Summary: NVIDIA's finally pulling the wraps off GeForce GTX 480. The final board features 480 rather than 512 cores, but still delivers blazing performance in today's latest games. See how it and the GTX 470 stack up against the Radeon 5800 and tons of other GPUs inside!
While we did get a really informative deep dive on the architecture back in January, the information blackout since then has left a void that the rumor mill has been more than happy to fill in.
Fortunately that all changes today. All the remaining info you’ve been dying to know the last two months (clock speeds, price points, and most importantly, performance) is about to get answered.
But first, we’re going to give you a really brief crash course on NVIDIA’s new architecture. The most dramatic new feature – for gamers at least – is the new PolyMorph Engine. The new PolyMorph Engine has been designed to deliver breakthrough levels of tessellation performance; tessellation is arguably DX11’s most defining new feature. Rather than handling geometry processing at the front of the pipeline, where it’s traditionally done, NVIDIA has incorporated it directly into the shading clusters found in GeForce GTX 480. NVIDIA refers to them as streaming multiprocessors (SMs). Each SM has its own dedicated hardware for tessellation and other geometry processing units. All told, GeForce GTX 480 has 15 tessellation units total.
GeForce GTX 480 sports twice the number of stream processors as its predecessor, with 480 CUDA Cores compared to GTX 285’s 240. And while its memory interface is reduced to 384-bit, thanks to the use of GDDR5 memory it actually features more available memory bandwidth.
NVIDIA has also made several tweaks to GeForce GTX 400’s ROP subsystem. Each ROP partition now contains eight units, double that of prior architectures. With 6 partitions, this adds up to 48 ROPs total for the GTX 480, 16 more than GT200. The ROPs also boast more efficient compression and higher clock speeds.
Between more ROPS, better compression, and higher clock speeds, this should improve GeForce GTX 400’s AA performance when compared to GeForce GTX 200, particularly when it comes to scaling up to 8xAA.
Going forward, game developers are going to increasingly incorporate GPU Compute into their newest titles, especially now that it’s a part of DX11. Games like Just Cause 2 and Metro 2033 utilize GPU compute today for eye candy effects like depth of field, and of course GPU compute can be used to tackle more realistic game physics. Along these lines, GeForce GTX 400 boasts faster context switching, allowing the GPU to switch between graphics and say PhysX quicker, and can execute multiple kernels simultaneously. NVIDIA even envisions a future where select portions of a scene like reflections may be handled by the GPU with ray tracing.
That’s just the really quick synopsis of GeForce GTX 400’s new architecture though. We discuss it in much greater detail in our GF100 “Fermi” Architecture Overview article, so you’ll want to head there for more specifics.
As you can imagine, there’s been tons of speculation surrounding GeForce GTX 480 and 470 specs. Clock speeds, shader counts, and more have been debated for months now. Today we can finally give you the full specs.
The following chart outlines everything, and how GTX 480 and 470 stack up to GTX 285:
On paper, GeForce GTX 480’s specs don’t look like dramatically improved over GTX 285. At least not what you’d expect from a next-generation product. Key performance metrics like texture and pixel fill-rate, and memory bandwidth aren’t an order of magnitude greater than GTX 285.
When compared against the Radeon 5870, the GTX 480 delivers more memory bandwidth (177.4GB/sec versus 153.6GB/sec) and pixel fill-rate (33.6 Gigapixels/sec versus 27.2 Gigapixels/sec), but is down significantly on texture fill-rate (42 Gigatexels/sec versus 68 Gigatexels/sec). This is surprising, as texture fill rate has traditionally favored NVIDIA’s architectures, so the GTX 480 is a pretty dramatic departure in this regard. NVIDIA does feel that GTX 480’s texture units are more efficient than their previous texture units though.
The other point you’ll no doubt notice is power consumption. With 3 billion transistors the GTX 480 consumes up to 250W of power. That’s considerably more juice than the Radeon 5870’s 188W, even though they both share the same 40-nm TSMC manufacturing process. Fortunately, NVIDIA’s power supply recommendations aren’t that extravagant, but because it consumes more power the GPU will also generate more heat. NVIDIA’s done their best to keep heat at bay, employing multiple heatpipes on the coolers for both the GTX 470 and GTX 480, but it’s something to consider.
NVIDIA says the GeForce GTX 480 will sell for an MSRP of $499, while the GTX 470 will go for $349. That’s a little more than ATI’s Radeon 5850 and 5870, but in exchange NVIDIA says their boards run a little faster, justifying the price premium. We’ll just have to see about that one in the benchmarks. Widespread availability at the retail/etail level will occur the week of April 12th.
While NVIDIA won’t disclose exact die sizes, we know for a fact that the GTX 480 contains more transistors, so it’s naturally going to be a bigger part. Therefore even if NVIDIA and ATI were getting the same yields (which is unlikely given the added complexity of GTX 480), the NVIDIA chip would be more expensive to produce, as you’ll get fewer dies per wafer. This article from Expreview says GTX 480’s die size could be as high as 529 square millimeters based on estimates of die shots of the chip. In comparison, AMD says the 5870’s die size is 334 mm2, so GTX 480 could be 1.58 times bigger than the Radeon 5870.
The GeForce GTX 480 Reference Board
NVIDIA’s reference board design for the GeForce GTX 480 looks downright menacing. Jutting out from the top of the board are four heatpipes, and the board is decked out in black and green, with a dual-slot heatsink/fan unit for cooling. That shiny metal surface just ahead of the card’s fan is actually the GPU heatsink. Part of it is exposed for dramatic effect. Board length is 10.5”, just like the GeForce GTX 285, and right around half an inch shorter than the Radeon 5870.
The plastic duct NVIDIA uses isn’t completely enclosed, but the majority of the hot air from the GPU does exhaust out the back of the case, with additional ventilation provided for the fan on the underside of the board.
Under load the board can get pretty toasty: we observed GPU temps hit 90-93 degrees Celsius on multiple occasions, with our sound level meter registering 53.7 decibels of noise. Under similar conditions and with the same load, ATI’s Radeon 5870 reference board was running at just 67 degrees Celsius while generating just 46.9 dB, so the GeForce board definitely runs both hotter and louder.
The GeForce GTX 480’s noise levels definitely aren’t overbearing, but this is the loudest reference design we can recall from NVIDIA since the GeForce 6800 Ultra days. Running two boards in SLI generated over 61.2 dB and it was definitely noticeable – not GeForce FX 5800 Ultra loud but certainly louder than we’re accustomed too (both GPUs were running over 90 degrees also). In comparison a pair of Radeon 5870 boards running in CrossFire produced 52.4dB of noise.
As we mentioned on the preceding page, you will need both a 6-pin and 8-pin PCIe power connector for the GTX 480, while video outputs include two dual-link DVIs and one mini-HDMI. The board supports 3-Way SLI, and will also support NVIDIA’s 3D Vision Surround with an upcoming driver release in April. This will allow you to run up to three displays in 3D stereo and is also supported by the GeForce GTX 470 as well as GeForce GTX 200 series boards.
The GeForce GTX 470 Reference Board
The GeForce GTX 470 board doesn’t look quite as intimidating as the 480, but underneath its dual-slot cooler lie five copper heatpipes. It isn’t as long as the GTX 480, measuring just 9.5” long from end-to-end, putting it on par with ATI’s Radeon 5850 in length. Both cards also need two 6-pin PCIe power connectors.
We observed quieter noise levels from the GTX 470, but still, in our testing the board ran hotter and louder than the competition from ATI. Under load, our GTX 470 board produced up to 48.5dB of noise, with a GPU temp of 87 degrees Celsius. ATI’s 5850 reference design was only slightly quieter in the noise department, at 47.6dB, but it ran significantly cooler – just 60 degrees Celsius.
Because of this, you’ll definitely want to run both GeForce boards in a case with optimal airflow. A cramped small form factor chassis with little or no airflow wouldn’t be a good idea – at least if you care about the longevity of your card. Aftermarket cooling manufacturers like Arctic Cooling and Zalman are likely salivating at the opportunity to improve on the GeForce GTX 480 and 470 coolers – this will likely end up being a nice little stimulus package for them in much the same way that the Radeon X1900 series was a few years ago.
We’ve been told that the first wave of GeForce GTX 400 series boards will be 100% reference designs, but eventually you will see NVIDIA’s board partners produce their own custom solutions with OC’ed clocks and unique cooling. ASUS is shipping their GTX 480 and 470 boards with their Voltage Tweak software, and we’re sure MSI will probably be doing the same with Afterburner.
For their part, EVGA has a wide range of GeForce GTX 480 and 470 boards planned, including factory OC'ed models with both air and water cooling.
Again, boards won’t be hitting shelves in volume until the week of April 12th. And hopefully street prices will be pretty close to NVIDIA’s MSRP of $349 for the GTX 470, and $499 for the GTX 480. NVIDIA says that tens of thousands of boards will be made available worldwide for the initial launch.
And what about ATI? They’re not making any changes in response to the arrival of GeForce GTX 400. That means no price cuts, and no Radeon 5890 SKU armed with higher clocks. ATI did however remind us that their board partners are now free to come up with their own custom Radeon 5870 and 5850 designs with better cooling, higher clocks, and more memory. To really send the point home, ATI sent over a factory OC’ed 5870 TurboX card from HIS that’s clocked at 900MHz core/1225MHz memory. We’ve included it in our testing today.
3D Performance Testbed
Intel Core i7-980X Extreme Edition @3.33GHz (Turbo Off)
Power Consumption Testbed
Intel Core i7-920
Call of Duty: Modern Warfare 2
Unigine Heaven 2.0 Benchmark
Metro 2033 – DirectX 11
CoD: MW2 – DirectX 9
Crysis – DirectX 10
Crysis – DirectX 10
Far Cry 2 – DirectX 10
STALKER Call of Pripyat – DirectX 10/11
DiRT 2 – DirectX11
Resident Evil 5 – Direct3D10
Bad Company 2 – DirectX
Batman:AA – DirectX 9
HAWX – DirectX 10
Far Cry 2
Resident Evil 5
Sure, the Radeon 5870 did manage to remain ahead of the GeForce GTX 480 in Crysis with very high settings, and the two boards finished in a virtual dead heat under the game’s high graphics settings, but the GTX 480 definitely delivered a better balance of performance across the 10 different games we tested.
As expected, the GTX 480 really shined in DirectX 11 apps. The card ran 16-19% faster than the Radeon 5870 in Metro 2033, and 13% faster in DiRT 2 at 1920x1200. The GTX 480 wasn’t quite as dominating in STALKER: Call of Pripyat, but a 6% win at 1920x1200 is still a victory. One interesting trend we noticed in STALKER, DiRT 2, Crysis with high settings, and many other titles though was that the Radeon 5870 tended to gain on the GTX 480 at 2560x1600. This could be a sign that the GeForce board isn’t as efficient at handling its available memory bandwidth as the Radeon 5870 is.
When compared to its direct predecessor, the GeForce GTX 285 and 280 (look at the GTX 275 scores to see how a 280 would perform, as we’ve found that it essentially delivers 99.9% of the performance of the GTX 280 in games) you can see that the GTX 480 doesn’t always deliver a 2X performance improvement over the prior generation. In fact, with only a handful of exceptions (most notably Far Cry 2, and thanks to DX11 shaders, STALKER) the GTX 275 SLI setup proved to be faster than GTX 480.
This is the same complaint we had with the Radeon 5870 back in September.
However, our biggest complaint with the GeForce GTX 480 isn’t the lack of a 2X performance improvement over the prior generation, instead it’s the enormous amount of heat the board generates. Under load the board gets scorching hot, you definitely wouldn’t want to make the mistake of touching the 480’s exposed heatpipes after an extended gaming session. GPU temps were typically in the 90 degree range – that’s more than 10 degrees higher than where we’d like them to be ideally.
As a result, this ends up creating an enormous hotspot in your chassis that you’re going to have to deal with. You’ll definitely want to pick up a case with a good side vent that can supply fresh, cool air from outside the chassis to the GPU. And although it’s definitely not overwhelming, the GTX 480 runs louder than the Radeon 5870 too. Quiet operation has been a hallmark of NVIDIA’s high-end boards dating all the way back to the GeForce 7900 GTX, so this is definitely a bit of a departure for NVIDIA.
And finally, $499 is a bit of high price to pay considering that the GTX 480 board doesn’t always deliver a significant improvement over the Radeon 5870, particularly at 2560x1600 as we noted earlier.
Still, if you simply must have the fastest GPU on the planet, look no further than the GeForce GTX 480.
Similarly, the GeForce GTX 470 generally outruns its competitor, the Radeon 5850. The showing isn’t quite as dominant as the GTX 480 was over the 5870 though, with the 5850 sweeping all of our performance testing with Crysis, and it also manages to pull ahead of the GTX 470 in games like Bad Company 2, Resident Evil 5, and DiRT 2 at 2560x1600. Again, this could be an indication that the Radeon boards are a little more efficient at handling their memory bandwidth. Technically the 470 has a slight bandwidth advantage over the 5850 (133.9GB/sec vs 128GB/sec).
The GTX 470 really shined against the 5850 in Metro 2033, Far Cry 2, Batman, and Modern Warfare 2 though. It even maintained a double-digit lead over the 5850 in the ATI sponsored DX10.1 game HAWX at 1600x1200 and 1920x1200. In DX11 apps like DiRT 2 the GTX 470’s edge ranged from 7-9% at those same resolutions, and slimmed to 6% in STALKER.
It doesn’t run as loud or as hot as the GeForce GTX 480 either, although we still think the GPU temps are a bit too high. Again, ideally you don’t want to see your GPU temps go above 80 degrees, and some enthusiasts even scoff at that. So NVIDIA’s definitely missed the mark in this regard with both GeForce GTX 400 boards.
One aspect of the GTX 400 boards that’s a little more difficult to cover today is the future. More specifically, how extensively will game developers incorporate DX11 features like tessellation? Clearly with their new PolyMorph Engines, NVIDIA’s betting big that game devs will crank the tessellation up to 11, as seen in something like Unigine’s Heaven benchmark, where the GTX 470 manages to outrun the Radeon 5870 even under the benchmark’s “moderate” tessellation setting. If this is indicative of the direction game developers will take with their upcoming DX11 titles, the future for the GeForce GTX 400 boards looks extremely bright – clearly this is the more future-proof architecture.
The question is, how soon should we expect scenes like Heaven to become the norm in games rather than the exception? It’s certainly challenging for someone like Unigine to put together a fancy benchmark filled with eye candy, but it’s an even greater challenge to deliver that level of visuals and still deliver a game that’s playable. Just ask Crytek.
By the time we do get games that really push the limits of what DX11 can do, the GTX 400 and Radeon 5800 cards may be a distant memory.
If you’re a performance junkie who must have the fastest hardware money can buy, or you need a long term investment that’s going to last you for a few years, and you don’t mind the power and heat concerns, the GeForce GTX 470 and 480 would be the safer bet, provided you can afford to pay the premium NVIDIA’s asking for them. ATI’s once again going for that sweet spot, price/performance gamer that doesn’t necessarily crave the absolute highest frame rates, but instead delivers good performance, good thermals, and more attractive pricing. And of course, for that ultra enthusiast who wants the very best, the Radeon 5970 is still the world’s fastest graphics card (even if they’re nearly impossible to find in stock online).
Going forward, it will be interesting to see which direction NVIDIA goes next. Will they follow ATI’s strategy, and target the $100-$200 segment dominated by the Radeon 5700, or go straight for the OEMs with a Radeon 5400/5500 competitor? Enthusiasts on a budget would love to see a competitor for the recently introduced Radeon 5830.
We’ll also be keeping a close eye on availability. Clearly TSMC still isn’t able to crank out enough 40-nm GPUs to keep up with demand, which has led to higher GPU prices. The GPU price wars from a year ago probably won’t return until this situation is resolved, which is obviously going to disappoint a lot of gamers.
At least the Radeon 5800 series finally has some competition though. Or at least, it will in a few weeks. For the sake of NVIDIA and their board partners, they can’t afford any more delays…
|© Copyright 2003 FS Media, Inc.|