AGP not dead; runs high detail Crysis at 25 FPS

We ditched our AGP-based test rigs several years ago, so we started our testing by raising some dead hardware from the labs graveyard and bolting it all together into a series of rebuilt Atomic Zombie test rigs.

Our first Zombie test rig comprised an Athlon64 3700+ (single core, circa 2005), 512MB of DDR400 and a PATA hard disk. Performance in this rig was actually pretty good; however the limited memory caused fantastic amounts of disk lag and grey hair.

3DMark confirmed this, breezing through graphic tests but coming to a sad halt in the CPU tests. The score reflected this, with our initial run with the 3700+ scraping in 5692 3DMarks. Bumping the memory up to 2GB with the single core was only rewarded with an extra 262 3DMarks – a little disappointing. It was apparent that for this card to fly, a little more CPU grunt was required.

Enter our second Zombie test rig, comprised of an Athlon64 X2 4200+ with 2GB of DDR400. Our world of pain was suddenly lifted, the milk and honey flowed again and we would happily use this rig for regular gaming. The X2 4200+ showed a marked improvement and collected 7916 3DMarks.

The AGP HD 3850 continued to eat everything we threw at it, but while we were testing on the X2 4200+, we swore we could feel a CPU bottleneck in-game. Crysis was still more than playable, averaging roughly 30fps with medium detail in 1280x1024, 20fps with high at 1680x1050. Bioshock, however was quite comfy, averaging at around 40fps when maxed out, also at 1680x1050.

Let’s just take a paragraph to repeat that: using an AGP card, Crysis is playable as long as you have a reasonably grunty CPU. And whats more, it wasn’t until we tried our third Zombie Test rig -- an FX-60 with 2GB of RAM -- that our old-school wonder showed its true colours. More on that on the next page.

You may have noticed a lack of Intel-based testing so far. This isn’t because we don’t like Intel (thanks to their 45nm quads, that couldn’t be further from the truth); we just didn’t have any Intel-based AGP gear to test with.

So how does AGP compare to PCI-e?