HP Builds GPU-Optimised Server

Hewlett-Packard has endorsed the use of graphics processing units (GPUs) for faster, cheaper scientific processing, and launched a modular processing system which uses them to pack 1 TeraFLOPS per unit of rack space.

GPUs are designed to increase graphics performance by handling the floating point operations associated with images, but have become increasingly popular as a low-cost way to provide scientific processing. Most GPUs have originally been designed for consumer gaming machines, but interest in GPUs is growing as they are now being built into business grade equipment.

A GPU system in the top five supercomputers?

The HP ProLiant SL6500, announced today in Barcelona, is a blade server system, whose modules include the SL390s, a blade which can have up to three GPUs in it (one GPU is shown on the left). This adds up to one TeraFLOPS – a thousand billion floating point operations per second.

While vendors have been building systems that can include GPUs, this is the first system from a mainstream vendor, that is optimised for GPUs, said Mark Potter, HP vice president for enterprise servers.

Dell has launched a chassis designed for GPUs using the PCIe expansion bus, but HP claims to have gone several steps further in its optimisation.

“A team in Tokyo is building a supercomputer with the SL390s, which will be in the top five supercomputers in the world, and could even be in the top two,” said Potter. “It will deliver two PetaFLOPS.

One SL6500 chassis can hold 100 GPUs, providing 77 TeraFLOPS, which would put it in the top 100 supercomputers in the world – at a fraction of the cost, said Potter. “These are game changers.”

“The use of GPUs in scientific computing started two or three years ago,” said Peter Groth, product line manager for industry standard servers. “We are bringing it to the mainstream. This server is designed for it and we can bring to bear all the strength of ProLiant and our ecosystem.”

The units are designed to fit well into HP’s POD data centres, and several already use the 6500 system. SL390s units include 10GB Ethernet, Fibre Channel and management connections. Compared with existing scientific computers it would use one third of the power and one quarter of the space, said Potter.

The popular Matlab software from Mathworks already has a version which takes advantage of GPUs, so most Matlab-based scientific programs can run virtually unchanged on the new systems, said Groth, running a demonstration where a 390s with 3GPUs did, indeed – with some impressively whirring fans – deliver a TeraFLOPS running the Linpack benchmark program.

As Groth pointed out, this was running in a hotel room, with warm air, not a data centre environment. “It’s truly mainstream,” he said.

Peter Judge

Peter Judge has been involved with tech B2B publishing in the UK for many years, working at Ziff-Davis, ZDNet, IDG and Reed. His main interests are networking security, mobility and cloud

View Comments

Recent Posts

EU Widens Investigations Into Chinese Imports, Subsidies

After the United States imposes 100 percent tariffs on certain Chinese goods, Europe widens its…

4 hours ago

Reddit Deal With OpenAI Gives ChatGPT Access To Content

OpenAI strikes deal with Reddit to train its AI tech on user posts and give…

5 hours ago

Microsoft Invests 4 Billion Euros In France For AI, Cloud

Global spending spree from Microsoft continues, with huge investment for new data centre to drive…

8 hours ago

Toshiba Axes 4,000 Staff In Post-Delisting Restructuring Operation

Workforce blow. Newly privatised Toshiba has embarked on a 'revitalisation plan' that will entail the…

1 day ago

European Union Opens Child Safety Probe Into Meta

European Commission opens an official child safety investigation into Facebook and Instagram-owner Meta Platforms

1 day ago