The GF110 chip shares a lot of its roots with the GF100 core. The GPU still packs an impressive 3.2 billion transistors in a die seize 520mm squared. Its default internal configuration has been ported from GF100 in that the GPU features 512 CUDA cores or shader processors as we are more accustomed to. There are organised in a 4x16x32 block where 16 denotes the number of stream multi-processors (SM) and 32 being the number of cores on each SM. Each SM has its own L1 Cache and a common L2 Cache is shared by all 16. The GPU is still manufactured on the 40nm process.


With the GTX480, only 15 of the total 16 SM were enabled to give it 480 CUDA cores and consequently, it boasted 60 texture units. The GTX580 uses a fully-fledged GF110 GPU and thus retains all 16 SM with none being disabled. The result is 512 fully operational CUDA cores and since the number of texture units is based on the number of stream-multiprocessors, the GTX580 sports 64 units instead of the 60. The number of ROPs has remained the same at 48 and the L2 Cache is still 768 KB. In terms of shader and texturing power, having 512 CUDA cores boosts performance by 6.6% over 480 CUDA cores. Nvidia have gone further with the GTX580 though. Not only have they increased the clock speed on the GTX580 compared to the GTX480, but they have also employed the GF104’s superior texture filtering capabilities in their tweaked GF110 GPU. Improvements have also been done to the tessellation engine so the already very capable performance that is expectant of the Fermi architecture has gone a step further.

The number of memory controller on the new GF110 GPU has remained untouched so the GTX580 has 6 memory controllers each sporting a 256MB framebuffer to give 1536MB of available memory. The GTX580, as with the GTX480 it is replacing, has a 384-bit memory interface that has been teamed up with GDDR5 memory for unmatched bandwidth.

GTX480: 15 SMs x 32 = 480 Core, 15 SMs x 4 = 60 texture units
32-bit channel x 12 = 384-bit interface, 256MB x 6 memory controllers = 1536MB memory

GTX580: 16 SMs x 32 = 512 Core, 16 SMs x 4 = 64 texture units
32-bit channel x 12 = 384-bit interface, 256MB x 6 memory controllers = 1536MB memory


At the same clock speed, the extra 32 CUDA cores alone give the GTX580 a 6.6% shader and texturing performance advantage over the GTX480. Given the 10% and 8% increase in core clock speed and memory clocks respectively, we can expect the GTX580 to widen its lead. Looking at the texture fill rate and shader operations, the GTX580 theoretically provides 17% more performance that the GTX480 it replaces, not a bad upgrade considering that the price points at launch are the same. At a retail price of $499, it is currently the most expensive graphics card on the market but it is yet to be seen if the price premium is worth the performance benefits.