The gruelling wait for Nvidia’s first DirectX 11 GPU is finally over. The company has now officially unleashed the GeForce GTX 470 and 480 at PAX East 2010; nearly six months after it first unveiled the Fermi architecture last year.
So what was the hold up? We put the question to Nvidia’s PR manager Ben Berraondo, who told us that “a part of the reason why there have been delays and delays is to make sure that we have enough chips to satisfy demand at launch.”
However, Nvidia is confident that it now has enough 40nm chips in order to satisfy demand. “There will be tens of thousands of cards at launch,” Berraondo assured us, “and we’re confident that during April that there will be enough cards.”
Why are there only 480 stream processors?
It looks as though supply hasn’t been the only issue either. When Nvidia first lifted the lid on Fermi back in September last year, it said the architecture supported up to 512 stream processors. However, the flagship GeForce GTX 480 is launching with just 480 stream processors. What happened to the extra 32?
“Fermi as an architecture obviously was built with 512 stream processors in mind,” explained Berraondo, “and all the GPUs that come off that will be various derivatives. Right now, in order to ensure we balance price, performance, and thermals, we decided to go with the 480 cores on the GTX 480.”
We’re not entirely convinced by this, and suspect that there may well have been problems getting sufficient yield of 40nm chips with the full 512 processors. Whatever the reason, though, we’re only getting 480 stream processors at the moment. Is it possible that a Fermi card with the full count of 512 stream processors could be launched in the future? We asked Nvidia’s senior technical marketing manager Lars Weinand, who confidently told us that “yes, it’s very possible,” although the company said it couldn’t comment on unannounced products.
The raw specs
On to the big question, what exactly do you get in terms of specs on a Fermi card? Let’s start with the GeForce GTX 480, which has 480 stream processors clocked at 1,401MHz, while the graphics core is clocked at 700MHz. The GPU will also be partnered by a buffer of 1,536MB of GDDR5 memory clocked at 1,848MHz, which is addressed via a 384-bit memory interface. The card has one eight-pin power connector, and one six-pin power socket, and it measures 10.5in across – the same length as ATI’s Radeon HD 5870.
Then, a notch down, we have the GeForce GTX 470, which has 448 stream processors clocked at 1,215MHz, while the graphics core is clocked at 607MHz. GTX 470 cards will also come with 1,280MB of GDDR5 memory clocked at 1,674MHz, which will is addressed by a slightly narrower 320-bit memory interface. These cards are a bit smaller than the 480s, measuring 9.5in across, and they only require two six-pin power connectors.
Nvidia’s performance claims
We’ve yet to see any independent benchmark results yet, but Nvidia claims that its Fermi architecture really pulls away from AMD’s Radeon HD 5870 when it comes to hardcore tessellation work. Tessellation is one of the main new features of DirectX 11, and one of its benefits is that it can make edges of objects look much more realistic. Using tessellation, game developers can create much more realistic bumps and edges. Check out the Unigine demo video to see what we’re talking about.
According to Nvidia’s quoted frame rates from the Unigine demo, the Radeon HD 5870 and GeForce GTX 480 are level in performance in some areas, but the GTX 480 pulls away dramatically in areas where there is much more tessellation. Nvidia also claims that the GTX 480’s average frame rate stayed above 30fps throughout the test, while the Radeon HD 5870 dropped well below 20fps at some stages.
According to Nvidia, Fermi also offers superior scaling in SLI mode when compared to its predecessors. The company quoted a number of figures for various games using a pair of GTX 480s in SLI mode, showing frame rates scaling by an average of 90 percent in games including Dirt 2, Battleforge, Alien vs Predator and S.T.A.L.K.E.R. Call of Pripyat.
All the tests were conducted at high resolutions of 1,920 x 1,200 and 2,560 x 1,600, and also had plenty of anti-aliasing and anisotropic filtering enabled. As with all manufacturers’ benchmarks, though, take these with a pinch of salt until you see some independent reviews.
Along with Intel, Nvidia is also making a big deal about ray tracing, and particularly showing that it can be effectively processed on stream processors as well as x86 cores. Fermi cards can process ray tracing via CUDA, and Nvidia has released its own Design Garage demo to show that you can indeed use a GPU to ray-trace in real-time.
Nvidia wouldn’t tell us whether it’s already working with game developers on ray tracing real games, but Lars Weinand did assure us that we “will hear some news about that very shortly.”
Elsewhere in the CUDA world, there’s also good news for those of you who like to use your spare GPU processing power for medical research via Stanford University’s Folding@home project. Nvidia claims that the GTX 480 processes a Villin work unit over 50 per cent quicker than a GeForce GTX 285. Meanwhile, users of the GPGPU video-encoding app, Badaboom, should also be able to expect a performance increase of over 50 per cent, says the company.
Finally, Nvidia has also spotted AMD’s EyeFinity technology for multiple displays, and points out that the GTX 470 and 480 can spread a game over three displays too. Not only that, but they can also do it in 3D at a resolution of up to 5,760 x 1,080 via Nvidia’s 3D Vision Surround technology.
Unsurprisingly, they won’t be cheap though. The RRP on the GeForce GTX 480 is £429 inc VAT, although the GTX 470 is coming in at a comparatively reasonable £299 inc VAT. As a point of comparison, you can currently pick up a Radeon HD 5870 for £308.81 inc VAT.