Nvidia has revealed its Fermi architecture at its GPU Technology Conference in San Francisco. Nvidia has revealed the first wave of chip details for the Fermi architecture and praised its GPU-computer capabilities.
The Fermi architecture features several features that have been seen in GPUs before. These features will open new frontiers for GPU computing and create new horizons of Nvidia’s already existing GeForce and Tesla product lines.
The Fermi architecture will have 16 streaming multiprocessors along with 512 discrete cores–more than twice the number of CUDA cores of the GT200.
The Fermi architectures makes use of 64-bit interfaces and six DRAM interfaces. This gives the GPU a total path to memory that is 384-bits in width. When compared to GPUs such as the GT200, this is fewer but it is compensated by delivering nearly two times the bandwidth per pin by using GDDR5 memory.
As of now very little information is available and so it is difficult to tell if Fermi would be able to compete with ATI’s Cypress or not. Nvidia has not made crucial information such as the clock speed of the Fermi GPU public.
Third party industry experts speculate that the Fermi GPU will run at about 1500Mhz–a quite reasonable frequency target for Fermi’s stream processing core. The GeForce GTX 285 also runs at the same frequency. If the Fermi GPU reaches this frequency, its throughput will approximately be about 1536 GFLOPS, which is about half the speed of the ATI Radeon HD 5870.
If Nvidia uses the same 4.8Gbps data rate for GDDR5 memory that AMD uses for Cypress, the Fermi GPU will have a peak memory bandwidth of 230Gbps. This is fifty percent more than that of the Radeon HD 5870, which uses a 256 bit memory bus.