More details of NVIDIA Tesla cGPU K20 “GK110″
Yesterday, Nvidia announced its new Tesla cGPUs K10 and K20 Tesla , Tesla its first products based on its architecture Kepler, aimed at intensive parallel applications for server segment and super-computers.
From VR-Zone we get some additional details of the future Tesla K20 based on the new core 2 nd generation Kepler called GK110.
GK110 will consist of a total of 7100 million transistors and like the current Kepler GPUs will be manufactured with the 28nm manufacturing process at TSMC. GK110 maintains the basic structure of the Kepler architecture, which has been reorganized to provide more performance intensive applications that use estimates of single and double precision.
The SMX (Next Generation Streaming Multiprocessors) remain the modular unit based on Kepler’s 2 nd generation, and continue to be shaped by 192 ALUs (Nvidia calls them CUDA cores) and an even undisclosed SFUs (Special Function Units) and units Load / Store, but unlike GK104 where each GPC (Graphic Processing Cluster) is comprised of 2 SMX, the CPGs are composed of GK110 SMX for 3 each, where we assume that each SMX should have a greater number of units dedicated to FP64 calculations (double precision) and GK110 is comprised of a total of 5 CPGs, this gives us a total of 2880 ALUs (shader processors).
GK110 is estimated to have a power calculation in double precision equivalent to between 50% to 85% power in single-precision (FP32).
The memory controller of GK110
Nvidia has not been very clear with that, but GK110 has a total of 6 drivers dual channel memory (64 bits) which total a 384-bit bus, Tesla K20 is available in versions with 6GB, 12GB and 24GB of ECC memory GDDR5.
GK110 introduces some new features, among them we can mention the following:
- Virtualization Silicon GPU (GPU-assisted virtualization hardware).
- Hyper-Q (greater use of multi-tasking GPU).
- Dynamic Parallelism (greater simplicity in programming the GPU and less dependence on CPU).
The die of GK110
Nvidia reveals itself most detailed images of GK110 die, thanks to the users of the Beyond3D forums have been analyzed and compared with those of GK104, showing the computation units:
GK110 is estimated to have an approximate area of 550mm.
We have prepared a table of specifications known GK110:
GK110 on Geforce and Quadro future
Nvidia will launch its Tesla cGPU GK110-based K20 during the last months of this year, so do not expect any version Geforce or Quadro GK110 based on well into next year, which will become the maximum exponent of the future family Geforce 700 Series products.