Nvidia GeForce GTX 960 Review: feat. Asus

January 22, 2015 | 14:56

Tags: #best-gtx-960 #directcu-ii #gm206 #gtx-960 #maxwell #strix

Companies: #asus #nvidia

The GM206 GPU

As with the GTX 970 launch, Nvidia has not produced a reference board, and went straight to a virtual design. As such, for water-cooling enthusiasts out there, you'll need to do some careful research to find compatible waterblocks. This is also a hard launch, with multiple SKUs from numerous board partners available immediately. We're featuring Asus for this launch article, with more to follow in the coming days and weeks.

 Nvidia GeForce GTX 980 4GBNvidia GeForce GTX 970 4GBNvidia GeForce GTX 960 2GBNvidia GeForce GTX 660 2GBNvidia GeForce GTX 750 Ti 2GB
GPU
ArchitectureMaxwellMaxwellMaxwellKeplerMaxwell
CodenameGM204GM204GM206GK106GM107
Base Clock1,126MHz1,050MHz1,127MHz980MHz1,020MHz
Boost Clock1,216MHz1,178MHz1,178MHz1,033MHz1,085MHz
Stream Processors2,0481,6641,024960640
Layout4 GPCs, 16 SMMs4 GPCs, 13 SMMs2 GPCs, 8 SMMs3 GPCs, 5 SMXs1 GPC, 5 SMMs
Rasterisers44231
Tesselation Units1613855
Texture Units128104648040
ROPs6464322416
Transistors5.2 billion5.2 billion2.94 billion2.54 billion1.87 billion
Die Size398mm2398mm2227mm2221mm2148mm2
Process28nm28nm28nm28nm28nm
Memory
Amount4GB GDDR54GB GDDR52GB GDDR52GB GDDR52GB GDDR5
Frequency1.75GHz (7GHz Effective)1.75GHz (7GHz Effective)1.75GHz (7GHz Effective)1.5GHz (6GHz effective)1.35GHz (5.4GHz effective)
Interface256-bit256-bit128-bit192-bit128-bit
Bandwidth224GB/sec224GB/sec112GB/sec144GB/sec86.4GB/sec
Card Specifications
Power Connectors2 x 6-pin PCI-E2 x 6-pin PCI-E1 x 6-pin PCI-E1 x 6-pin PCI-ENone
Stock Card Length267mm267mmN/A241mm147mm
TDP165W145W120W140W60W

The reference display outputs from the GTX 980 are carried over, namely three DisplayPort 1.2 connections, a dual-link DVI-I and HDMI 2.0, which is a newcomer for this market segment. The GTX 960 also features Nvidia's new display engine, with support for up to 5K resolutions and a maximum of four monitors, including four 4K MST displays. A new addition is H.265 encode and decode (previously encode only on the GTX 980), which may appeal to home threatre users, as well as native support for HDCP 2.2 content protection over HDMI.

On to the nitty gritty details, the GTX 960 utilises a new GPU called GM206, the direct successor to the Kepler-based GK106 used by the GTX 660. It is of course a 28nm part with 2.94 billion transistors and a 227mm2 die size. It uses the same Maxwell streaming multiprocessor (SMM) design introduced with the GTX 980 (see here for the full details), whereby each SMM has four blocks of 32 CUDA cores. GM206 has two graphics processing clusters (GPCs) with four SMMs apiece, for a total of 1,024 CUDA cores and 64 texture units. This is only 7 percent more cores than the GTX 660, and 128 less than GTX 760 (and less texture units than both too). However, the efficiency of the SMM design means the cores are utilised a lot more, with Nvidia saying that each GM206 core does approximately 1.4 times the work of a GK106 core, and delivers roughly twice the performance-per-watt. The reference base clock speed is 1,127MHz, with a rated boost clock of 1,178MHz – expect this to vary significantly depending on board partner and SKU.

Nvidia GeForce GTX 960 Review: feat. Asus Nvidia GeForce GTX 960 Review - The GM206 GPU
Click to enlarge - The GM206 die shot

As you can see, there are only two memory controllers in the GM206 GPU, giving it a 128-bit interface compared to 192-bit on GTX 660 and 256-bit on GTX 760. The use of faster memory is one simple way to help overcome this, with 7GHz being the default speed. L2 cache has also been massively increased, from 384KB to 1MB, reducing the number of calls to the GDDR5 memory. Speaking of, the GTX 960 has a 2GB frame buffer. No 4GB variants have yet been confirmed.

Finally, GTX 960 also utilises Nvidia's third-generation delta colour compression, a lossless compression algorithm that leads to, on average, 25 percent fewer bytes per frame and thus a memory interface that is a third faster than Kepler's (again, on average). This has led to Nvidia quoting its raw 112GB/sec memory bandwidth as 148.8GB/sec “effective”, which is a little higher than the GTX 660's 144.2GB/sec and thus not a downgrade – either way, the results should show us whether the lower interface is a limiting factor. Finally, with 16 ROPs per memory controller now, these have increased from 24 to 32 total.
Discuss this in the forums
YouTube logo
MSI MPG Velox 100R Chassis Review

October 14 2021 | 15:04

TOP STORIES

SUGGESTED FOR YOU