CPUS GPUS MOBOS & CHIPSETS CASES & POWER STORAGE MOBILE FORUMS

3dfx vs. NVIDIA: Doing the math
by Scott Wasson — 1:41 AM on May 3, 2000

Earlier this week, Andy and I were kicking around all the info from the deluge of graphics-related announcements and reviews in the past couple of weeks. The question that stuck in Andy's mind was how the 3dfx stuff would measure up to the latest from NVIDIA. I told him I thought the Voodoo 5 5500, 3dfx's $300 mid-level card, probably wouldn't beat a GeForce DDR.

But then we started looking around at benchmarks, and some interesting things became clear. The new Voodoos have killer fill rates, so they do very well in higher resolutions and in 32-bit color. However, at lower resolutions, they don't do so well, even in a machine with a fast CPU. The consensus is that 3dfx's early drivers for the new Voodoos aren't yet up to speed, and once they get a little better, the cards should be much more competitive—even without hardware transform and lighting like the NVIDIA cards.

But how competitive? The question, really, is whether or not NVIDIA has left any room in the market for 3dfx's bread and butter, the V5 5500. And is a dual-chip setup a total kludge, like cr0bar's rant says? Or is it a sophisticated multi-processing architecture with its own sort of elegance?

I started looking at the numbers, trying to get a reasonable idea what we can expect from the hardware, assuming the drivers are there to push it. We'll leave triangle throughput aside for the time being, because 1) everyone knows NVIDIA's got the edge there, and 2) with a fast CPU and most current games, triangle throughput isn't an issue. (If you want to run professional 3D modeling apps, don't get a 3dfx.)

So we are left with two primary bottlenecks: a card's peak fill rate and its memory throughput. We can bust fill rate into two pieces: pixels/second and texels/second. "Texels" are "texture pixels," and in games that render using lots of passes, like Quake III, texel throughput really matters. In some other scenarios, with simpler renderers, pixel fill rates are more important, as I understand it. In this context, 3dfx's multi-chip architecture does have some merits.

Let's do the math and see how the contenders stack up. (Here's hoping I got my math right....)

	Core clock	Pixel pipelines	Fill rate Mpixels/sec	Textures per pixel	Fill rate Mtexels/sec
GeForce DDR	120MHz	4	480	1	480
GeForce 2 GTS	200MHz	4	800	2	1600
Voodoo 5 5500	166MHz	2 * 2	667	1	667
Voodoo 5 6000	166MHz	2 * 4	1333	1	1333

Clearly, the Voodoo 5 5500 has the potential to match or beat a GeForce DDR, but the 5500 is simply outclassed by the monster 1.6 Gigatexels/sec that the GeForce 2 can pump out. Once the $600 Voodoo 5 6000 hits the streets (with a very loud thud), 3dfx will have a strong contender, but it will probably cost about twice what a GeForce 2 GTS-based card will cost—and the GeForce 2 GTS will still have a theoretical edge in texel throughput.

Not that it's likely to matter, since the cards' peak theoretical fill rates will probably be choked off by a memory access bottleneck. Memory bandwidth is an intriguing question, because 3dfx and NVIDIA have taken very different approaches here. NVIDIA uses DDR RAM that sends twice the data in a "double-pumped" setup, while 3dfx uses plain ol' SDRAM, but lots of it. 3dfx gives each VSA-100 chip in its "array o' chips" its own 32MB bank of SDRAM, so a dual-chip setup gets roughly the same amount of bandwidth, total, that a single-chip DDR solution would. This approach isn't terribly efficient, but it's very effective.

We can do the math and compare the memory bandwidth available on each card, although, again, these numbers are nice for comparison, but differences in the implementations will make for differences in real-world performance. Memory bandwidth will be eaten up by geometry, texture management, and the like, and those numbers will vary from card to card. Again, I just hope I got the numbers right....

	Memory clock	Bus width	Memory	Memory bandwidth
GeForce DDR	300MHz (150MHz DDR)	128 bits	32MB	4.8GB/sec
GeForce 2 GTS	333MHz (166Mhz DDR)	128 bits	32MB	5.2GB/sec
Voodoo 5 5500	166MHz	128 bits * 2	32MB * 2	5.2GB/sec
Voodoo 5 6000	166MHz	128 bits * 4	32MB * 4	10.4GB/sec

If you fixate on those GB/sec numbers long enough, you can almost see why Intel engineers got all weak-kneed when Rambus came a knockin'.

The GeForce 2 GTS, despite its huge pixel and texel fill rate advantages over the GeForce DDR, has gained very little in memory bandwidth. The V5 5500 looks pretty good on this front, matching the GeForce 2 in theory and threatening to beat it in some real-world scenarios. (It's tough to say how these things will play out in terms of real-world performance, of course, but the possibility is there.) And 3dfx's monster, the V5 6000, might just rule the 3D roost (while doubling as a space heater). (Check the V5 6000 pictures here and let 3dfx strike fear in your heart.)

Whatever you think of their approach, 3dfx has undeniably given themselves a means of coping with the memory bottleneck problem and with NVIDIA's relentless onslaught of ever-faster one-chip solutions. None of this changes the fact that NVIDIA has cards on the shelves now, while 3dfx is still talkin'. Or the fact that the new Voodoos are still Voodoos at heart, with some very nice—and sometimes innovative—additions, but no T&L, and not enough integration to win the hearts of PC OEMs like Dell and Gateway.

But there is something to be said for the brute-force approach. It may keep 3dfx in the game a while longer.

Tags: GPUs