NVIDIA GTX580: a Fermi saver?

After coming close to screwing up Fermi the first time round, has NVIDIA finally released a card that can redeem it? Read our review!

NVIDIA came really close to screwing up Fermi at the last minute. In an effort to remain relevant in the battle for DX11 hardware it rushed out the GeForce GTX 480, delivering a product that felt slightly undercooked. The 480 had less CUDA cores than NVIDIA had initially announced, and ran noticeably hot and loud – all signs that the GPU had been rushed out a few revisions short
of perfection.

This was seemingly confirmed by the fact that the latter variants of the Fermi architecture ran incredibly well, with no sign of the heatwave that emanated from the GF100-based GTX 480 and 470 cards. These latter cards used new cores, ones that were less ambitious than the GF100 and its three billion transistors.

When we first saw the rumours leak out about the GeForce GTX 580, our assumption was that NVIDIA had kept back the best GF100 chips and were launching a tweaked up version of the GTX 480 to compete with AMD’s imminent 6900 series of GPUs. In reality, the GTX 580 uses a new variant of the GF100 silicon, codenamed GF110, and while the architecture is still Fermi, the GPU has had a lot of attention paid to the major weaknesses in the GTX 480. Both heat and power draw have allegedly been solved with the GF110 silicon, and new additions to the reference design of the cards further act to improve cooling and TDP.

Besides these improvements, NVIDIA has also managed to get all the CUDA cores running on the GTX 580. When we reviewed the GTX 480 we noted that the initial briefings called for 512 CUDA cores, but the actual card had one block of 32 cores disabled. In what is a rarity this meant that the GeForce GTX 480 actually referenced the presence of 480 CUDA cores inside the GPU. Given that graphics card model numbering seems more about the illusion of ZOMG massive numbers than anything approaching reality, it is no surprise that NVIDIA has gone for a number 100 more than 480 rather than reference the 512 cores.

These minor architectural tweaks mean that we don’t expect huge performance leaps with the GTX 580. This is a card with a dual purpose – keep competitive with AMD’s top end, while erasing the major criticisms of the GTX 480. 

One other significant side effect of the focus on improved power consumption and cooling is that NVIDIA has put in hardware and software optimisations for programs like Furmark, which are purely designed to load up a GPU to maximum. These optimisations throttle performance in order to stop the card overheating. When a program like Furmark or OCCT is detected, hardware VRMs kick in and reduce the card’s performance. Some websites have managed to bypass these optimisations and it seems that the card will spiral out to 350W if the hardware isn’t keeping it in check.

There is huge potential for debate over this move. Unlike optimisations made to improve benchmark scores, these ones seem deeply tied to the new power management paradigm. In the end it’s only working on software with very little real world relevance – Furmark et al. are designed as worst case scenarios, and have never had any pretence of representing real world graphics workloads.