New Free Software Makes Supercomputers Faster Using Cheaper Memory

Modern supercomputers contain two kinds of memory. One is HBM — super-fast memory stacked right next to the processor, but expensive and available in small amounts. The other is ordinary DRAM — cheap, plentiful, but slower. Today, deciding what goes where is mostly guesswork: programmers tune it by hand, or the operating system uses rules written long before these two-tier machines existed. Get it wrong and your expensive fast memory sits half-empty while the machine crawls.

A team led by Marios Asiminakis and Manolis Marazakis at Greece's FORTH research institute built SHAMBLES, a piece of software baked into the operating system that watches which chunks of data a program actually touches, then quietly shuffles the busy data into the fast tier and the idle data out. Crucially, the program itself never has to be changed. Think of a chef who keeps the ingredients they're actively using on the counter and the rest in the pantry — and re-stocks the counter automatically as the recipe changes.

The results are striking. On a demanding physics benchmark called HPCG, keeping just 40% of the problem in fast memory preserved 93.75% of the speed you'd get if everything fit there. On DGEMM, a heavy math test, one third of the data in fast memory delivered up to 99% of full speed. A third test, Himeno, showed there's no magic number: the right split depends on both the workload and its size, so the software has to adapt on the fly.

The catch: this is a research prototype, not something you can download and install at home. It needs hardware that already has both memory types, which is mostly big servers and supercomputers. And it works best on scientific simulations, not everyday apps. Still, the direction is clear — squeezing more speed out of hardware people already own.