|
Today appeared the detailed characteristics of Tesla K20 - calculations accelerator, based on the new processor GK110, which contains 7.1 billion transistors. Recall that for the GK104, the figure is 3.54 billion transistors. Unfortunately, Tesla K20 will be released only in the fourth quarter this year, Judging by the specs, the processor is more focused on the computing market.
In particular, the performance level in floating point double precision will be three times higher (more gigaflops) than in generation Fermi (Tesla M2090), but only in terms of watt. As explained NVIDIA, the effectiveness of streaming multiprocessor SMX has been raised with respect to Fermi by fourfold greater of CUDA cores number while reducing the frequency of each core, powering down GPU parts and increase the area of GPU, designed for parallel computations for the cores instead of logic control . Thus, GK110 in total may include from 1920 to 2880 cores CUDA, if every SMX have 128 and 192 cores, respectively.
According to preliminary data, GK110 will have 384-bit memory bus, NVIDIA itself does not specify this characteristic. The amount of memory can be equal to 6, 12 or 24 GB. Everything depends only on how much it will cost . In the first supercomputer based on Tesla K20 will be applied solutions with 6 GB of memory.
Tesla K20 is focused on the use of server systems, there are six-and 8 pin power connectors. Power consumption should not exceed 300 watts, though it may be lower.
At the architectural level NVIDIA Tesla K20 is ready to offer the following innovations:
- SMX streaming multiprocessor - As a basic building material for each GPU, SMX streaming multiprocessor has been created from scratch for high performance and efficiency.
- Dynamic Parallelism- This function allows threads to GPU dynamically generate new streams to dynamically adapt to the data. The new technology greatly simplifies parallel programming by using GPU-acceleration for a wide range of popular algorithms such as adaptive mesh refinement, fast multipole methods.
- Hyper-Q- This feature allows multiple CPU cores simultaneously to use the same core CUDA GPU Kepler. The load on the GPU grows significantly decreases . Hyper-Q - is the ideal solution for cluster tasks using MPI.
Tesla K10 only supports SMX, two other features are unique to Tesla K20. Related Products :
|