Nvidia Fermi GF100 review
The new architecture isn’t designed solely for 3D gaming. Fermi also boosts the GPU’s ability to perform non-graphical calculations, such as those used in cryptography or data analysis, via the proprietary CUDA programming framework. Nvidia claims Fermi can perform some functions eight times faster than the GT200 core found in the previous top-of-the-range cards – and it’s also the first card from any company that supports C++ natively (with CUDA extensions to support the card’s highly parallel architecture).
Burying bad news?
Mike Jennings raises some questions about Nvidia’s low-profile launch of the Fermi GF100
This sophistication is doubtless a major reason why the card has taken so long to produce, but Nvidia hopes the investment will give Fermi a strong foothold in the workstation market, where its 512 potential cores can tear through data at hundreds of times the rate of a desktop CPU. And for the home market, Nvidia is encouraging developers to use CUDA for sophisticated in-game processes such as physics simulation and artificial intelligence.
We carried out a test using CyberLink PowerDirector 8 – the only piece of software we know of to support both CUDA and ATI Stream – and found little clear ground between the two. We rendered out a 1080p video clip to AVCHD format with a single accelerated filter effect applied to it, and found the GTX 470 and 480 finished the test in just under 3mins 30secs. That’s a boost of around 35% over non-accelerated performance. By contrast, the ATI HD 5870 was slower doing the standard encode, but took 3mins 47secs when accelerated, which was a boost of 39%.
So CUDA can significantly reduce the time taken for long encoding jobs, but at this stage doesn’t offer much over the competition, and you only see the benefit when making heavy use of a select group of supported filter effects. Indeed, it remains to be seen whether Fermi’s computing powers will be widely utilised by mainstream applications. Many developers may prefer to stick with the basic but universal computing functions built into DirectX, rather than investing in CUDA routines that run only on Nvidia hardware.
Will it sell?
We don’t doubt Fermi’s appeal in certain specialised areas, and Nvidia’s plans for Quadro and Tesla variants with up to 6GB of RAM are arguably more interesting. Whether the GTX 480 and GTX 470 will win back the legions of fanboys who long ago ran out of patience is a different question entirely. The GTX 480 is expensive at £429 inc VAT, while the more mainstream GTX 470 will still set you back £299 inc VAT. Does that 5% to 10% speed boost justify a £70 premium over a quieter, much more efficient HD 5850? We’d argue not.