We have a saying up here, "we have many names for the things we love", and what NVIDIA calls Fermi has been known as GT300 and GF100 in the news, but now we know even more. Earlier his week new information hinted that NVIDIA's new GPU would consist of no less than 3 billion transistors and that seems to be very much the case, which would make Radeon HD 5870 and its 2.15 billion transistors seem pretty puny. But there's more.
According to the story published at BSN.com the specifications are as follows;
- 3.0 billion transistors
- 40nm TSMC
- 384-bit memory bus
- 512 shader cores [now called CUDA Cores]
- 32 CUDA cores per Shader Cluster
- 1MB L1 cache [divided into 16KB Cache - Shared Memory]
- 768KB L2 unified cache
- Up to 6GB GDDR5 memory
- Half Speed IEEE 754 Double Precision
The transistor count aside we we see that the shader count, or CUDA cores, have increased drastically since GT200. Up from 240 to 512 pieces. This is far from the 1600 stream processors AMD has with RV870, but this is a very small indication of performance since NVIDIA and AMD use completely different shader architectures.
The wide 384-bit memory bus will be paired with up to 6GB GDDR5 memory, but you should perhaps first of all count on models with 1.5GB and 3.0GB memory.
Though the things that really stands out with GT300 is that it is starting to look more and more like a CPU, and with that said NVIDIA is working harder and harder to take over previously CPU specific workloads with C++ and Fortran support. Of course there is also support for DirectX 11, DirectCompute, CUDA, OpenCL and OpenGL 3.2.
Interestingly it seems like NVIDIA and Intel are meeting halfway in their plans for the future to create a hybrid for what we call CPU and GPU.
In a follow up post BSN has revealed the TDP of Fermi/GT300/GF100, and it looks like it is circulating around the 225W mark as the 6GB version fails to meet the requirements for 225W branding while lower capacity versions does. Expect 8+6-pin PCIe power connectors for the 6GB version and and 6+6-pin with the lower capacity models.
NVIDIA talked a bit more about GT300/Fermi in a press briefing where it showed something that looked like a factory built card, though its authenticity has been questioned (for a revamped GTX 285).
Picture courtesy of bit-tech.net
Comment 
Send to a friend
Related news:
2009-11-19 Galaxy planning dual-GPU NVIDIA card for Christmas
2009-11-18 NVIDIA Expands Award-Winning 3D Vision Technology To Notebook Market
2009-11-18 Picture of real Fermi card posted
2009-11-17 NVIDIA sneaks out GeForce GT240
2009-11-12 NVIDIA welcomes Lucid
GT300/Fermi specifications and power consumption
Noctua
