ppc FP16 |
56ppc
bilinear 28ppc FP16 |
| ROPs | 48 | 40 | 48 | 40 |
| ROP rate | 37.1 | 29.3 | 33.6 | 24.28 |
| GTexel/s bilinear | 49.4 | 35.1 | 42 | 33.99 |
| FP16 rate | 49.4 | 35.1 | 21 | 17 |
| Board power (TDP) | 244W | 219W | 250W | 215W |
| HDMI | 1.4a | 1.4a | 1.4a | 1.4a |
| Retail price | £399 | £289 | £300 | £185 |
Analysis - GTX 570 vs. GTX
480: fight
The new chip uses almost all of the GF110 die. In NVIDIA-speak this translates to four General Processing Clusters, 15 PolyMorph engines (out of a possible 16) and an identical number of Streaming Multiprocessors that each hold 32 CUDA cores. In this respect, the GTX 570 is identical to the GTX 480.
NVIDIA, too, has made the same architecture trade-offs as with the GTX 480/470. This means that core, shader and memory clocks are reduced; the memory interface is cut from 384 bits to 320 bits; shaders see a snip from 512 to 480, and raster back-ends fall from 48 to 40. It doesn't take an M.I.T. graduate student to figure out that the GTX 570 is going to be between 10-20 per cent slower than the range-topping model.
The speed reduction means that the really interesting comparison isn't 570 vs. 580. Rather, it's how the new GPU matches up against the GTX 480. The numbers show that the GTX 570 loses out in terms of memory bandwidth - a faster speed isn't enough compensation for a bus-width snip - and falls foul of the GTX 480's 48 raster back-ends when looking at the pure ROP rate and bilinear fill-rate. But where the GTX 570 hits back is with respect to more shading power - down to a higher shader speed and identical number of CUDA cores - and significantly higher FP16 processing rate.
What this tells us is that the GTX 570 vs. GTX 480 battle, if we can call it that, will be very close. We expect the GTX 570 to steal a march if titles use lots of HDR rendering, while the GTX 480 will nose ahead if the game engine is partial to lots of memory bandwidth, especially at higher resolutions, where the 570's smaller frame-buffer may impose limitations.
Greener power credentials, and a look to the future
Based on the same vapour-chamber cooling technology as GTX 580, the GTX 570 should be quieter than GTX 480. What's more, the lower TDP generally matches that of the GTX 470.
We reckon that NVIDIA's GTX 570's core and memory frequencies have been set with one eye on the probable performance of AMD's upcoming Radeon HD 6950 part. A 732MHz core speed and associated 1,464MHz shader clock is fairly aggressive for an NVIDIA reference part. The memory speed, too, is pretty close to GTX 580's.
Projected to hit retail shelves at £289, thereby just undercutting the cheapest GTX 480s, the new 5-series product stack is slowly taking shape. The numbers match up well with GTX 580, so let's take a look at the card and then head on over to the benchmarks.
