Review: Nvidia’s F104 - GTX 460 mini-fur-me

Posted by Jon Peddie on July 19th 2010 | Discuss
Categories: Hardware Review
Tags:

A $200 board that packs a lot of wallop

One of the fastest airplanes ever made was the Lockheed F104 clocking in at 1.7 mach with a 48k ft/min climb rate it was  called a missile with a pilot in it. We’ve been testing another F104 - the new F104-based Nvidia AIBs, the 1GB GTX 460 and the 768MB GTX 460 both units configured with GDDR5 memory. These are consumer derivative versions of the famous Fermi chip, minus the super computer parts like EC memory management, smaller caches, and a larger number of texture units per FP unit. The new GPU also has more head room for over clocking.

Figure 1: Another very fast F104 (source Lockheed)

Designated the F104 GPU by Nvidia, the GeForce GTX 460 AIB ships with 336 CUDA cores and 56 texture units; two memory configurations are available:

Clock speeds between both boards are identical. The graphics core clock is 675MHz, while the CUDA Cores run at 1350MHz. Memory speed is 1800MHz.

The GPU has seven of what Nvidia is marketing as PolyMorph Engines for tessellation. Nvidia likes to point out that their competitor’s architecture is only outfitted with a single tessellation engine.

Some of the early speculation on the (unnamed at the time) 460 was Nvidia would simply cut the GeForce GTX 480 GPU in half, however, Nvidia surprised folks and instead made some changes to the design of their Shader Modules (SM) for GF104, resulting in a product that’s Fermi-based but also notably different from GF100. The GTX 460 is equipped with 336 CUDA Cores (48 per SM) in order to hit the performance, and power requirements of the Performance gaming segment.

Figure 2: Comparison of the relative size difference between a GTX480 and a GTX 460 GPU

Designed for:

Enthusiasts

The Gamers’ Sweet Spot

CUDA Cores

480

336

Polymorph Engines

15

7

Texture Units

60

56

Board Length

10.5”

8.25”

Max Board Power

250W

160W

MSRP

$499

$229 ‐ $199

In order to feed the CUDA cores with data, Nvidia included four dispatch units per SM. As a result, two instructions can be dispatched per warp, for a grand total of four instructions per clock per SM. Along similar lines, the number of special function units (SFUs) and texture units is eight per SM.

Figure 3 : CUDA cores in a shader module

GT in gaming performance, and speedups of up to 4.5x were found with PhysX applications.

The tests

The GTX 460 with 768MB frame buffer goes head to head with a ATI Radeon HD 5830, a product that for a similar price delivers a larger frame buffer (very important both for today’s games and even richer, more detailed DirectX 11 games to come); However, although the ATI AIB supports ATI’s Eyefinity and can drive three displays, two are DVI and one is either a Display Port (DP) or HDMI. And to use the DP the user will have to get a DP to DVI adaptor (~$25.) However, the GTX 460 only can drive two monitors, so the additional cost for a third monitor adaptor is not too burdensome. The Radeon 5830 also offers 7.1 HBR audio pass through.

The good news is there is no single test to use in evaluating a graphics AIB. The bad news is there no single test to use in evaluating a graphics AIB. It used to be if you wanted a rock solid test you used the application you wanted to run on the graphics AIB, in this case the game. Synthetic benchmarks like 3DMark, and Unigine’s Heaven benchmark can give you a very good idea of an AIB’s general performance and solid idea of its tessellation capabilities but the actual experience of your favorite application may vary.

That means folks like us have to run a lot of tests to get a feeling for how an AIB really performs. The problem is the testing is complex, compounding, and can be confounding due to the range of control that most games offer. Of course you can vary resolution, but you can also turn anti-aliasing off, or on, and when on you can choose between 2x, 4x, or 8x samples per pixel. You can turn on and off anisotropic filtering, some games let you turn on or off tessellation, and/or vary it by complexity (normal, extreme), and some games will run different lighting conditions. The combinations of conditions are staggering. Add to that Nvidia and AMD come out with new drivers almost weekly.

To make life a little easier on us (not the AIB) we in general choose the most sever conditions to test the AIBs. We ran three boards through a suite of tests: An AMD Radeon 5830 1GB, an Nvidia GeForce GTX 460 1GB, and a GTX 460 768MB. We ran four tests: Aliens vs. Predators, Resident Evil 5, Stalker Call of Pripyat, and Unigine’s Heaven at three resolutions: 1680x1050, 1920x1200, and 2560x1600.

We ran the tests in the most extreme conditions we could and discovered as good as these mid-range priced AIBs are, they can’t quite drive a ultra high-resolution screen (i.e., 2560x1600) at useful frame rates (i.e., >25 fps) so for clarity we have not included those scores in this report.

The results are shown in the following charts.

Figure 4: Benchmark suite results

For the tests we ran, on average the GTX 460 0.8GM board was 21% higher in FPS than the Radeon 5830, and the GTX 460 1GB was 32.6% and that seems significant. As the data shows the extreme conditions knock the AIBs down below 30 fps and so the user will have to be careful when selecting which features to turn on.

Extra memory

As the DX11 data shows, the added cost for a GB of memory is not clear. The difference in cost to add 232 MB of memory (~25%) will cost $30 or 15.1%. However, the average performance improvement was only 9%. In Stalker COP we saw improvements as high as 17.4%, but all the other scores were 10% or less.

Pmark

To calculate the Pmark which uses Price, Performance, and Power we have in the past used Futuremark’s Vantage test. However, Vantage only tests for DirectX 10. Since there is no standard synthetic benchmark, we took the average of all the DirectX 11 scores we observed (in frames-per-second) and used that for the Performance component of the evaluation. This gives a good overall expectation of the AIBs in general. Then for those who want to find something they can point to show their superiority they can pick the particular test that works for their objectives.

Figure 5: Pmark for $199 to $229 AIBs

The Nvidia GTX 460 768MB AIB is the clear winner in the $199 to $229 category. The raw data that went into the Pmark calculation is shown in the table.

 

HD 5830

GTX 460 768MB

GTX 460 1GB

Pmark

0.64

0.88

0.80

Power (W)

175

150

160

Price ($US)

$199

$199

$229

Performance (Ave Dx11 FPS)

22.16

26.34

29.22

Pmark

The PMark

Where:

  • Performance is expressed in 3DMark Vantage score
  • Price is expressed in US dollars
  • Power is expressed in watts of the AIB

About the power consumption. Nvidia’s published power usage (which is the figure we used) is 150 to 160 watts. Several sites have said they measured the 0.8GB GTX 460 at 160W. This too is a variance of manufacturing like over-clocking range. We monitored the power of the system without an AIB in it to get the “tare-weight.” Then we put the board in and let it just sit with only Explorer running, that drew ~ 25 W. The we ran Heaven the most demanding program and found the power consumption (from the tare value) was 100 to 135 W - so much for empirical measurements.

Comment on the Pmark

The Pmark is designed to be a general consumer guide, and should be useful for people who want a balanced view of a graphics board. . An analogy would be the buyer of an automobile. For an average buyer with a limited budget and life style that involvs children, groceries, and pets, an automobile would not be purchased on its horsepower rating. That may be factor, but gas mileage, and cargo space would probably get equal if not greater weighting. However, for a more well off person, with either no kids, or the budget for an expensive second (or third car) performance may be the main criterion and so he or she will buy a high-powered sports car and look at gas mileage as a secondary issue.

The same holds true for the purchase of an AIB. For people with limited budgets other criteria may come into consideration such as power consumption (we all want to help the environment) and the subsequent noise a big fan will make. This is the buyer the Pmark is designed for. For enthusiasts where performance is the only criteria then the Pmark has little utility.

DirectX 10 results

There’s still mostly DirectX 10 games being sold and played and so a DirectX 10 score (Resident Evil 5) is shown for comparison.

Figure 6 : Resident Evil 5 DirectX 10 standard and over-clocked scores

Looking at the Resident Evil Test 1680x1050 4X AA (GTX 480 768) the test appears to be CPU bound – Nvidia observed a FPS score of 91, almost 10 frames higher, on Gulftown as opposed to the Phenom II – we confirmed these finding when we ran the test on the Gulftown based system as well.

As might be expected the new gen AIBs do very well on the older games, Direct X 10 as well and DX 9.

Over-clocking

We ran all the boards through all the tests with the GPU and memory clocks turned up. And although there was some improvement in FPS it wasn’t significant enough, nor universally reproducible to warrant a report. The GTX 460 has significant head room and can be cranked up to 850+ MHz but it’s an individual thing – that is, some parts may come out capable of that and some won’t so it can’t be guaranteed- it if could Nvidia would offer the part at those speeds.

I want to hear it too

Besides gaming, users who buy AIBs in this segment will use their computer for other tasks such as photo-editing, MP3 music, and watching movies with the computer possibly connected to a large screen HDTVs Nvidia provide audio pass-though via HDMI for such users. To satisfy the needs of this audience, the GeForce GTX 460 also features enhanced audio support over HDMI, this includes bitstreaming support for Dolby True HD and DTS-HD Master Audio over HDMI.

What do we think?

There one heck of a lot of bang for the buck in Nvidia’s $199 to $229 F104-based AIBs: good gaming scores, good overall capabilities, modest power consumption, and quiet running. The AIBs can be ganged up in SLI mode (but they have to be of the same memory size) for additional performance. Showing its Starfighter qualities, the F104-based GTX 460 0.8GM board scored on average 21% higher FPS than the Radeon 5830, and the GTX 460 1GB was 32.6% and that’s damn impressive.

Discuss this entry