A look at NVIDIA's Geforce GTX 580
To be honest, on the surface the GTX 580 doesn't look too different from previous high-end NVIDIA GPUs, and that's not a bad thing. NVIDIA's reference card is an ominous black and green slab that looks like it means business, business, ... and possibly even more business.
The 10.5” board-length (same as GTX 480) GTX 580 has a 2-slot design, it features two SLI connectors for up to Tri-SLI capability, and receives auxiliary power from an 8-pin plus a six-pin connector to satisfy the 244W TDP specification of the GF110/512SP. Again, par-for-the course.
It is worth noting that unlike the GTX 480, NVIDIA's GTX 580 lacks the PCB ventilation cut-out holes seen in the older card. This appears to signal NVIDIA's confidence in the heat output of the GTX 580. Furthermore, the power circuitry looks quite a bit different than what is present on the GTX 580. Since the newer GPU has a few new features concerning power/voltage regulation, this is understandable. As such, the GTX 580 PCB appears to be almost sparsely laid-out compared to previous high-end graphics cards.
GTX 580 with fan shroud removed
Fan shrouds are usually unremarkable, and the GTX 580's is almost no different. However, it appears that more thought has been put into the design of this piece (more than what I saw in use on a GTX 465 anyway), as there seems to be some pieces serving a sort of an “air guide” functionality which could better remove warm air from the card. NVIDIA has mentioned that the angled end of this piece has proved to help in removing warm air as well.
The board's non-GPU heatsink/mounting plate has a design that looks similar to previous GTX 4XX cards. Since the GTX 580 has a lower TDP than the GTX 480, it can be assumed that the thermal management supplied by this part is going to be more than adequate in cooling power circuitry and the GDDR5 memory. The big story here is the fan. NVIDIA is proud of the new fan on the GTX 580, which is supposed to run at a lower pitch and with less vibration than previous GPU fans. Another benefit that will be covered later is a new fan control algorithm that is supposed to “smooth out” transitions between low and high fans speeds. NVIDIA measured noise levels on this new GPU while running Crysis Warhead at 1920x1200, 4xAA/16xAF using Release 260 drivers, the 580 registered at 47 dba, with the GTX 285, 280 and 480 splitting eardrums (hint: sarcasm) at 47.5, 48 and 52 dba, respectively. Great acoustics for a high-end NVIDIA GPU? Where do I sign up?
GPU Heat Sink
Wait, where are the heatpipes? What, does the GTX 580 run THAT cool? No, not quite. But look closer at the chip side and you'll see a vapor chamber in use. Hey, that's pretty cool (points for pun usage please?)! … And I had thought that new high-end GPU's would start sprouting heatpipes ON their heatpipes, with what amounts to a bundle of snakes or tentacles emerging from the board (and I shall dub thee... Cthulhu!).Nope. It looks like NVIDIA went the “simple, elegant, practical” route on this one. The vapor chamber makes use of an evaporator and a condenser to facilitate liquid return, and the air blowing over the heat sink takes care of business as all good heat sinks are supposed to. A preliminary thumbs-up to NVIDIA from the author is in order.
NVIDIA's vapor chamber cooling
More Information on GTX 580:
NVIDIA made it very clear when describing some of the technical aspects of the new GTX 580. The GF110 may have started life as a GF100, but the end product is quite different. Here's their take on just what the GF110 is:
To improve performance/watt, GTX 580 was re-engineered down to the transistor level. We evaluated every block of the GPU, using lower-leakage transistors on less timing sensitive processing paths, and higher speed transistors on more critical processing paths. A very large percentage of transistors on the chip were modified. Through this redesign, we were able to achieve faster clocks with less power.
As a result of this effort, we managed to deliver a product that runs at higher clocks with 512 CUDA Cores and 16 SMs, yet still generates less power than GTX 480.
That sounds pretty good, but is it just PR-speak or might these advancements be able to couple with Fermi's already solid tessellation and geometry performance and raw power to deliver a GPU that pushes pixels but does not warm the room?
SLI, PhysX 3D Vision & Surround Gaming:
NVIDIA is aware that most high-end GPU consumers already know about most of these features, and with the primary focus of the GF110 GPU being performance and power consumption/heat efficiency, I think that NVIDIA has not keyed quite so much on 3D Vision, Surround Gaming and PhysX. Obviously with its advanced architecture, the GTX 580 should be able to run PhysX-enabled games in single-GPU mode with little drop in performance. NV News will have a more in-depth look at GF110 PhysX performance in a later review. As far as SLI is concerned, NVIDIA is proud enough of the GTX 580's SLI performance to toot their horns a little. They claim that the GTX 580 offers 'tremendous” SLI scaling. They showed a graph that showed a bunch of newer games and their scaling performance relative to two HD 5870s in Crossfire. Obviously with the HD 5870s not-very-good Crossfire performance being in direct firing range of NVIDIA, even with most of the listed games exhibiting SLI scaling around 1.4 to 1.6 (with some below, some above), it still looked a lot more desirable than the AMD alternative.
For several generations, NVIDIA graphics cards have featured GPU throttling techniques to keep cards from overheating. With the GTX 580 there are a number of new features that expand on this idea in new ways. There is some new circuitry on the board designed to monitor and adjust current and voltage on each 12V rail of the card.
GTX 580 power monitoring circuitry
The graphics driver also plays a role, helping to dynamically adjust voltages in high-stress situations. Personally, I didn't see this feature coming, although I did recently spot some Chinese Radeon HD 5830's sporting a “Worry-Free” monkier and employing similar on-the fly voltage protection. If this is a measure to increase stability and product life, that's great, I'm all for that. But if it can also help attain more stable overclocks, I'm doubly for that.
If GTX 480 is "The 2010 Tank", is the GTX 580 "The 2010 1/2 Tank"?
512 CUDA Cores. Finally. This (according to the jury that is called the Internet) is what Fermi was supposed to be, right? Along with the jump in CUDA cores is an increase in SM/PolyMorph Engine count, up one to 16, and an increase of four texture filtering units to 64 over the older GTX 480. Memory speed has increased by around 300 mhz (indicating improvements in the memory controller), and core/shader clock speeds have also increased (by 72 mhz over the GTX 480). Pixel fillrate is fast approaching 40 GP/s and texel fillrate is approaching 50 GT/s. This is an astounding amount of tessellation, pixel and geometry-processing horsepower, and appears to further build on the GTX series' 3D gaming and multi-monitor gaming capabilities, where raw power is needed to handle the super-high resolutions. So, YES. This is the Fermi that we wanted. So shouldn't this be heavier, bigger and hotter than a GTX 480? Maybe, but obviously NVIDIA has worked hard to take this 3 billion-transistor GPU and tweak/optimize where they could in order to actually deliver a Fermi refresh that is faster, yet consumes less power and runs cooler.
It's time to kick the tires and see what this “Tank with a Weapons Upgrade” GPU can do in some gaming and benchmarking scenarios.