Linux Quake3 rocks Win-XP Quake3 on new P4

The new Northwood 2.2gHz 0.13-micron P4, as I mentioned earlier, seems made for Windows-XP. It's got special chipset drivers; it's got an 'application accelerator'; it's got Rambus working overtime.

For Linux it's got nothing special to offer -- no accelerators or drivers. Just 2.2 in clock speed and a memory controller that exploits RDRAM nicely, which is definitely nothing to sneeze at. But it's got that on Windows as well.

So imagine my surprise when I benchmarked it with the only test I know that crosses the great divide between Linux and Windows -- the Quake-3 FPS benchmark -- and found that the performance of this Windows-loving kit was considerably better on Linux, at least in that context.

A brief re-cap of the hardware:

One Intel D850MVSE mobo with Northwood P4; 512M PC800 RDRAM; two Maxtor D740X 20G ATA-133 drives on the mobo's onboard ATA-100 controller, one booting Win-XP Pro on FAT and one booting SuSE 7.3 Pro on ReiserFS and both installed clean and subsequently patched; and a 64M DDR GeForce AGP4.

The Windows drive is patched with whatever the MS auto-update cloak-and-dagger process does to it. The Linux drive is patched to kernel 2.4.17. The video and OpenGL drivers for both OS's were upgraded with the most recent files from Nvidia's Web site.

On the Windows drive I installed all the Intel chipset drivers and the Application Accelerator. The Linux kernel is reasonably optimized for the HDD and the P4, but with APIC disabled, as it just won't run on the 850 mobo otherwise.

But that's hardly a problem.

Both operating systems, obviously, had to be running at the same level of display detail, and the limitations of XFree86 pretty well determined that for me. Both desktops were set at 16-bit color depth, and in both cases Quake was set with the following display options for the first series of runs:

Mode: 1024x786
Color depth: 16-bit
Lighting: lightmap
Geometric detail: high
Texture Detail: maximum
Texture quality: 16-bit
Filter: trilinear

It seems a bit skimpy, but rich detail takes more from the graphics accelerator whereas less detail gives us a better look at the CPU, chipset and system memory.

Win-XP returned an average of 72.7 FPS, which is worse than I'd expect from a P3 800 on '98 with about 128M RAM, or a 486DX 100 on Win 3.1 with about 16M RAM. (You see the pattern here....)

Linux returned an average of 80.2 FPS, which is significantly better, though hardly brilliant. But let's keep in mind that the system I'm using here is virtually Linux-hostile. The next one won't be.

With even less detail, further reducing dependency on the graphics card, we got better numbers from the CPU. The breakdown was similar, though Windows narrowed the gap a bit.

Mode: 640x480
Color depth: 16-bit
Lighting: vertex
Geometric detail: low
Texture Detail: minimum
Texture quality: 16-bit
Filter: bilinear

This gave us averages of 272.2 FPS on Windows and 304.7 on Linux.

We can infer that Win-XP is so greedy for system resources that even the most potent (and most expensive) CPU on the market, coupled with a hefty chunk of very fast RDRAM (also very expensive), only suffices to make it work nicely.

Other considerations

It's quite difficult to compare the performance of a given system on both Linux and Windows. The Quake benchmark is a rare exception, but basically it's apples and oranges. For example, what can we learn from evaluating the performance of Photoshop on Windows and the Gimp on Linux? Damn little, I reckon.

For that matter, what can we learn from running Netscape, StarOffice and the Gimp on both? It's entirely possible that these applications could have performance issues on a given OS which would skew the results.

I'll certainly try a number of tests like that during the weekend. This way, I hope, a single performance oddity won't cause too much distortion.

It's also worthwhile trying to match a system with an OS. For insight we can look at some of the everyday tasks common to both OS's, and compare them on different systems. I've taken a few common-sense measurements on both SuSE and Win-XP with the Intel 850/Northwood combo, but these won't have meaning until I repeat them on a different system and see where they differ.

Which I'll do, early next week. ®