It’s no understatement to say that NVIDIA’s GeForce GTX 980 took the PC gaming world by storm in August. Raw performance for the new cards was of course impressive, matching or surpassing GPUs priced considerably higher, but the largest stride NVIDIA’s Maxwell architecture took was in the arena of power efficiency, trouncing even Kepler by 2:1. As the flagship and its lesser sibling ruled the roost a new question emerged – how well does could the new architecture scale? Today we find out as NVIDIA release the GeForce GTX 960.
NVIDIA describe their GTX x60 class of GPUs as fitting into a performance and price ‘sweet spot’, and that’s a pretty fair assessment. Since the GTX 460 each GPU of this class has been popular with gamers who opt for a more balanced system configuration over investing in an especially high-end GPU. That’s not to say that it’s especially budget-oriented – in previous generations the x60-class has approached £200 at launch – but it was excellent bang-for-your-buck when playing the latest games at 1080p.
Even amongst this class of cards the GTX 660 has been especially noteworthy. The most popular gaming class GPU according to the Steam Hardware Survey, it will cast a long shadow as NVIDIA transition the entirety of their lineup to Maxwell. The performance envelope and price is going to be critical to tempt gamers satisfied with their older mid-range, and our take on this point is available in the companion reviews. But for those looking for a more general overview, read on.
ASUS GTX 960 STRIX DirectCU II Review
GIGABYTE G1 GTX 960 Review
At the heart of the Geforce GTX 960 is the new GM206 GPU, a design based on NVIDIA Maxwell architecture but of considerably smaller size than the GTX 980’s GM204. Although theorised as originally designed for 20nm-based manufacturing processes, Maxwell has proved highly successful in increasing power efficiency on even the now mature TSMC 28nm node. In terms of specifications the GM206 very much puts you in mind of half a GM204, but the key question that reviews will answer is how well performance scales for this new slim-line part.
The Kepler architecture organised elements within the GPU into Graphics Processor Clusters, and each GPC into two streaming multiprocessors known as SMXs. Maxwell by contrast has a much more aggressive partition: four SSM’s of 128 CUDA cores each with yet further internal partitions.
Each Maxwell SMM is subdivided four ways into what can be thought of as a further sub-SM, such that each SMM now has four sets of instruction buffers, warp schedulers, dispatch units and registers rather than a single set per SMX in Kepler. This, in addition to offloading some scheduling to the CPU where appropriate, is why the card can be so power-efficient and perhaps more importantly make use of resources more intelligently.
One further improvement NVIDIA have made to the memory architecture is the implementation of Third Generation Delta Colour Compression. The new lossless algorithm allows them to make better use of the memory bandwidth available, which is essential for higher resolution rendering, texture sizes and IQ settings.
Click Here for larger image
Benefits realised by using this algorithm largely depend on the scene but NVIDIA estimate that the effective bandwidth is increased by between 17 and 25%. It’s for this reason that many reviews and overviews will express the GTX 960’s memory bandwidth as an effective 9.3GHz rather than actual 7GHz.
Reference specifications for the GTX 960 top out at 120W TDP, requiring just one 6-Pin PCIe power connector. Factory overclocked and overclockable variants are set to use us just one 8-pin power connector. Critically this means that those upgrading from previous mid-range cards such as the GTX 560, 660 or AMD equivalent won’t need to change PSU despite large performance gains.
Additional Features: DSR and MFAA and more
As well as supporting up to four monitors, and for the first time will bringing HDMI 2.0 to the mid-range, major new driver-level technologies that were introduced alongside the GTX 980 make a repeat appearance with today's GTX 960.
As a form of super-sampling, Dynamic Super Resolution is a driver-level feature which allows excellent anti-aliasing in games with relatively low performance overhead or poor even-level anti-aliasing support. By essentially rendering a scene at higher resolutions and then downscaling to the native resolution of your monitor, DSR greatly reduces the appearance of 'jaggies' but at a hefty performance cost. Aimed very much at the higher performance bracket, it's not yet clear how well suited DSR is to the GTX 960; however it's certainly a welcome free addition.
Multi-Frame Temporal Anti-Aliasing by contrast is a big addition to the mid-range. Only compatible with Maxwell-based GPUs thanks to a change in the underlying operation of memory hardware, MFAA performs anti-aliasing operations of a frame by comparing it to the preview output frame. Normally this would be a recipe for aliasing artefacts due to pixel sample positions being identical between frames, however Maxwell is able to sidestep the problem by randomising the sample positions on a frame-by-frame basis. As no part of the scene is being re-rendered MFAA is potentially far more time efficient than MSAA (the most common alternative) yet has higher fidelity than the previous in-house FXAA technique.
Voxel Global Illumination (VXGI) is the last of the big three, and unlocks more accurate simulation of lighting, including diffuse lighting, specular lighting and reflections. VXGI is being baked into NVIDIA GAMEWORKS as well as an optional component for Unreal Engine 4, and Maxwell is the first architecture to support it in hardware.
Although we’re utilising reference specifications above, it’s worth noting that there is no specific manufactured reference design for the GTX 960. NVIDIA’s graphics partners, i.e. GIGABYTE, MSI, EVGA, ASUS etc., will instead be producing their own designs based on their own manufacturing strengths and target markets. Not only will size and cooling vary significantly between the different brands, so will the factory overclock on each GTX 960 SKU.
Of course, the horsepower available to the GTX 960 won’t be needed in every game. Some titles, especially those frame-limited internally or via V-Sync, place a very low overall load on the GPU. Recognising this NVIDIA have implemented a new low power mode which shuts off the GPU fan entirely when not needed, relying instead on the passive cooling of the GPU cooler and case air flow.
This low power mode will only be supported by certain manufacturers who deem their cooling solution suitable, but could be a boon to gamers seeking the truest low-noise environment for their gaming setup. It will be interesting to see if this new tech. is put in place for NVIDIA’s future performance parts, and indeed if totally passive solutions are possible for the GTX 960 as it has been for the GTX 750Ti.
Finally, we come to the price: NVIDIA’s MSRP for the GeForce GTX 960 is £159.99, which if we’re honest is surprisingly aggressive. At the time of writing the GTX 760, which the 960 ostensively replaces, hits £179.99 at many retailers and AMD’s nearest comparable product is the Radeon R9 285 and has a similar price tag. With the release of the GTX 960 we could be looking at as significant a shape-up for the graphics card market as that which occurred upon the release of the GTX 980 and 970.
Looking ahead, there's an awful lot of room between the GTX 960 and 970 for a Ti variant to slot into. Furthermore the GTX 950 lineup also has yet to be released and might feature either a rehashed GM107 or slim-line GM206. And of course we can't forget the rumoured 'Big Maxwell' GM200. There are certainly interesting times ahead for NVIDIA.