By: Michael Feldman

For those of you anxiously anticipating the first exaflop supercomputer, your wait is over – sort of. HPC cloud specialist Rescale is already offering more than 1.4 exaflops of computing power across its global network. The company, which casts its cloud as a “unified HPC simulation platform for the enterprise IT environment,” says its infrastructure currently encompasses 8 million servers spread across 30 datacenters. In aggregate, that works out to over 1400 petaflops of peak computing power, according to the Rescale website.

Assuming the 8 million server number is accurate, their 1.4 exaflop mark certainly seems reasonable. A back-of-the-envelope calculation that assumes one teraflop per five dual-socket servers would easily get you into that territory. That’s assuming straight-up x86 CPUs, although Rescale does offers GPU and Xeon Phi configurations as well.

To put that in perspective, the aggregate computing power of all the supercomputers on the TOP500 list is currently just (!) 418 petaflops. Even Google, which claims 15 data centers, is estimated to have just a million or so servers, although no one knows the exact number (and Google isn’t saying).

Rescale specializes in supporting engineering simulations and modeling, primarily CFD and FEA, although it also has customers who are using the platform for deep neural network work and other more research-oriented applications. It partners with most of the popular ISVs in this area, including the usual suspects like ANSYS, CD-adapco, LS-DYNA, MathWorks, Dassault Systèmes, and MSC Software. It also offers access to a wide array of HPC libraries for visualization, machine learning, molecular modeling, and others.

The actual hardware infrastructure under Rescale is almost certainly the aggregation of one or more third-party cloud providers, although that’s not exactly spelled out on their website. (If the company actually controlled all these servers and they were running Rescale jobs at full capacity, they could be able to accrue well over $200 million per day in revenue.) Access to the hardware is done via Rescale’s SaaS platform, which they call ScaleX.

Given all that, it would be quite an exercise to actually amass an exaflop of computing for a single application under Rescale. But it’s not outside the realm of possibilities that the first exa-application could actually execute in the cloud well before it’s run on one of the first exascale machines expected in 2020 or thereabouts. Could be an interesting research project…