How Nvidia's Quadro FX5600 and FX4600 run on Harpertown

Product Leadtek Quadro FX5600 (avg US$ 3,500 - price varies)
Leadtek Quadro FX4600 (avg US$ 1,800 - price varies)

Web www.leadtek.com

LEADTEK, TOGETHER with PNY, is the only 3-D card brand endorsed by Nvidia to sell the highest-end and by far highest margin grade of them all: the Quadro line of professional OpenGL optimised cards.

It may be a reward for Leadtek's unyielding loyalty to Nvidia over the past many years, or their overall 3-D GPU expertise, but in any case, it is still the Satan Clara "green goblin" making far more dosh per each Quadro card than Leadtek itself, so if you need to haggle over the prices, knock (bang) on the correct door.

The two top end Leadtek Quadro cards right now are short-size FX4600, in a format and GPU/memory speed similar to GeForce 8800GTS but with full 384-bit bus and 768 MB GDDR3 RAM, and FX5600, with 20% faster GPU and 25% faster memory speeds, closer to the 8800GTX, but on a full-length card with a whopping 1.5 GB GDDR3 RAM.

The performance claims for the cards are 300 million triangles per second and 19M texels/s for the FX5600, as well as 250 million triangles/s and 12M texels/sfor the FX4600.

Besides the Quadro-specific drivers with all the Pro 3-D application presets, the cards come with Autodesk-specific utilities MaXtreme for 3DStudio Max and PowerDraft for AutoCAD - keep in mind that, despite this being just a slightly modified GeForce, lots of those expensive apps (often above $2,000 per CPU core license) require full OpenGL and application certification to get the (rip-off priced) support from their vendors.

That's why Nvidia can still afford to price the Quadros outrageously, beyond just the driver cost - ATI's new FireGL cards may be faster from raw hardware point of view, but they still lag somewhat on that critical driver and apps sup port front.

Now, expensive fast cards should go into even more expensive, faster systems - right now, the fastest workstation in this corner of Planet Earth seems to be that dual CPU Harpertown system I started running a day before IDF Fall. Eight 45 nm 3 GHz cores coupled with 16 GB FBD-800 DIMMs over dual FSB1600 should provide enough juice for literally anything.

One slight problem is that the floor-bound 60-pounder workstation couldn't easily be moved, while the full-length FX5600 had to be inserted carefully (see photo against the much smaller FX4600), so I bent a bit too much, and something in my back - old bones I guess - gave way. I could barely walk with all the pain, and three days of Chinese acupuncture had to come in...

So, I ran the cards under identical conditions with identical drivers - ForceWare 162 Quadro series on both WinXP 32 (for 3DMark06) and WinXP 64 (SpecOPC ViewPerf, and CineBench 10 OpenGL). 3DMark 06 was run as a reference to check if there is a big penalty in DirectX optimisation compared to similar gaming card speeds of 8800GTS and GTX, while the OpenGL 64-bit stuff was more of the real thing relevant to the users who may actually buy these cards for productive use.

Here are the results in the screenshots and tables:

BENCHMARK FX4600 FX5600
3DMark06 10657 13048
CPU 6890 6912
Shader2 4068 5107
Shader3 3920 4898
OGL 6027 6537

As you can see, the ViewPerf 10 test doesn't show much difference between the two cards - even though, in this case, the two fast Harpertowns leave only GPUs as the possible limitation on the benchmark results. It is interesting that 3DMark06 CPU does get a consistent couple percent extra jump on the faster OpenGL card, despite being a CPU-only test. The 6912 score is one of the fastest ever on a 3DMark06 CPU, and the fastest ever at my hands.

However, in large data set applications, whether in terms of model complexit y or textures, or some GPGPU work, the doubled memory of FX5600, coupled with the extra raw speed, could well pay for double the price. Even if using SLI with such cards, it makes far more sense than on the GeForce gaming front.

Why? Well, in games, what you may get is like 150 FPS vs 100 FPS speed record, or extra 'fancy effects' on to boost your ego at a LAN party. But, in workstation or HPC use, polygon-rich solid models will get more speed boost from triangle processing SLI parallelism, and multithreaded GPGPU jobs may get near twice as fast due to dual GPUs and double the memory. All these mean saving time, a "monetisable" category in competitive engineering and science arena - if you're rushing a visualisation of a new aircraft engine for a big client, or a next-generation game virtual environment, you will appreciate the extra speed as it may literally thicken your wallet...

In summary, both card handle the OpenGL benchmarks as well as 3DMark reference well. The thing is - the price has to come down, and that is Nvidia, not Leadtek, decision. The "DAAMIT wounded tiger" of AMD/ATI parentage has finally released the FireGL refresh, and those look promising speed-wise - at almost half the price in respective categories. The financially strapped AMD management has no choice but to support more aggressive market share fight in higher-profit segments like the professional OpenGL.

It will be even more interesting to see how this market competition goes when, in about six months, Nvidia (hopefully) releases Quadro versions of its upcoming G92 family chips - with, supposedly, full native double-precision floating point capability for even more GPGPU task acceleration. In the meantime, Leadtek Quadro FX5600 still carries the performance torch on the Nvidia OpenGL side, with the most complete pro application support driver among PC OpenGL cards. µ

Good Large memory, high OpenGL performance
Bad Large, long card taking 2 slot widths all the way
Ugly The price, obviously - ask JenHsun Huang about it, though

Barbird's verdict