Our experience with the test system was marred somewhat by two unrelated video card issues.
Unstable Video Overclocking.
Initially, our two GTX 275 cards were each clocked at 700MHz, a speed the company claimed it had verified 100 percent. In real world 3D tests, however, that speed proved unsustainable; the system crashed almost immediately in every test. Oddly, each card individually was capable of running at 700MHz in a non-SLI configuration, but enabling the feature immediately destabilized the system until the GPU clockspeed was lowered. We contacted the company and backed the cards down to a core clock of 675MHz, at which point we were able to test without further incident.
When we inquired about the company's method of verifying video cards were stable at overclocked speeds, we were told that Digital Storm uses both Crysis and Furmark as stability tests, with Furmark considered the higher bar. After downloading and running Furmark (the program bills itself as an overclocking stability test and OpenGL benchmark) on the two GTX 275's in SLI at 700MHz, I can confirm that the cards are stable at that speed in the program's stress test mode. Try to run the actual benchmark, or any benchmark, and the cards promptly crash.
Digital Storm shipped its test suite (and several screenshots) installed to our review system, which allowed us to do a bit of cross comparison. Based on the Furmark test results reflected in one of the screen captures, it appears the stress test was run on just one card (or with SLI disabled).
We don't know if Digital Storm tested each card individually at 700MHz and then assumed they'd run correctly at that speed in SLI or if another factor confounded our results, but we've recommended the company use a more robust set of benchmarks for determining overclock stability. Furmark does an excellent job at heating up a GPU and draws a prodigious amount of power, but this test should be augmented by at least one of the mainstream 3DMark tests. It's not always possible to predict which programs or games are going to destabilize an overclocked component. The 3DMark programs are pretty good GPU stability tests—all of them since '03 have been sensitive to GPU instabilities—but even a 3DMark run does not, in and of itself, prove stability. The only way to do that is one program at a time; it's for this reason that any company (including OEMs) shipping video cards in an overclocked configuration should have a suite of tests.
Incorrect GPU Configuration:
The second problem is a bit more complex. When the Digital Storm i750-based system arrived, it was configured exactly as shown in our photos, with a GTX 275 in each of the first two PCIe physical x16 slots. At first, we didn't realize this was incorrect until after putting the system through a complete series of 3D benchmarks. It wasn't until we were actually writing this review that we stumbled across the following on page 22 of the FTW's manual. (PDF) "For SLI use, it is very important to use only PCI Express Slots 1 and 3 (primary and third) for optimal performance."
They aren't kidding. Here's why it matters:
The P55 FTW is built on Intel's P55 chipset, as shown above. 16 of the chipset's 24 PCIe lanes are Gen 2 and attach directly to the processor. The remaining eight are first-generation lanes and hang off the southbridge. The Gen 2 lanes directly off the CPU can be allocated to provide either a single video card with an electrical x16 connection or split between two cards at x8/x8. Since PCIe Gen 2 offers twice the bandwidth of PCIe Gen 1, two x8 connections offer each card 3.2GB/s of unidirectional bandwidth when in SLI mode (6.4GB/s bidirectional). In the case of EVGA's P55 FTW, it's the first and third slots that run at x8/x8. Digital Storm actually put its own review system at a severe disadvantage when it shipped the video cards in a Slot 1/Slot 2 configuration. Not only is the second slot limited to a fraction of the first's bandwidth (800MB/s, 1.6GB/s bidirectional), it's also attached through the southbridge rather than directly to the CPU. This adds additional latency to an already-lopsided linkage and hamstrings the configuration further.
EVGA didn't just put this bit of critical information in the manual—they actually silkscreened it on the board.
Pop out the first two video cards and there it is, plain as day.
It's hard to know what to make of this. Obviously it's not the sort of thing any company does on purpose, but it's not the sort of mistake that an enthusiast-oriented, custom-designed OEM wants to make either. If EVGA had simply noted the issue in a manual footnote that might be one thing, but the requisite information is plainly written. On other EVGA motherboards, the primary PCIe slots are a separate color from the secondary southbridge connection (which the company occasionally refers to as a PhysX slot.). If the difference between a Slot 1/Slot 2 configuration and a Slot 1 / Slot 3 configuration was minimal that would be one thing, but as you'll see, the performance gap is significant.
This is also where the ill-placed red-button from Page 3 causes trouble. As shipped, the button blocks the use of the third PCIe x16 slot. Normally, moving the backplate would be the work of seconds, but Digital Storm's tight cabling work prevents this. Moving the button to a different slots required rewiring and partial removal of the power supply, as well as adjustment at multiple points—the wiring to that button runs behind the entire case before connecting to one of the cold cathodes and the front fan.