Supermicro New Zealand
NVIDIA Tesla K10 for GPU Computing
30 June 2012
NVIDIA has recently released their latest scientific GPU co-processor, the Tesla K10 GPU accelerator, designed to address the needs of two high-performance computing (HPC) markets: seismic data
processing in oil and gas exploration, and signal and image processing
in the defense industry. In this newsletter we would like to address two points of interest: (1) What is new about the K10 architecture? (2) How does it compare with the previous generation and what should we expect in the near future?
The latest generation GPU from NVIDIA is named Kepler and will eventually replace all products based on the previous generation Fermi architecture. The Tesla K10 is based on Kepler and is now the world's fastest GPU accelerator for several HPC applications. There are three main architectural differences:
(1) Architecture (SMX) - It has 7.1 billion transistors, made possible by the latest manufacturing process (a die shrink from 40 nanometer to 28 nanometer). This enabled NVIDIA to increase the amount of processing cores in the Kepler while reducing power usage by lowering the clock speeds of each core. The result - we get 3x the performance pet watt of energy consumed by the cards.
(2) Dynamic Parallelism - The Kepler GPU can now launch nested kernels, eliminating the need to communicate with the CPU as often. The previous architecture required the CPU to launch every kernel to the GPU.
(3) Hyper Q enables multiple CPU cores to launch work on a single GPU simultaneously, effectively maximising GPU utilisation time and slashing CPU idle times. This feature increase the total number of connections between the host and the Kepler GPU by allowing 32 simultaneous connections, compared with the single connection with Fermi.
NVIDIA has only released the Kepler GK104 core as of last month - these were designed for maximum single-precision math performance which benefits seismic, signal, image, and life sciences workloads. If you or your customer is currently dealing with such workloads, expect to see a 3x performance improvement over the previous generation top-end Tesla M2090 card.
If you are doing finite element analysis, computational fluid dynamics, various
physics simulations, and financial calculations and simulations that are
dependent on double-precision floating point math you should wait for
the Tesla K20 cards using the Kepler2 GPUs, expected to arrive in Quarter 4 of 2012.
Supermicro New Zealand
Managed by Compucon New Zealand