In the past decade supercomputers were dressed-up versions of Intel’s x86 machines, but increasingly supercomputers are borrowing innovations (and silicon in the form of ARM-based chips or DSPs) from the mobile and big data realms to add speed without guzzling too much power.
Prior to this century many supercomputers really were a different animal entirely, sporting specialty chips and software. But the industry turned to commodity chips in the early 2000s. Now, to meet the demands of exascale computing at low power, chip makers are taking inspiration from the cloud computing and mobile industries.
ARM tries supercomputing on for size
As the Supercomputing 2011 show gets under way in Seattle, Nvidia, Texas Instruments, ARM and others are announcing new silicon to power the machines we rely on for science, climate prediction and high-end simulations in industries that range from oil production to car design.
Nvidia is a fairly recent newcomer to the supercomputing market, but it has made huge strides since 2008, when it first starting pushing its graphics processors (GPUs) as a way to boost speed while keeping energy usage in check. It said it would use its high-end GPUs and its new GPU-plus-ARM chip to build a new supercomputer in Spain. This is the first time an ARM-based processor has made its way into a supercomputer. ARM thus far has been the chip of choice inside cell phones and tablets.
Accelerator chips advance in supercomputers
Nvidia is doing well with its GPUs, given that in the top 500 ranking of the world’s fastest supercomputers, 39 systems use GPUs as accelerators and 35 of these use Nvidia chips. The graphics processors are used in supercomputers because they can handle massively parallel tasks that high-end computing requires while using less energy than the typical CPUs made by Intel and AMD. Nvidia and its GPUs made their first appearance on the list in 2008, and the last time the top 500 list was published, six months ago, Nvidia chips were in 17 machines. To go to 35 today is a pretty big uptake.
Perhaps inspired by Nvidia’s success in getting its GPUs onto supercomputers, Texas Instruments is bringing its digital signal processors to the mix for high-performance computing. DSP chips are really good at math, and they are used in telecommunications chips and in routers. TI has been thinking about this for a while, but Monday was its first launch into the market formally.
New chips for the cloud
The same power-efficiency issues that plague those trying to advance supercomputing are hitting those who run webscale applications, from Facebook to Amazon Web Services. And while the cloud and web-scale data center operators aren’t looking for specialty gear, like Infiniband for networking, they are running one or a few applications on their hardware, similar in some ways to a supercomputer, where all workloads are optimized for speed.
This is why certain chip and hardware companies, such as Tilera, Calxeda and Applied Micro, see an opportunity to redesign the silicon and gear inside the cloud. Meanwhile, companies such as Adapteva, which makes a massively multicore chip for cell phones and HPC, see an opportunity in pushing into supercomputers and mobile handsets, where the need for more-powerful processors and lower power consumption are always at war. And with ARM piggybacking on this trend thanks to Nvidia, it’s clear that supercomputers want to be super without the influence of PCs.