All Mt. Tiburon Testing Labs reports

Mt. Tiburon Testing Labs

Graphics boards and killer computers

FIGURE 1: 3DMark06 tests of three AIBs on an AMD Spider platform.
(Courtesy CBS Studios)

FIGURE 2: Relative performance gain over a single HD 3870 on a Spider platform.
(Source: Jon Peddie Research)

TABLE 1: Analysis of the data and benchmarks per dollar.

It has taken us almost three weeks to do all the testing needed for this article. It was a labor of love.

We had two motherboards with new CPUs, and three AIBs, not totally equal in all specifications, but all with interesting features and prices.

The equipment:

• AMD Spider with Phenom quad-core processor.
  • Two Radeon HD 3870 AIBs.
  • One Radeon HD 3870 2x AIB.
• Intel Skulltrail with dual quad-core Xeon processors.
  • Nvidia 8800 GTX AIBs.

Due to the newness of these systems, we couldn't do mix and match comparison, for example we couldn't run Nvidia AIBs in SLI on the Spider, and we couldn't run the new 3870 AIBs on the Skulltrail. It will take a little time for all the drivers to catch up. However, what we have is still pretty interesting.

AMD Spider and 3870 AIBs

This system is an Asus motherboard with a 2.6Ghz Phenom quad processor. We had a lot of trouble bringing up this system, and had to learn a few things the hard way. For one, AMD does not have Vista 64-bit drivers. When we finally understood that, we yanked our HDD and installed a 32-bit version of Vista Ultimate that AMD provided to us so we'd stop calling them for new drivers; it worked, we shut up.

3Dmark06 tests

The criticism of this test is that it's 3Dmark06, a two-year old test for DirectX 9. We ran three sets of tests on this system, a single HD 3870, dual crossfire HD 3870s, and a single HD 3870x2. The results using 3dMark06 are in Figure 1. As the chart indicates the 3870x2 is the best of the lot in all cases in terms of raw performance.

We normalized the data to the single 3870 and made a comparison, the results of which are shown in Figure 2

This is an interesting figure because it clearly shows three things, first the gain in performance of two AIBs over one, the gain of a closely coupled dual GPUs over dual AIBs, and the relative gain in performance of multiple GPUs over one at higher resolutions. It also showed a strange dip in performance at 1600 x 1200. So far we haven't found an explanation for this.

We ran one more analysis of the data, benchmarks per dollar. Not everyone can afford the maximum performance and has to shop efficiently. We searched the web for the latest and lowest price on the AIBs and came up with the data show in Table 1.

We then normalized to the 3870x2 and produced the results you can see in Figure 3.

Here the interesting thing to notice is that as resolution increases the extra cost of the higher performance AIB(s) is diminished. This clearly shows the value of a low-cost good performance AIB to a casual gamer and the benefit per dollar of a high performance and more expensive AIB to game enthusiast.

Intel Skulltrail and Nvidia 8800GTX

FIGURE 1: Benchmarks per dollar normalized to the 3870x2.
(Source: Jon Peddie Research)

TABLE 1: comes with eleven fans and gigundous heat sinks.
(Source: Jon Peddie Research)

This system is now the most powerful system we have, or ever had. It displaces our beloved Darth Vader the AMD QuadFX, which will live near it, slightly behind. As Robert Dow, our long suffering co-tester said when we fired up the Skulltrail, “if the QuadFX is Darth Vader, this is the Death Star. “

The machine looks as powerful as it is.

There are twelve fans in this beast—four on the door, one on each processor, one on each AIB, one in power supply, a back panel, a front panel, and one in the top, and remarkably it is reasonably quiet. Not only that, Intel says they can make it even quieter, we're waiting for that patch.

The Skulltrail is a D5400XS-based 1600MHz bus chipset motherboard with 4 GB of DDR3, two 45nm, quad-core QX9775 Core 2 Extreme processors running at 3.2GHz, processors giving the system eight cores with a total of 25.6GB/s of bandwidth. This is one powerful system! We expect this will be the benchmark system for a long time. However, it's not cheap.

The QX9775 is priced around costs $1,500 per processor, the D5400XS motherboard costs around $600, plus 4GB of FB-DIMM DDR3 memory which is $800, plus a $1,000 1500w power supply, you've got a system that costs $3,900 without a case, HDD, OS, peripherals or AIBs.

3Dmark06 tests

Running 3DMark06 on this machine was pure pleasure. Everything was easy, snappy fast and produced amazing numbers. The system was equipped with two Nvidia 8800 GTXs running SLI. We've been told it is also capable of running ATI HD 3870s in Crossfire, but at the time of this writing we couldn't get a commitment from AMD that it would work, and we were very fearful about messing with the SLI setup since it's not exactly a sanctioned deal, and under the pressure of deadline, this was not the moment to engage in a science project.

We ran two sets of tests on the system, SLI on and SLI off. The comparison is shown in Figure 4

When normalized to the single 8800 GTX AIB, you cans see the gain from SLI operation. Notice that SLI also improves the relative performance at higher resolutions, as shown in Figure 5.

A not apples to apples comparison

Plotting the 3DMark06 scores of the Radeon HD 3870x2 AIB in the Spider platform to the 3DMark06 scores of the dual Nvidia GeForce 8800 GTXs in SLI mode in the Skulltrail platform we got the results shown in Figure 6.

On a percentage comparison, relative to the HD3870x2, we found the differences to be surprisingly tight, within 5%, as shown in Figure 7

Given the relative closeness of the AIBs when tested on the same platform, we believe this is result primarily shows the CPU's contribution.

DirectX 10 tests

With the 3DMark06 test results as a baseline we then wanted to look at the relative performance of these systems in DirectX 10 operation. We ran “Crysis,” “Call of Juarez,” and “Company of Heroes,”—or I should say, we tried to.

AMD Spider and 3870 AIBs

We ran Crysis on the AMD platform and got the following results

The AMD cards would not run (WNR) the benchmark at 4X 2560 x 1600; it would crash on the third or fourth pass.

The jump in frame rate at 1600 x 1200 is completely inexplicable and we checked it four times—it is consistent, and complimentary to what we saw in the 3DMark06 tests.

Intel Skulltrail and Nvidia 8800GTX

The Skulltrail system proved to be more temperamental than the AMD Spider. We ran “Company of Heroes” (COH) and got the following average FPS results

It was hard work to get the numbers because when the program ran there was no cursor.

“Call of Juarez” wouldn't run at all, and “Crysis” loaded and ran but with no cursor making it impossible to change settings.

What do we think?

All in all this has been a very tiresome and frustrating experience and really exposes some major weaknesses in these new systems The hardware is there, but the software leaves lot to be desired and no one is exempt. There were problems on Microsoft's side, from the drivers from AMD and Nvidia, and from the games themselves.—JP

• Back to Top

• More Mt. Tiburon Testing Labs reports