Inno3D provided us with their branded card for this review and on the front of the packaging we are given some basic feature information, a note of the bundled items and a sticker has been applied which tells us this is an overclocked model. Inside consumers will find a Molex to PCIe power convertor, software disc, short manual, StarCraft 2 trial and a free cloth/foam mouse pad.
The card itself follows the NVIDIA reference design which means a dual slot cooler and 10.5” black PCB. Inno3D have added their own branding to the heatsink shroud and fan which also has the same ventilation holes which were included on the 580 version. Flipping the card over we find a relatively standard PCB design with two SLI connectors in the top right corner which tell us this card supports 3-Way SLI.
Underneath the cooler, which has received a complete redesign since the 400 series (see below), is a GF110 core. This is a 40nm part and an evolution of the original GF100 used on the GTX 480. In creating the GTX 570/580 NVIDIA have made a number of changes and enhancements to the design which benefit performance.
Firstly the GTX 500 based cards support full speed FP16 texture filtering which enhances performance in texture heavy applications. Next is the use of new tile formats which improve Z-cull efficiency and then we have the more obvious changes such as additional CUDA cores and texture units in addition to a clock speed bump over the GTX 480/470. All of these combine to make a product which, in the case of the GTX 570 and based on NVIDIA figures, can be over 25% faster than the GTX 470.
While the cooler may look similar to previous high end GTX 400 models it significantly different beneath the shroud. The differences start with the fan where the centre section has eight holes which pass air to the inside of the fan, maximising airflow around this area. NVIDIA also state that they have optimised the fan design to reduce noise and vibrations as well as smoothing out the fan profiles to ensure a less intrusive experience as the fan speed changes.
Then, sitting on the GPU, we have a vapor chamber cooler. Essentially this is a copper GPU block which sits beneath a chamber filled with water. As the GPU heats up the water turns into vapour, moves the heat away taking it to the condenser plate where it liquefies again. Heat is drawn away from the condenser plate to the aluminium heatsink fins which have air passed through them by the fan.
One of the other significant changes on the GTX 500 is the use of new power circuitry which sits underneath the fan area. The three chips positioned here and these report the status of the 12v rails to the driver. The driver is then able to dynamically adjust the cards performance in specified applications (e.g. Furmark or OCCT) if the power level exceeds NVIDIAs certified level. This is a similar type of feature to the temperature monitoring which has been present in graphics cards for some time now, NVIDIA are essentially applying the same protection to the power monitoring and they state that at this time there are no games specified in the driver so we receive full gaming speeds at all times on GTX 500 and we can overclock in this scenario without driver limitations also.
Rounding off the power level changes to the GPU we have a re-engineering of the GPU down to the transistor level. Essentially NVIDIA looked at the design of the GTX 480/470 which used fast, leaky transistors throughout, and replaced the transistors on paths which were not timing sensitive with low leakage versions. This reduces the power loss/requirement of the GPU and allows it to run faster than the GTX 480/470 but with a lower power draw, even with the additional CUDA cores and Polymorph Engines (which enhance tessellation performance).
So that’s the ins and outs of the GTX 500 series in general so back to the Inno3D model where they have gone for the standard layout of 2xDVI and one mini HDMI. These can power two displays at any one time and when the card is combined in SLI we are able to use three displays. Reference design GTX 570s require two 6-pin PCIe power connectors and that remains the same on this overclocked version.
The core on this model is set to 800MHz, 1000MHz is the memory speed and 1600MHz is the shader clock. This is of course on the 480 CUDA cores which connect to the 1280MB of GDDR5 via a 320-bit memory bus. PCIe 2.0, DirectX 11, DirectCompute 5.0, CUDA, PhysX and OpenCL are all supported on this card as is acceleration of high definition content and the card can output 7.1 audio over HDMI.
For those interested in the specifications of the reference design cards, they use the same core with its 60 texture units and 40 ROPs but the memory runs at 950MHz, core at 732MHz and CUDA cores at 1464MHz.