By differentiating their technologies to fit data-intensive workloads in HPC, smaller processor companies are giving IBM and Intel a run for their money. Robert Roe examines some of the technologies on display at the ISC High Performance event in Frankfurt last week.
Data centric workloads are growing in importance in high-performance computing and, in an industry that has been dominated by a handful of technologies for several years, this has led users to look for new technologies which better suit such jobs. The demand now is for low power, high memory, and I/O-intensive solutions, so there is a growing niche which can be addressed by solutions which are less focused on Flops performance.
The exhibition at the ISC High Performance event in Frankfurt last week provided the idea showcase for what is being achieved in HPC technology. This complemented the conference, which offered a platform to air policy issues such as Why do smaller companies shun HPC? and national policy strategies for Easing access to HPC for the SME.
ARM processors and IBM’s OpenPower have so far gained the most traction in the HPC market, but other technologies such as Altera’s FPGAs also have the potential to disrupt the status quo.
IBM has seen its OpenPower framework rise quickly into the HPC limelight after it acquired two of the three contracts in the US Department of Energy’s Coral programme earlier this year.
In an interview at ISC, the company’s newly appointed VP, HPC & OpenPower Operations, Summit Gupta, explained that IBM had decided to collaborate closely with partners such as Nvidia and Mellanox rather than ‘reinventing the wheel’ when it came to developing OpenPower. Gupta had been working on this project while he was at Nvidia and he has recently moved across from Nvidia to IBM.
IBM had already announced that its Power8 CPU would support Nvidia’s NV Link technology, a high speed interconnect which allows up to four GPUs to be connected along with a Power8 processor. Gupta also explained that IBM has also been working closely with Mellanox to develop the networking side of the platform.
Gupta said: ‘Mellanox is going to be putting CAPI into its next generation Infiniband switches, and that will bring the Infiniband library into OpenPower systems. So you will start to see that network integration in OpenPower as well.’
By opening out the framework and inviting experts from other companies to collaborate, IBM can effectively cherry-pick what it sees as the best technology to support its ambitions in HPC. Gupta concluded: ‘We are bringing together all the R&D dollars from all of these companies and it brings many more brains to the table.’
Smaller companies in HPC have to take a more conservative approach, as they cannot hope to match the financial might of companies such as Nvidia, Intel, or IBM. Instead they leverage their own strengths, through IP or expertise picked up in other industries, to develop more specialist platforms that can address specific workloads.
E4 has established two new products, based on ARM and Cavium’s technology, which it calls the ARKA series. Both are 2U solutions but one makes use of Applied Micro’s eight-core X-Gene processor and the other uses Cavium’s Thunder-X 48-core processor, based on 64 bit ARM technology. Both couple these processors with an Nvidia GPU and Infiniband interconnects, although not with the same type of integration afforded to IBM.
These are low energy consumption appliances, specifically designed to appeal to the HPC industry, as demonstrated by the choice to pair them with Nvidia GPUs and up to 128GB of Ram for the X-Gene and up to 256GB of Ram for the Cavium based system.
Cavium is a relatively new addition to HPC, having originally developed technology for the telecoms industry. Its newest product, the Thunder-X is a portfolio of workload-optimised SOCs that integrate ARM processors, hardware accelerators, scalable Ethernet, high I/O bandwidth along with high memory and full level virtualisation – making them suitable for applications that are generally bottlenecked by data rather than flops performance.
Gopal Hegde, Cavium’s VP of data-centre processing, explained that it was Cavium’s rich history in telecoms which led it down this path. Having developed its own IP for networking, the company released quickly that there was a demand for a high memory bandwidth compute platform in the HPC market.
Leading up to ISC15, the company announced several developments to the platform, including a 64 bit ARM toolkit from Allinea and support for Mellanox InfiniBand Adapters, demonstrating that this technology is reacting quickly to the demands of the HPC industry.
Applied Micro also been developing its X-Gene platform to make it not only competitive against other HPC solutions, but to differentiate it from some of the more established Flops-dominant technologies. Making best use of that specific platforms strength rather than trying to chase the industry’s dominant technologies creates a plethora of specialist solutions that are tailored to more specific workloads.
By differentiating their technology platforms, through aiming for more memory intensive applications for example, these smaller companies can overcome the disadvantage that they do not have the sheer financial muscle to match Intel or IBM when it comes to processor development and the huge costs associated with the development and fabrication of ever smaller semiconductors.
Other reports from ISC High Performance have focused on the issues of Why do smaller companies shun HPC? national policy strategies for Easing access to HPC for the SME and on how A portal opens to German HPC centres.
