GeForce GTX 570: Now That's More Like It
Tom’s Hardware reader nevertell in response to my review of Nvidia’s GeForce GTX 580:
“So it's basically what the 480 should have been. Fair enough, I'll wait for the 470 version of the GF110 and buy that.”
If nevertell stuck to his guns, then he’s probably pretty happy right about now. After all, I’ve been playing with “the 470 version of the GF110” for three days now, and have to say I’m genuinely embarrassed for the GeForce GTX 480 cards selling for $450 online. Why? Well, here’s a bit of a spoiler alert: the GeForce GTX 570 is every bit as fast, and in some cases faster. Moreover, Nvidia’s pricing the thing at $349. As soon as I found out about this board, I tried to warn those of you following me on Twitter to abandon any plans to scoop up a discounted GTX 480 ahead of the holidays. The GTX 570 is far more attractive.
Then, before sitting down to write this piece up, I went back to read all 10 pages of feedback on the GeForce GTX 580 launch. The one theme that came up over and over was 6800-series cards in CrossFire. I didn’t have the boards to make that happen at the time (they were hanging out over at Don’s place in Canada), but I do now. And so this time around, you’ll see Nvidia’s GeForce GTX 570 compared to a pair of Radeon HD 6850s yoked together. The AMD cards are $30 more expensive, but if they can also manage to put down better performance, perhaps that’ll make for an even better value.
Not Every Flower Can Be A Rose
With GeForce GTX 580, Nvidia showed us what its Fermi architecture was intended to look like nine months ago.
All 16 Shader Multiprocessors are enabled, yielding 512 CUDA cores, 64 texture units, and 16 PolyMorph engines. And while the back-end still consists of six ROP partitions associated with six 64-bit memory controllers (384-bit aggregate), it runs at a slightly higher data rate, improving memory bandwidth. The graphics and shader clocks are also a fair bit quicker. Lo and behold, a more efficient heatsink and better fan design keep the thermals and noise in check too, even as the GeForce GTX 580 uses just about as much power as its predecessor.
Unfortunately, not every GF110 GPU cut from TSMC’s silicon can grow up to be a princess (that’s not to say the rest of them have to be GeForce GTX 465 toads either, though).
The GeForce GTX 570 is what you’d get if GeForce GTX 480 slept with GeForce GTX 470, and the offspring somehow ended up with GF110 DNA. That is to say, the card’s GPU features 15 Shader Multiprocessors, yielding 480 CUDA cores and 60 texture units. That’s GTX 480’s bosom. It also wields five ROP partitions, a 320-bit memory bus, and 1.25 GB of GDDR5 memory. That’s GTX 470’s derriere.
And of course you end up with the architectural enhancements made to GF110. Mainly, FP16 texture filtering happens in one clock cycle, just as it does on the GF104 found in GeForce GTX 460, and not the two cycles endured by GF100. Nvidia also made improvements to GF110’s Z-culling efficiency. This means the GPU is smarter about discarding pixels that don’t need to be rendered in a scene (because they’re behind other objects), conserving memory bandwidth and compute muscle.