GPU Computing For Servers
16 December 2009 Source: In house
The term personal supercomputer has been coined to describe the gains made possible by accelerating specifically parallel processing bound applications.
Using this technology it is possible to have 4 Teraflops of processing power in a single workstation system, but how did this massive technology leap come about?
Essentially this was a logical step made by these vendors who are more typically better known in the graphics processor markets, powered by the highly versatile architecture of their GPU's.
Over many years the GPU architecture has evolved considerably to become an extremely powerful processing component. One of the key
features of a modern GPU is that it can incorporate hundreds of shaders which in a graphics application would be used to generate visual aspects - GPGPU takes these shaders and effectively utilises them as computing cores.
The large core count coupled with onboard memory utilising cutting edge technology provides excellent memory bandwidth making the GPU an extremely serious consideration.
This architecture can bring massive advantages to compute applications where hundreds of threads can be opened and run simultaneously meaning parallel applications can be completed in record times.
In some cases applications have been accelerated by as much as 150x when compared to a standard CPU; actual gains can vary dramatically depending on the level of optimisation and code suitability however.
A GPU can have up to 60x more cores than a CPU
It's important to recognise that not all code can be run and accelerated by GPGPU, applications that are not highly parallel are usually better suited to CPU processing or can only be accelerated in part by GPGPU. As a result it's often better to have both GPU and CPU available simultaneously in order that they can complement each other and allow the application to receive the benefits of both worlds.
Additionally those applications which are positioned to take advantage of GPU also need to be coded and compiled so that they can appropriately run on GPU.
To facilitate this nVIDIA promote the well known and widely accepted CUDA architecture to help programmers experienced with C to write in a familiar environment. Other architectures are emerging too which should allow for a vendor agnostic approach; OpenCL and Direct X Compute are tipped to be used more frequently in future applications for compute and even in everyday computing.
Venom T4000 GPU Personal Supercomputer
- 4 TFLOPS of computational power
- Support for over 960+ GPU cores
- Intel Xeon (Nehalem) processors
- 1400W high efficiency redundant PSU
Venom T1000 GPU Supercomputer
- Two TFLOP GPU's in a 1U enclosure
- Multi-core architecture delivers optimum scaling across HPC applications
- The worlds 1st & only GPU optimised 1U server
- Gold level 93% high efficiency power supplies
These cutting edge server and workstation solutions combine standard x86 processing with the massively parallel processing power of up to either 2 or 4 GPGPU cards.
The Venom T1000 can process at 2 Teraflops using 2 nVIDIA Tesla C1060 cards. It too can have dual Intel Xeon 5500 processors and 48GB of DDR- 3 ECC Registered memory.
Using 4x nVIDIA Tesla C1060 cards the Venom T4000 can process at 4 Teraflops of computational power and arrives with 960 GPGPU cores. In addition it is also possible to equip dual Intel Xeon 5500 processors and up to 96GB of high performance DDR-3 ECC registered memory.
Utilising these technologies it is entirely possible to have 4 teraflops of processing power in a tower system that fits neatly under a desk. No longer is it necessary to purchase a large and expensive cluster of power and cooling hungry x86 systems in order to run complex parallel tasks in a reasonable time frame.
The processing power which is needed by the end user is brought directly to them for a fraction of the price and is made permanently available; no queuing for time on the cluster or sending jobs for overnight processing; the personal supercomputer is certainly here today!
Some examples where GPU compute has already made significant gains
