Review: NVIDIA GeForce GTX 670 SLI vs. The Beast™

by Tarinder Sandhu on 18 May 2012, 14:49

Tags: NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qabgvv

Add to My Vault: x

Take two, baby

Recommending a high-end graphics card used to be a fairly straightforward business. You'd take a look at NVIDIA and AMD's best products, run a few benchmarks, and pass judgement. We did just this when the impressive GeForce GTX 680 arrived on the scene. It was, and remains, the best single-GPU option available. In a move that kept us guessing until the last moment, NVIDIA then launched the card we like the call the Beast™, the twin-GPU GeForce GTX 690, which became the best graphics card around. But just last week the GeForce GTX 670 rolled into town. It offers about 90 per cent of the GTX 680's performance at 80 per cent of its price, making it, even at £325, a better value proposition.

We already know how the £900 Beast™ of a GTX 690 stacks up against a couple of GTX 680s in SLI and two of AMD's finest single-GPU cards in CrossFire. If you can't be bothered to go back through previous reviews, and who'd blame you, GTX 690 performs just as well as two GTX 680s and, on average, better than a couple of HD 7970s. Got all that?

But we've left a few questions unanswered on the premium graphics-card front. The GTX 690 is all about ultimate performance and really needs to be tested when having to run three monitors in tandem; it scoffs at a single-screen, 1,920x1,080 resolution. We've described the £325 GTX 670 as the best premium card around, once value is factored into the mix, so how do two perform when in SLI, how close do they get to the GTX 690 juggernaut, and are two second-rung Kepler cards good enough to provide a smooth experience when gaming on three screens? Only one way to find out, right?

Dust off the test system, Sandhu

GPU comparisons

Graphics card GPU clock
(MHz)
Stream
processors
Shader clock
(MHz)
Memory clock
(MHz)
Memory bus
(bits)
Graphics driver
NVIDIA GeForce GTX 690 (2 x 2,048MB) 915+ 3,072 915+ 6,008 256 x 2 ForceWare 301.33
NVIDIA GeForce GTX 670 SLI (2 x 2,048MB) 915+ 2,688 915+ 6,008 256 x 2 ForceWare 301.33
NVIDIA GeForce GTX 670 (2,048MB) 915+ 1,344 915+ 6,008 256 ForceWare 301.33

HEXUS high-end test bench

Processor Intel Core i5-2500K (3.30GHz, 6MB smart cache, quad-core, LGA1155)
Motherboard Intel Desktop Board DP67BG
Memory 8GB Corsair Vengeance (9-9-9-24 @ 1,600MHz)
Power Supply Corsair AX750W
Monitor Dell 30in 3007WFP
Disk drive(s) Crucial RealSSD C300 (256GB)
Chassis Corsair Graphite Series 600T
Operating system Windows 7 Ultimate (64-bit, SP1)

HEXUS high-end benchmark suite

Aliens vs. Predator DX11, 1,920x1,080, 2,560x1,600, and 5,760x1,080 resolutions, 4xAA, 16xAF, very high quality
Battlefield 3 DX11, 1,920x1,080, 2,560x1,600, and 5,760x1,080 resolutions, 4xMSAA, 16xAF, ultra quality
Batman: Arkham City DX11, 1,920x1,080, 2,560x1,600, and 5,760x1,080 resolutions, 8xMSAA, extreme quality
Crysis 2 DX11, 1,920x1,080, 2,560x1,600, and 5,760x1,080 resolutions, 4xAA, ultra quality
Just Cause 2 DX10, 1,920x1,080, 2,560x1,600, and 5,760x1,080 resolutions, 4xAA, 16xAF, very high quality
Total War: Shogun 2 DX11, 1,920x1,080, 2,560x1,600, and 5,760x1,080 resolutions, 4xMSAA, 16xAF, high quality
Power Consumption To emulate real-world usage scenarios, we record mains power draw both when idle and whilst playing Batman: Arkham City
Temperature To emulate real-world usage scenarios, we record GPU core temperature both when idle and whilst playing Batman: AC
Noise A PCE-318 meter is used to record noise levels when idle and whilst playing Batman: AC

Notes

Theory tells us that two GeForce GTX 670s should not be as rapido as GTX 690 Beast™. Both are based on the same architecture, GK104 Kepler, but GTX 690 uses it in its fullest sense. GTX 670, meanwhile, drops one of the SMX units and therefore a couple of cards, combined, still have fewer shaders/cores - 2,688 vs. 3,072. But it isn't as clear cut as making predictions on these numbers alone. Our two GTX 670s have a tendency to GPU Boost to 1,050MHz-plus in each of our games; GTX 690's auto-overclocking isn't as robust, meaning the scores may well be closer than you suspect.

We've benchmarked our chosen trio at three resolutions. Of particular importance is the 5,760x1,080 setting, made up of three BenQ screens placed alongside each other, and we've logged the per-second and per-frame framerates at this resolution and graphed/tabled them up appropriately.

We'll be able to observe the scaling from one GTX 670 (£325) to two (£650) and also compare their performance against the £900 Beast™.