Supercomputing in previous gen was mainly covered by IBM and Intel for production, distribution, R&D, and corporate client solutions etc. But recently the ball has fallen into a separate court. AMD at SIGGRAPH presented a supercomputer that’s able to process a PetaFLOPs in a single cycle, Project 47.
Project 47 is a server rack that is capable of processing a whopping 1 Peta floating point operations per second. It incorporates 10 TBs of Samsung 2.66GHz memory, 20 AMD EPYC 7601 CPUs, 80 Radeon Instinct GPUs and 20 MELLANOX 100G IB cards with 1 switch. All of this fit in a single rack that’s roughly the height of 6 feet. Available from Inventec and its main distributor AMAX later in Quarter four of 2017.
Ten years ago, there was a supercomputer which offered the same processing power built by AMD and IBM. It had a spanning area of 6000 sq. feet consuming 2.35 Megawatts of electricity. The cluster consisted of 64000 dual core Opteron CPUs & some accelerators on 696 racks.
It was called IBM’s Roadrunner.
For the sake of comparison, looking at the technological advancement now we realize that over a span of 10 years AMD has managed to build a system that consumes 98% less power without taking 99.9% of the space. It will probably cost much lesser than the Roadrunner that was sold for $100,000,000 (one hundred million us dollars).
With 20 EPYC 7601 CPUs, they’ll give a core count of 640 cores which are nothing compared to the 128,000 cores of the original Roadrunner. It will have 80 Radeon Instincts inside the rack that makes roughly 1000 TFLOPs as per P47s rating. With 128 PCIe lanes, EPYC processors will basically act on support drivers for Radeon Instinct GPUs, which shows that P47 is an all-GPU based solution.
The scaling possibilities of core count and node shrinks have developed a new era of strategic effective investments in the corporate sector that require supercomputing to handle enormous data streams.