GPU memory vs FP64 gap
- Analysis noted NVIDIA's AI focus deprioritizes FP64, leaving HPC and large-scene rendering markets under-served. - Newcomer Bolt unveiled a test GPU chip with a 384GB memory card aimed at 'scenes don't fit' limitations on RTX-class cards. - For aerospace simulation and high-fidelity modelling, larger memory and FP64 capability remain important system-level requirements. (x.com)
Bolt Graphics says it has taped out a test Zeus GPU that can be configured with up to 384 GB of memory to fix “scenes don’t fit.” (prnewswire.com 1) (prnewswire.com 2) Double‑precision (FP64) is a 64‑bit floating‑point format used for high‑fidelity science and engineering solvers; many CFD and structural codes recommend double precision for convergence and stability. (ansyshelp.ansys.com) GPU video memory (VRAM) holds geometry, textures and simulation state; Bolt’s prototype advertises a mixed LPDDR5X+DDR5 design that reaches 384 GB of combined GPU memory. (techspot.com 1) (techspot.com 2) Bolt announced the tape‑out on April 22, 2026, says developer kits are due in 2026 and mass production is targeted for 2027, and the company lists up to ~20 FP64 teraFLOPS in its marketing. (prnewswire.com) (prnewswire.com) NVIDIA’s recent product mix has emphasized low‑precision AI throughput over native FP64 hardware, and the company has leaned on software emulation to boost FP64 matrix rates—reports say emulation can reach much higher FP64 TFLOPS than native modes. (theregister.com) That trade‑off matters for aerospace simulation and high‑fidelity modelling because large meshes and stiff numerical solvers both demand ample VRAM and sustained FP64 throughput to avoid divergence or accuracy loss. (compute.hivenet.com) AMD has positioned its Instinct family around stronger FP64 and HBM memory for HPC use cases, and vendors point to those cards as alternatives where native double‑precision matters. (amd.com) Bolt’s hardware design replaces fixed GDDR layouts with modular LPDDR5X and DDR5 SO‑DIMM slots to scale VRAM to 384 GB, but analysts and Jon Peddie note most performance numbers remain pre‑silicon and require independent benchmarking. (en.gamegpu.com) NVIDIA tells HPCWire it plans to add more FP64 capability in future generations, while some researchers and rival engineers question whether emulation matches hardware for production aerospace simulations. (techpowerup.com) Bolt’s April 22, 2026 tape‑out moves a memory‑focused GPU concept into silicon; independent developer‑kit benchmarks in 2026 — and how emulated vs. native FP64 performs in real CFD and structural runs — will determine if Zeus meets aerospace and HPC system requirements. (prnewswire.com) (prnewswire.com)
