Nvidia fights back against AMD with fresh GPUs

AMD'S GPU DIVISION, that is, ATI, led the performance race against Nvidia for most of 2009 and at least the first half of 2010. If you count the dual GPU cards as one unit, it still leads now with the year-old HD5970 and is expected to continue to do so with the upcoming HD6990 next month.

However, on the per-GPU die scale, Nvidia finally got rid of most of the yield and performance issues it had with Fermi, and took over the DirectX 11 graphics performance crown at the very high end with its GF110 die, or Geforce GTX580, and the generic high end with its GF114 die, or GTX560 Ti cards.

Of course, the lead comes at the cost of somewhat large chips, where even the smaller GF114 die is within five per cent of the size of the Cayman die in the HD6970, but nevertheless it does move AMD to provide an HD6950 1GB card with similar performance and price. On the other hand, there are also 2GB versions of the GTX560 Ti availale from vendors like Gainward, as well as, at the high end, a $520 souped-up 3GB version of the GTX580, again from Gainward, to counter the overclocked versions of the HD6970 2GB Radeon.

So, let's look at it from several points of view. First, the performance. Absolute per-GPU chip performance and absolute per-card performance in a variety of key benchmarks, biased or not, would be a good starting metric. For the first one, as of now, the GTX580 holds that place in the majority of DX11 graphics benchmarks, but not the double precision (DP) floating-point (FP) compute benchmarks. Why? Well, Nvidia didn't enable full DP FP performance on the Geforce parts, in order to protect its highly priced professional Tesla compute cards, while AMD left the full speed DP FP enabled in the HD6900 series.

So, it's 1.5 TFLOPS single precision FP and about 150 GFLOPs DP FP on the Nvidia GPUs versus 3 TFLOPs single precision FP and 750 GFLOPs DP FP on the AMD GPUs. That's one angle that you won't find in 3Dmark or Heaven benches, but will be important for new compute application uses of GPUs.

From the per-card performance point of view, the AMD HD6990, with roughly 70 per cent more performance than the HD6970 using the slightly clocked-down dual GPU configuration, should keep the title for AMD overall. Vendors like Asus and Sapphire will likely release full-speed overclocked versions of that card fairly quickly, as Nvidia and AMD now both seemingly allow OEMs to launch customised versions almost on the heels of their reference cards' announcement dates.

Using a two-GPU card to beat the one-GPU one might not sound like an apples-to-apples comparison, however AMD is able to achieve that. Even though we saw some prototype PCBs for possible dual GF104/114 cards, and Asus also attempted to create a truly monstruous dual GF100/110 card in its MARS series, with three 8-pin connectors, the existence of a dual Fermi GPU in the open market is still just speculation. This includes the most updated rumours of a 'very secret' dual GF110 GF590 3GB card sometime this month, on time to take on the AMD HD6990 Radeon.

Then, we look at the total power, size and cost versus performance. The Nvidia GF110 and GF114 GPUs have somewhat lowered the numbers there, with the GF114 on the GTX560 Ti going down to 160W, giving a target for the AMD HD6950 to meet. Also, while the GF110-based GTX580 Ti card is still a big and clumsy package that can inflict real injury if wielded as a weapon, the smaller GTX560 series are as compact as any midrange cards can be.

Either way, AMD should take note of the resurgent Nvidia. The AMD 40nm process experience - and it's fair to guess also its relationship - with TSMC is definitely longer and more mature than Nvidia's, and the HD6900 series Radeons provide nicely balanced high end performance in both 3D and compute usage. But Nvidia's still very much alive Geforce branding strength and the success of its Tegra smartphone processors will rub off each other to help the often-troubled company get back on its feet.

AMD would do well to look into releasing some fine-tuned speedup flavours of the HD6900 series cards first, until the next semiconductor process shrink is ready. Even though top end cards are but a small part of total sales, the performance crown holder always enjoys the 'waterfall effect' with the mainstream part sales benefitting from the top speed position. So 2011 should definitely be more exciting than the last one, as both main graphics processor vendors have to speed up their new product introductions.

At the same time, AMD can use its substantial compute FP performance advantage, as well as large on board memory - useful for offloading larger tasks locally from the CPU without non-stop hopping over the slow PCIe connection - to show the benefits of its designs in more generalised graphics applications besides gaming. After all, unless 10 megapixel 4096x2560 or such 16:10 screens become commonplace, with the corresponding need for a high-end GPU to fill them, the 3D gaming benchmark performance difference between Nvidia's GTX580 and AMD's HD6970 - or for that matter their upcoming dual chip GTX590 and HD6990 incarnations - will be mostly academic. µ