nVidia’s new GeForce GTX680: You win some, you lose some?

Today, nVidia has introduced the first of its new 600-series of videocards, based on the new Kepler architecture (as opposed to the Fermi architecture used in the 400- and 500-series).

All the major review sites are putting up their reviews of this new card. Let’s look at the Anandtech review, as usual: http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review

In short, it seems to be a very strong performer in games. It mostly outperforms AMD’s Radeon 7970, with just a few games where the 680 ties, or loses slightly. Looking specifically at tessellation, we see it is a bit hit-and-miss:

At normal tessellation settings, the 680 loses out to the 7970 (which is not too surprising, as AMD’s approach has very high peak-throughput at low tessellation factors, but drops off exponentially as the factors get higher, as discussed previously). It also loses out to the 580, which is… odd…

When we up the tessellation however, things change:

Apparently the 680 has not lost the ability to scale, and goes straight past the 580. The 7970 is still faster however, but that would probably change at higher tessellation settings, this sample does not go all the way to the maximum setting of 64x. Namely, when we look at Unigine Heaven, we see the 680 coming out on top:

When we look at the compute performance, we see a similar hit-and-miss characteristic… In some areas the 680 is very fast, in others it struggles to keep up with the 580, let alone the 7970.

A matter of efficiency

Is there an explanation to these performance characteristics? Well yes. Namely, the GTX680 is not quite an all-out high-end design like the GTX580 and 7970 are. For starters, it only has a 256-bit bus, where the GTX580 and 7970 both have a 384-bit bus. So it could be that the 680 is slightly more bandwidth-limited in certain situations. It tries to compensate for that with slightly higher clocked memory, but if we compare the numbers:

• Radeon 7970 has 384-bit memory at 5.5 GHz
• GeForce GTX680 has 256-bit memory at 6 GHz
• GeForce GTX580 has 384-bit memory at 4 GHz

This means that theoretically the 7970 has about 37.5% more bandwidth than the GTX680. The GTX580 and GTX680 are tied. Aside from that, the GTX680 has ‘only’ 2 GB of memory where the 7970 has 3 GB.

The GTX680 is also relatively small for a 28 nm chip, partly because of the 256-bit bus. It has a transistorcount of 3.5B. Which is only slightly more than the 3B of the GTX580 (which was manufactured at 40 nm), and considerably less than the 4.31B transistors in the 7970. So this is more of a lean-and-mean design, instead of the incredibly large GPUs that nVidia has been making the past years. nVidia also deliberately limits the 64-bit float performance, so that their consumer cards will not threaten their professional products (Quadro and Tesla series).

So that explains why the GTX680 can not outperform the 7970 and GTX580 in every situation. And it makes it all the more impressive that it outperforms the 7970 in most other situations. This architecture is smaller, and consumes considerably less power, yet it performs better. So it is quite an efficient architecture. Then again, in my coverage of the 7970 introduction I already pointed out that the 7970 did not perform quite as impressively as its specs on paper would indicate, by extrapolating the Fermi specs to 28 nm and faster GDDR5 memory.

There are rumours circulating however, that the GTX680 is not the largest Kepler-based GPU that nVidia has planned. Which could be true. If nVidia still wants to go with large high-end GPUs, like they did with Fermi, they have almost 1B transistors to spend before they even get to 7970-level. If they opt for a 384-bit bus with 5.5 or 6 GHz memory, and add a bunch of extra stream processors, such a GPU is likely to outperform the 7970 across the board.

I think then, that this GTX680 is a reincarnation of the GeForce 8800GT. The 8800GT was also a leaner-and-meaner version of the original 8800 architecture. nVidia opted for a 256-bit bus instead of 384-bit as well, yet because of architectural improvements, the 8800GT (and the later 8800GTS based on the same die) could keep up with the older, larger and more power-hungry 8800GTX/8800Ultra cards. nVidia later released the GTX 280 cards, which were the true ‘high-end’ variation of the 8800GT/GTS, taking the 65 nm process to large scale again.

So for now, the GTX680 is a great card, combining class-leading performance with reasonable price and power consumption levels (much like the 8800GT in its day), and if the rumours are true, nVidia has a truly killer Kepler GPU planned for the future.

Advertisement