nVidia has revealed the first titbits of information regarding its next-gen processor at the its GPU Technology Conference, which is being held in San Francisco. They decided to focus on the GPU computational capabilities, with little info on the specification and actual graphical performance of the chip.
Thanks to TechReport we know that the chip sports 16 steam processors and 512 “CUDA cores” (the GT200 had 240). Based on clock speed estimates, throughput should be around 760 GFLOPS, roughly a 50% increase over the Radeon HD 5870.
Each CUDA core features a double precision mathematical processing unit, which may not be ideal for real-time gaming applications, but is important for certain data critical processing applications, such as protein folding.
The memory bus is reduced to 384 bits, down from 512 in the GT200. Fermi makes up for the loss by doubling the bandwidth per pin via GDDR5 memory support. Making some assumptions, if a 4.8 Gbps rate is used for the GDDR5, the peak memory bandwidth could be around 230 GB/s, roughly 50% higher than the Radeon HD 5870, which has a memory bus width of 256 bits.
nVidia has estimated the Fermi to have 3 billion transistors, but hasn’t given information on the die size. Taking the Cypress as a baseline, assuming a 40nm process and a similar transistor density the Fermi could be around 460mm² in size.
Big features include the ability to simultaneously run multiple kernel threads – an ability already present on the Cypress – improved OpenCL and DirectCompute APIs, and IEEE 754-2008-compliant mathematical precision (the double precision in the CUDA cores). The GPU Technology Conference continues until the 2nd of October.
Based on these specifications, Fermi is expected to outperform AMD’s recently released hi-end product; the HD 5870. Likely to be dubbed the GTX 300 series, Fermi will mark Nvidia’s entrance to DX11 gaming and is expected to be released before the end of the year.
Discuss Fermi in the forums
Loading ...
