ATI CrossFire and NVIDIA SLI in Terms of CPU Performance

We proceed with the series of articles, which can be entitled "Nitty-Gritty of iXBT.com Testlab", devoted to results of our inhouse tests - they can be interesting to some of our readers. This article will be devoted to consequences of one our experiment: what if equip CPU testbeds with the most powerful graphic solution - two video cards in CrossFire (ATI) or SLI (NVIDIA) mode? There were some disappointments, but the results were funny. OK, let's see...

Hardware and Software

Testbed configurations

CPU	Motherboard	Memory
Core 2 Duo X6800	ECS nFORCE 570 SLIT-A	Corsair CM2X1024-6400 (5-5-5-12)
Core 2 Duo X6800	Intel D975XBX (BIOS 1334)	Corsair CM2X1024-6400 (5-5-5-12)

A small note: as the CrossFire technology is officially supported only by motherboards with ATI and Intel 975X chipsets, and SLI is supported only by motherboards with NVIDIA chipsets, we couldn't assemble identical testbeds for ATI and NVIDIA video cards. But as it's impossible (using unofficial hacks in testbeds is not acceptable), our comparisons of results are quite well posed.

• Video cards: GeForce 7800GTX 256 MB (Gigabyte), ATI Radeon X1900XT (Sapphire)
• Memory: 2 GB (2 modules)
• HDD: Samsung SP1614C (SATA)
• Coolers: standard models that come shipped with processors
• PSU: CHIEFTEC GPS-550AB A

Processor	Core 2 eXtreme X6800
Process Technology	65 nm
Core Clock, GHz	2.93
Number of Cores	2
L2 Cache*, KB	4096
FSB clock**, MHz	266 QP
Multiplier	10
Socket	LGA775
Typical thermal emission***	>75 W
AMD64/EM64T	+
Hyper-Threading	-
Virtualization Technology	+

* — "2x..." means "per each core"
** — in AMD processors it's frequency of the memory controller bus
*** — it's measured differently in Intel and AMD processors, so a direct comparison is not correct

Software

1. Windows XP Professional x64 Edition SP1.
2. 3ds max 7.0
3. Maya 6.5
4. Pro/ENGINEER Wildfire 2.0
5. SolidWorks 2005
6. F.E.A.R. 1.3
7. Half-Life 2
8. Unreal Tournament 2004 build 3339
9. Quake 4 Point Release 1.1

Drivers

1. NVIDIA ForceWare 91.31
2. ATI CATALYST 6.6
3. NVIDIA nForce SMBus Driver 4.50
4. Intel INF Update 7.2.2.1006

Testing

In this case we decided to depart from the concept of "a total score in a subgroup of tests", which is used in our new test procedure by default, and to focus on the most illustrative results.

3ds max 7.0

Insipid results, both for CrossFire (CF) and for SLI. SLI is formally at advantage, but the gain is so meager in both cases that we shouldn't really take it into account.

Maya 6.5

Complete fiasco again. Neither CF nor SLI provide any performance gain in visualization in 3D modeling packages.

Pro/ENGINEER Wildfire 2.0

The situation in a professional engineering package is even worse: CF and SLI reduce (!) performance in wireframe mode. In Graphics Shade test, performance slumps only on the testbed with CF. But NVIDIA SLI brings a meager performance gain.

SolidWorks 2005

The smaller results, the better. Complete fiasco: CF and SLI result in a performance drop. To be precise: NVIDIA SLI drops performance only insignificantly, unlike the ATI technology.

F.E.A.R.

Everything falls into place in the first game: SLI and CF work as they should. Performance gains do not reach twofold, but they are still significant.

Half-Life 2

One of the most CPU-intensive games is still limited by a processor even in high quality mode at 1024x768, video cards have nothing to do with it. At least they don't make it worse, like in professional packages ;).

Quake 4

Quake 4 benefits from SLI as well as CF.

Unreal Tournament 2004

Old UT2004 also reacts well, but not as much as we'd like. The situation would have certainly become better at higher resolutions or with AA. But it's hardly relevant for us, as we analyze CPU performance, not GPU performance.

Conclusions

Conclusions are crystal clear for those who have scrutinized the diagrams. It's quite obvious that ATI CrossFire and NVIDIA SLI are purely gaming technologies so far. They are intended for dynamic 3D games. Both technologies do not perform well in professional applications. It would be logical to assume that the problem is not in these technologies, but in the lack of attention to this tandem from developers. An indirect proof of this fact is the lack of ATI FireGL for CF. What concerns NVIDIA Quadro with SLI, we can only point at the drivers: to all appearances, SLI optimization for professional applications is disabled, if a regular desktop GeForce is detected (we all hate the so called positioning so much, don't we? :)).

Thus, considering that the CPU test procedure used in iXBT.com testlab does not pay major attention to games, it's not expedient to use these technologies in CPU testbeds. Similar advice can be given to those users who thought about building CF or SLI tandems for powerful graphics workstations intended for professional 3D modeling applications and CAD.