facebook rss twitter

Review: MSI GeForce FX5900

by Ryszard Sommefeldt on 29 June 2003, 00:00 4.5

Tags: MSI

Quick Link: HEXUS.net/qash

Add to My Vault: x


It's been graphics card city at Hexus recently, with samples from a good range of manufacturers that cover all ends of the consumer graphics spectrum. We've had the high end, the midrange and we've even got a card or two from the budget sector lined up for you as well.

What's nice about that is that it helps refocus the way you think about consumer graphics, for me at least. Being able to sample a piece from every part of the pie, genuinely gives me a better overall view of things. You get a better feel for where manufacturers are shooting at with new models, and their pricing strategies start to make a lot more sense.

But without giving the game away before I've even started, we should have a look at the board that's the focus of this review, MSI's brand new GeForce FX 5900.

Let's take a look at some basic GPU 101 first, to get you familiar with what makes up an NV35 solution, and we'll compare it to its immediate competitors along the way.

GeForce FX 5900Radeon 9800 ProRadeon 9700 Pro
Render pipelines 8** 8 8
Texture units per pipeline 1** 1 1
Shader units per pipeline 1** 1 1
Core clock 400MHz 380MHz 325MHz
Memory clock 850MHz DDR 680MHz DDR 630MHz DDR
Memory bus width 256-bit/32-byte 256-bit/32-byte 256-bit/32-byte
Memory bandwidth ~27.20GB/sec ~21.76GB/sec ~20.16GB/sec
Pixel fillrate 3200Mpixels/sec 3040Mpixels/sec 2600Mpixels/sec
Texel fillrate 3200Mtexels/sec 3040Mtexels/sec 2600Mtexels/sec

A quick glance at the table shows us that in terms of raw pixel pushing power and memory bandwidth, the FX 5900 has the definite edge over the other two GPU solutions from ATI.

But a quick glance doesn't show the full story. While NV35 (the FX 5900's internal GPU codename) is a definite 8x1 pipeline and TMU layout, 8 pipes and a single texture unit per pipeline, it doesn't operate that way under all conditions.

The GPU design means that NV35 drops to only processing 4 pixels per clock in a number of situations, effectively halving its pixel pushing performance (it keeps all its texturing performance by using 2 TMU's per pipeline). That occurs when doing regular colour and Z rendering, a rendering approach undertaken by the majority of titles on the market today. Future titles that render differently will allow NV35 to stretch its legs, using all 8 pixel pipelines to let you get the full benefit of all that pixel pushing and shader horsepower. A fearsome beast, but somewhat handicapped in some games.

As Tarinder mentioned in his close look at the FX 5900 Ultra back in May, NVIDIA effectively get around this imposed handicap in their design via a number of methods. The most obvious is the clock speed and memory bandwidth advantage it has over its competitors. Like the Ultra, regular FX 5900 has the same memory clock as its bigger brother, 850MHz, for a whopping 27.2GB/sec of memory bandwidth. That's a full 25% more than Radeon 9800 Pro. Clocking the GPU itself at 400MHz hasn't hurt either, so even in 4 pixel per clock mode, it still has a fearsome amount of ability.

Combine the raw speed with bandwidth efficiency measures in the form of Intellisample HCT, which gives colour and Z buffer compression at better ratios than was possible with NV30, and NV35 can make as much of that theoretical peak bandwidth figure as possible.

In terms of anti-aliasing and texture filtering performance, which is what consumer graphics is all about these days, it purports to offer an uprated version of Intellisample, their AA and AF engine. The new version of Intellisample applies AF to all textured surfaces in quality mode, something that ATI's recent GPU's don't. If implemented properly, this should offer better image quality than an ATI solution which doesn't texture filter all textured surfaces in a scene, instead dynamically choosing AF sample methods for each surface to give good IQ and good speed at the same time. NV35's Intellisample tech lets them emulate the ATI method of AF for a balance between IQ and speed too, so you theoretically get the best of both worlds. An objective IQ comparison is needed to confirm this however, something we'll endeavour to work up at Hexus soon.

So at the risk of not further regurgitating Tarinder's recent work with FX 5900 Ultra, I'll just point you at the front page of his article if you need any more information on how NV35 goes about its business.

I'd have liked to cover shader performance a bit more, but I'm rapidly running out of room, again the original Ultra review will help you get a handle on things in that respect.

So, we've talked about the GPU, its features, how it gets performance and how it stacks up against ATI's big guns. We better take a look at how MSI dress it up. And dress it up they do.