by

AMD says Nvidia cheats at maths

AMD has taken a swipe at Nvidia’s approach to scientific computing, arguing that using software tricks to reproduce high-precision calculations is a poor substitute for hardware designed to perform them properly.

At the SC26 supercomputing conference in Chicago, AMD promoted its Instinct MI430X accelerator as an alternative for scientific workloads requiring native double-precision floating-point calculations, known as FP64.

According to Tom’s Hardware, AMD boffins have challenged techniques that reproduce FP64 calculations using lower-precision operations, particularly the Ozaki method. This allows processors designed for less precise arithmetic to perform certain double-precision calculations by breaking them into smaller operations.

FP64 is particularly important in scientific computing, where climate simulations, molecular modelling, and engineering calculations often require much greater numerical precision than typical AI workloads. A small rounding error can become a serious problem when a simulation performs billions of calculations.

Nvidia’s recent AI accelerators have concentrated heavily on lower-precision arithmetic, which is useful for training and running artificial intelligence models. AMD reckons the industry has become so obsessed with AI performance that some chip designers are treating traditional scientific computing as an inconvenient afterthought.

AMD’s researchers examined 47 workload configurations across 11 scientific applications and benchmarks to see whether emulated FP64 could provide a practical alternative to native hardware. Their findings suggest the technique works best with large, dense matrix multiplication operations, which make up a relatively small share of scientific computing.

The boffins estimate that fewer than 10 per cent of high-performance computing workloads are sufficiently dominated by these operations to benefit from the approach. Other applications involve sparse matrices, irregular memory access and calculations that cannot easily be converted into the operations favoured by AI accelerators.

Ozaki-style emulation can require up to three times as much memory as native FP64 calculations. For supercomputing operators already paying fortunes for high-bandwidth memory, that could mean buying additional accelerators simply to accommodate the increased memory requirements.

AMD is promoting its Instinct MI430X as a processor that can handle conventional scientific calculations alongside AI workloads. Based on the CDNA 5 architecture, the chip promises up to 288 TFLOPS of native FP64 performance, 432GB of HBM4 memory and memory bandwidth reaching 23.3TB/s.

The company claims the additional memory capacity and native double-precision hardware could reduce the number of GPUs required for certain scientific applications. Its proposed four-GPU configuration uses Infinity Fabric to connect the accelerators directly, providing up to 12.3Tbps of aggregate scale-up bandwidth across the node.

Naturally, AMD has a commercial interest in demonstrating that its architecture is better suited to scientific computing, and its performance figures for the MI430X remain projections. AMD researchers conducted the research, so independent testing will be needed before accepting every claim about its advantages over competing hardware.

AMD has already secured important supercomputing customers for the MI430X, including the Discovery system at Oak Ridge National Laboratory and France’s Alice Recoque exascale project. These machines must support demanding scientific applications where numerical accuracy remains important, regardless of the industry’s enthusiasm for artificial intelligence.

AMD’s research concludes that Ozaki emulation can help with suitable matrix calculations but has substantial limitations across broader scientific workloads. The company’s tests found that only one of its 11 profiled applications had a dominant dense matrix multiplication component that could potentially benefit from the technique.

 

 

TOPICS:
AMD  ·  FP64  ·  gpu architecture  ·  High Performance Computing  ·  instinct mi430x  ·  Nvidia  ·  SC26  ·  scientific computing  ·  supercomputers

Latest articles

Share

Featured articles

Hot topics

No results found.

Latest reviews