Advanced Micro Devices Inc. today pulled back the curtains on the MI100, a graphics processing unit for supercomputers touted as the fastest chip in its class with more than 10 teraflops of peak performance.

Supercomputers are typically used to run scientific workloads such as weather prediction models and physics simulations. Once, they were built mainly with central processing units.

However, new supercomputers are nowadays often commissioned with large numbers of GPUs as well because researchers are increasingly making use of machine learning in scientific computing. The trend is fueling growing demand for chips such as AMD’s MI100.

The chip, based on a seven-nanometer manufacturing process, provides what AMD describes as record-breaking peak performance of 11.5 teraflops when processing FP64 data. A teraflop equals a trillion calculations per second. FP64 is shorthand for the double-precision floating-point format, a type of data unit that is popular in scientific computing because it lends itself well to representing large numbers.

The MI100 can provide even faster speeds when working with smaller data units. For computations involving FP32 numbers, AMD is promising peak speeds of 46.1 teraflops under certain circumstances.

Part of the credit for the chip’s impressive performance goes to the new Matrix Core Technology included in the package. It’s a specialized compute engine optimized for AI that, according to AMD, can boost the speeds at which neural networks crunch data by a factor of nearly seven depending on the use case.

The basic building of AMD’s GPUs are stream processors, which serve an analogous role to the cores of a central processing unit. The MI100 ships with 7680 such stream processors that are organized into 120 larger chip sections known as computing units. The circuits are supported by 32 gigabytes of memory that the GPU uses to store the data it’s processing for quick access.

A supercomputer is made of multiple servers that can each have multiple GPUs inside depending on performance requirements. To support customers taking this approach, the MI100 packs a technology called Infinity Fabric that can link up to four GPUs together into a so-called hive. MI100 chips set up in such a hive configuration share data with one another faster, and thereby perform calculations faster as well, because Infinity Fabric provides up to twice the peak bandwidth of the PCIe 4.0 technology normally used for interchip connectivity.

AMD is launching the MI100 during an uptick in supercomputing-related spending. Nations around the world are building a new generation of “exascale” supercomputers with performance exceeding a billion billion operations per second, several times more than what today’s fastest systems can manage. AMD has already won deals to supply chips for multiple upcoming exascale systems and, with the MI100, it may be in a stronger position to win future contracts.

AMD’s bold performance claims for the MI100 could mean more competition for Nvidia Corp. in this important market. Only last month, Nvidia announced it has been selected to supply GPUs for four new supercomputers to be built in Europe, one of which is expected to be the world’s most powerful AI supercomputer.

Image: AMD