• AI has driven an explosion of new number formats-the ways in which numbers are represented digitally. • Engineers are looking at every possible way to save computation time and energy, including shortening the number of bits used to represent data. • But what works for AI doesn’t necessarily work for scientific computing, be it for computational physics, biology, fluid dynamics, or engineering simulations. • IEEE Spectrum spoke with Laslo Hunhold, who recently joined Barcelona-based Openchip as an AI engineer, about his efforts to develop a bespoke number format for scientific computing. • LASLO HUNHOLD Laslo Hunhold is a senior AI accelerator engineer at Barcelona-based startup Openchip. • He recently completed a Ph.D.

Article Summaries:

  • IEEE Spectrum reports that the surge in AI‑driven number formats-designed to cut bit‑width and save energy-does not translate to scientific computing. Laslo Hunhold, a senior AI accelerator engineer at Openchip, explains that scientific simulations require a wide dynamic range and high precision for both very large and very small values, which 64‑bit floating‑point formats provide but AI formats do not. To address this gap, Hunhold has developed “takum,” a custom format inspired by posits but re‑engineered to allocate more density to the value ranges common in physics, biology, and engineering simulations. The new format aims to improve energy efficiency while preserving the accuracy needed for scientific workloads.
  • AI researchers have introduced a wide array of low‑bit number formats to cut energy use in machine‑learning workloads, but these designs often fail for scientific computing, which demands high dynamic range and precision. Laslo Hunhold, a senior AI accelerator engineer at Barcelona‑based Openchip, argues that the 64‑bit IEEE standard is still too large for many scientific applications, yet the new AI‑centric formats (e.g., 16‑bit, 8‑bit, or even 2‑bit) lack the range needed for physics, biology, and engineering simulations. To address this gap, Hunhold has developed “takum,” a custom format inspired by posits but engineered to provide dense representation across the wide value ranges typical in scientific calculations.
  • IEEE Spectrum reports that the surge in AI‑driven number formats-shorter, energy‑efficient bit representations-does not translate to scientific computing. Laslo Hunhold, a senior AI accelerator engineer at Barcelona‑based Openchip, argues that scientific simulations require a wide dynamic range and high accuracy, unlike AI workloads that tolerate lower precision. Hunhold has developed the “takum” format, inspired by posits but redesigned to allocate more density to the broad range of values common in physics, biology, and engineering simulations. His work aims to make numerical representations both energy‑efficient and suitable for high‑precision scientific calculations.
  • Scientists and engineers are developing new digital number formats to reduce computation time and energy use, but the tricks that work for artificial‑intelligence workloads do not translate to scientific computing. Laslo Hunhold, a senior AI accelerator engineer at Barcelona‑based Openchip, explains that scientific simulations require a wide dynamic range and high precision for both very large and very small values-something the 64‑bit IEEE standard can handle but is inefficient for many AI tasks. Hunhold has created a new format, “takum,” inspired by posits but redesigned to allocate more representation density across the broad range of values typical in physics, biology, and engineering simulations. The goal is to achieve up to 10 % energy savings while maintaining accuracy for scientific workloads.

Sources: