It’s been a long six-month wait and we finally have GF100, the first GPU based on the Fermi architecture that will power the next generation of GeForce cards.
The GeForce GTX 480 is Nvidia’s first card based on this new architecture. Here’s a brief look at what’s inside the GF100 before we get to the card itself. A full-blown GF100 features four graphics processing clusters (GPCs), each containing four Streaming Multiprocessors (SMs) and a Raster Engine. Each SM has 32 cores, a Polymorph Engine and four Texture Units. So in all, the GF100 features 512 cores organized in 16 SMs of 32 cores each, 16 Polymorph Engines and 64 Texture Units. These coupled with other units churn out all the eye candy and handle CUDA applications. Further, the GF100 has six GDDR 5 memory controllers that run on a 64-bit wide bus each, making for a total bus width of 384 bits. There’s also a 768 KB unified L2 cache that provides high speed data sharing (relating to physics, ray tracing, etc) across the GPU thereby improving the memory bandwidth efficiency.
The GF100 will be scaled down to create graphics cards at various performance levels and price points. Nvidia will simply cut down on the SMs, memory and memory controller units enabled for each spec. For example, the most entry level GPU might have just one SM with 32 cores and a 64-bit memory bus width.
What we have with us is a GeFore GTX 480 which is one out of the first two GF100-based GPUs. The other is the GTX 470 which we will get to shortly. The GTX 480 features 480 CUDA cores (15 SMs x 32 cores). Note that it has one SM less than the total capacity of GF100, which means you could expect an even faster GPU with 512 cores (could be a GTX 490, who knows). It has 1536 MB of GDDR5 memory that runs on 384-bit wide bus. The GPU is clocked at 700 MHz and the memory runs at 924 MHz. The rear panel has two DVI outputs and one mini HDMI output, and most brands will ship a mini-to-full size HDMI adapter in the box. Only two outputs can be used at a time for a dual-display setup unlike the Radeon HD 5000 series that supports triple-display using AMD’s Eyefinity technology.
We ripped open the card to get a good view of the PCB on which the massive GTX 480 and memory chips reside. First, the colossal cooler: the heatsink that covers the GPU is exposed and protrudes through the shell that covers the massive cooler. The finned aluminum heatsink has five heatpipes that dissipate heat from GPU to the fins. The shell comes off easily by releasing the clamps that hold it in place. We then got rid of the metal plate with the fan that draws heat from memory chips and voltage regulators on the PCB. Check out the size of the GTX 480 in proportion to the size of the card—it’s a massive 529 mm2 die (23 mm x 23 mm)! The card draws power from one 8-pin as well as one 6-pin PCIe power connector. Also note the cutout on the back of the PCB to allow the fan to suck in more air—a clever way to facilitate airflow, especially in an SLI setup wherein there’s hardly any space at the top for the card to breathe.
The first retail sample of the GTX 480 was sent by Zotac prior to which we received a demo unit from Nvidia. So we got a golden chance to club the two in SLI and check out the results. We tested the card on a Gigabyte P55A-UD7 motherboard which supports SLI in dual-x16 mode as well as 3-way SLI in triple-x8 mode. The processor used was an Intel Core i7-870 along with 4 GB DDR3-1333 memory and an 80 GB Intel X25-M SSD. We needed a 1300W Tagan PSU to power our rig.