The nVidia GTX 680

[Dirk Broer]

nVidia has a new card, the GTX 680, which they claim is world's fastest card at the moment. But seeing the relevant (for crunchers that is) benchmarks, I'm less than impressed. Why?

The good:
nVidia finally has a card with a decent SINGLE precision performance and as a bonus with a good GFLOP/Watt (in tdp) ratio

3090/195=15.85
The 3,090 is between the performance of the HD 7950 (2867) and the HD 7970 (3789), the 15.85 is an unprecedented good value for nVidia
(their previous best performer in the GLOPS per Watt was the GTX 560Ti with a score of 7.43, my GTX 260 does 3.54...)

Compared with the GTX 580 this card has three times the Cuda cores, twice the texture units, but slightly less ROPs. (32 compared to 48)
GPU Core runs at same speed as the shaders now (shaders used to do twice the core speed with nVidia's earlier designs) while memory runs at an insane 6000 Mhz.

The bad:
For some reason some marketing guru at nVidia decided to cut the DOUBLE precision performance to a shocking 1/24th of the single precision performance. Fermi had 1/8 (not all), older designs did with 1/12. As a comparision: Ati/AMD cards used to have 1/4 or 1/5 ratios.

As a result of this way of pushing their Quadro and Tesla (who have the ratio, using the same basic chip, at 1/2) the GTX 680 has a double precision performance of a meagre 129 GFLOP, about the level of a HD 4830 (which scores 147.2 GFLOP DP actually)
The GFLOP/Watt DP thus ends with a dissapointing 0.66

The HD 7750, a mid-range AMD card, does 819 GFLOP single precision, but due to a change in AMD strategy cutting DP performance to 1/16th of SP, manages 51.2 GFLOP double precision, doing so at a mere 55 Watt,
giving a score of 14.89 GFLOP/Watt SP and 0.93 GFLOP/Watt DP. I already declared this card a bit prematurely the new king of budget crunching, but that title remains with the HD 4770. The cutting of DP to 1/16th of SP is for both Cape Verde (77xx) and Pitcairn (78xx) GPUs, the Tahiti (79xx) gets 1/4 and remains a good cruncher both SP and DP

The HD 7970, the rival to the GTX 680, does 3789 GFLOP single precision and 947 GFLOP double precision, this at a tdp of 230 Watt.
While that wattage might frighten you -and your wallet-, it means 16.47 GFLOP/Watt SP and 4.12 GFLOP/Watt DP, surely more attractive figures than nVidia's flagship.