All Mt. Tiburon Testing Labs reports

Mt. Tiburon Testing Labs

Graphics boards and chips

Testing the AMD HD2900XT

AMD and Nvidia try to excite—no sale

In the last week of April AMD held a little party first in Tunisia and later in the week in Sunnyvale, because those AMD folks really like to party and stack up the mileage points. To attend either of these shindigs you had to sign a NDA promising you wouldn’t say or write anything about what you might hear at the event until May 14, 2007. On Tuesday April 24 we were given details about the R600 chip and the AIB it would grace, the Radeon HD2900. We were not given any boards (which is normally what is done) and were told they would be shipped to us in a few days.

Table 1. Benchmark results for AMD and Nvidia AIBs. (Source: JPR)

The next day, in the early morning, we checked the web and found test results for the HD 2900—so much for NDAs and the loyalty of “partners.” Furthermore, the results that were posted were not too good, indicating that they were probably done with green drivers and possibly rev 1 silicon. We experienced folks saw that right away, but the ignorant investor community and the newbie web boys went wild with gleeful sarcasms and told-ya-so rants about AMD going into the toilet and Nvidia eating their lunch. It was a most disgusting waste of bandwidth and little more than graffiti.

In the meantime Nvidia had their own announcement coming up and decided to take an unusual position,:they would make the release announcement and then give out boards to reviewers.

On May 2 2007, the GeForce 8800 Ultra was announced. A supercharged version of the GF8800, it was supposed to be a benchmark score killer—the world held its breath waiting for the first web coverage; we knew the pizza-eating, Jolt-drinking webbies would be hard at work testing the Ultra, an updated version of the GTX with cranked-up GPU and memory clocks. The Ultra features a 612-MHz GPU clock, 128 stream processors at 1500 MHz, and 768 MBytes of GDDR3 memory at 2160 MHz and a new super-cooler, all of which made the recommended retail price $830.

Then on the May 14 the world was given the official news about the AMD Radeon HD 2900 XT. Most of the world yawned. It was the Nvidia killer everyone thought it would be and the early leaks weren’t that far off. Yes, there were new drivers, and yes, the AIB did get very good scores, but not Nvidia killer scores—but it was the best price: $399 SRP.

Like everyone else in the industry we ran some tests on these parts, and you’ll find more information in this issue on the testing aspects. For now, take a look at the table on the preceding page.

The tests were run on two systems (one Crossfire, the other SLI), both with FX60 AMD processors, and we used the latest drivers. We did not run the GPU vendor’s favorite games.

This is bewildering

We knew the G80 had some headroom in it, and the GTX was already being overclocked by happy enthusiasts (thank goodness for thermal diodes). After the Internet lit up with news about AMD’s R600 arrival on May 14, we expected Nvidia would launch a preemptive product or special version, and they did, the GeForce 8800 Ultra. They also learned from the AMD Tunisia leak to hold off letting the souped-up AIBs out of their control until they were ready to have the data and test results released.

But we also knew the specification differences and thought the souped-up Ultra would be able to match the HD2900.

Quite frankly, the whole bunch of them are disappointing. The HD2900 with its higher clock, wider memory, over 2x the processors, and innovative new ring system should have blasted the benchmarks. And Nvidia’s sky-high price for an AIB that’s only a few percentage points better than the regular unit is absurd.

If you already own a GeForce 800 GTX there’s no reason to buy either of these new boards. If you don’t own a GTX, then the HD2900 is the best performance per dollar unit.

And the SLI Ultra is even more of a waste—a 8.7% increase in benchmark performance for $830.

Maybe next rev

Getting the HD2900 out was hard for AMD—the delays speak to that, and the fact that it came out based on an 80-nm process shows when it was started. The R600 design really does look good and should scale well, so predicting a 65-nm rev is no leap of imagination, especially when the R600’s little brothers are built in 65-nm. And we fully expect Nvidia to come later this year with a 65-nm version of the G80.

And what’s all this crap about DirectX10 anyway?

I’m sorry, I just can’t seem to see the emperor’s new clothes, what are they called again? Oh, right, DirectX10. But not to worry because I will see them soon, is that right?

For the folks who live in Europe or have friends there (or are friends of AMD), there is a game they can run from Ubisoft, “The Call of Juarez.” It was launched in September 2006 and now has, we’ve been told, a patch for Dx10. Yippee!

Nvidia is expected to host new DirectX 10 content on nZone.com in the form of a “Lost Planet” benchmark. AMD points out that “Lost Planet” is an Nvidia-sponsored title, and one that Nvidia has had a chance to look at and use to optimize their drivers. AMD complains that the developer has not made the company aware of this new benchmark, and as such the ATI Radeon driver team has not had the opportunity to explore how the benchmark uses its hardware and optimize in a similar fashion.

So all the folks who have Vista Ultimate, and an 8800 or a 2900, might be able to run some tests. And then, you know what? Yeah, the web wieniees will beat themselves silly arguing which board screws up the most in Dx10—and that’s supposed to influence sales of AIBs?

What do we think?

Yeah, I know, I know, I just don’t get it. I’m out of touch with reality, why hell, I’m not even using a supercharged Core2Duo for our benchmarking—what a Neanderthal.

Well, I can’t hang here with you and take this abuse; I’ve got to go find the emperor’s clothes.—JP

Testing the AMD HD2900XT Ted takes the R600 on a little joy ride By Ted Pollak, JPR Games Analyst
Figure 1. The ATI 2900 XT in Crossfire mode is a race bred champion. (Source: rFactor)	Figure 2. A lotta power and a lotta heat. My custom temperature fighter—I call it the Swamp Cooler. (Source: JPR)

When Jon called to ask if I was interested in helping test the ATI R600, the first thing that came to mind was a completely stutterfree frame rate while maxing out everything, using my favorite AIB torture tester—”rFactor.”

I was not disappointed (that is until Jon showed up to take the boards back for Vista testing).

Running two of these cards in Crossfire mode was a sight to behold (and experience) and provided the most immersive virtual racing experience yet.

In this scene my desk is slightly shaking from subwoofer reproduced engine and exhaust noise and my arms are actually in pain from fighting the force feedback wheel. In the time it took you to read the last sentence the R600s have rendered over 700 frames.

The test platform I used is the Asus A8R32-MVP Deluxe motherboard, an AMD Athlon 64 FX-60 dual-core processor running around 2.6 GHz, 2 Gigs of PC 3200 RAM, a Tagan TG1100-U6 (1100W) power supply, and of course the two 2900 XTs. The OS for my part of the test was Windows XP SP2.

But first, benchmarking

Before getting to the game testing I ran these babies through 3DMark06 for which the results are summarized in Jon’s table (p. 13), benchmark results for AMD and Nvidia AIBs. They put up fairly impressive numbers considering the price to performance ratio, though the power supply will be part of the equation for many purchasers.

Figure 3. The Devil is not in the Details. New features in Catalyst Control Center and the accompanying drivers bring out amazing details that are difficult to render at speed. (Source: rFactor

In rFactor these cards all but eliminated the slightest hint of jerkiness. From within the cockpit of the cars, frame rates on many tracks approached 120; running at 1680 x 1050 with maximum in-game graphics settings (as a matter of practice I avoid commanding the cards to override in-game graphics settings). We are talking silky smooth. On replays in out-of–car scenes, which are the most demanding for graphics and processing hardware, I think I detected the “jerks” maybe once or twice. Most people would probably not even notice, but I’m convinced that I’m one of those eagle-eyed superhumans that can sense details and movement better than others (wait, where did my cursor go? … oh there it is).

Raw processing power is useless unless you have the software to complete the package. The new CCC and drivers give the user an amazing amount of control ranging from simplified “speed vs. detail” settings all the way up to some fairly advanced overclocking functions. As I previously mentioned I let the game software manage advanced graphics settings but there was one feature that I chose to engage called “adaptive anti-aliasing.” It may be difficult to see the details in the Porsche screenshot in Figure 3 (below), but the individual links of the cyclone fence on the right seemed to render more accurately with this feature turned on. (As this track comes from the mod community we will forgive the bush anomaly on the fencepost to the right—it’s not like they had an army of bug testers at their disposal.)

What do we think?

The R600 is a kick-ass AIB, and in Crossfire array it may compete with a single Nvidia 8800 Ultra for price/performance ratio. (Power supply expense can affect this ratio.) Of course this is just from the perspective of Windows XP DX9. Most are buying this generation of cards for Vista and DX10.—TP

• Back to Top

• More Mt. Tiburon Testing Labs reports