Based on information gathered so far about GT300/Fermi:
- Transistors ca 3 billion
- Built on the 40 nm technnology process
- 512 shader processors (cores)
- 32 cores per core cluster
- 384-bit GDDR5 memory
- 1 MB L1 cache memory, 768 KB L2 unified cache memory
- Up to 6 GB of total memory, 1.5 GB expected for the consumer product
- Half Speed IEEE 754 Double Precision floating point
- Native support for execution of C (CUDA), C++, Fortran, support for DirectCompute 11, DirectX 11, OpenGL 3.1, and OpenCL
Based on techPowerUp news
Read today good article about DirectCompute, OpenGL, DirectX 11, NVIDIA Cuda, ATI Firestream and other hot topics:
Up to 2,72 TFLOPS per card, new ATI Radeon HD 5800, fully Win 7, OpenGL supported and DirectX 11 compatible. This is the firs card to support Directcompute 11 standard. As we know, GT300 from NVIDIA will not be ready before December this year. So, ATI got some 4-5 month lead in next get DirectX 11 product development.
Until now, ATI has not performed well in Folding@Home, delivering lower performance levels per dollar as NVIDIA. Hope to see change in this area also.
Directcompute 11 is Microsofts application programming interface (API) that takes advantage of the massively parallel processing power of a GPU to accelerate PC application performance in Microsoft Windows Vista or Windows 7. DirectCompute is part of the Microsoft DirectX collection of APIs. This means, that ATI Radeon HD 5800 equipped PC fill be significantly faster in certain tasks, cause it’s streaming processors are used in parallel with PC-s main processor.
As You see from this image, new SGI Octane III has 10 blades, each 2 x 4-core Intel 5500 series CPU’s, total 80 cores per on small rack. Pretty neat supercomputer. All Blades are interconnected by Gigabit Ethernet or InfiniBand. Supercomputer supports Red Hat or Suse, SGI Isle cluster manager and Altair PBS patch scheduler.
GT300 announcement is going to be in December – NVIDIA head during October will visit TSMC for checking the readiness degree of this video chip for mass deliveries according to several sources.
NVIDIA will simultaneously discuss the prices of 40 nm production , which is released by TSMC. During July the last company was possible to raise the release level of suitable 40 nm chips from 30% to 60%.
GT300 will be the most powerful CUDA chip ever. Based on technicad data about NVIDIA GT300 shader units frequencies, memory speed and other, people fom Hardware-Infos (Germany) have decided, that G300 performance will be 2,4 TFLOPS. It’s more than twice GTX280 cards 0,933 TFLOPS. Comparision is not direct, as G300 supports MIMD architecture, but GTX280 only SIMD. GPU is rumoured to have rather tremendous die area of 452mm², which means that only about 104 – 105 of such dies can fit onto one 300mm wafer.
Nvidia received first working samples from TSMC (Taiwan Semiconductor Manufacturing Company) in the beginning of September.
After long summer vacation I restarted my GPU cluster @ 166K PPD.
ca 165 000 - 166 000 PPD