analiza aiqbench
This article proposes a theoretical framework where FP8 tensor-core operations, augmented by specific reconstruction schemes, can match native FP64 accuracy on next-generation NVIDIA GPUs. The authors argue that high-precision floating-point hardware is no longer a strict requirement but rather an emergent property of optimized lower-precision workflows and data movement strategies. This research directly addresses the architectural evolution of AI accelerators and the feasibility of replacing legacy precision standards with more efficient formats for large-scale matrix computations.
FP8 is All You Need (Part 1): Debunking Hardware FP64 as the HPC Holy Grail (Sep 3rd version)
arXiv:2606.06510v4 Announce Type: replace-cross Abstract: We argue that on AI-optimised GPUs of the NVIDIA B300 generation and beyond, the FP8 tensor-core matrix operation, composed through CRT-based Ozaki Scheme II, can serve as the dominant matrix-work substrate for the surveyed matrix-dominated FP64 kernel classes at FP64-grade accuracy, with native FP64 recast from a hardware requirement into a derived accuracy guarantee. The claim is conditional: the FP8 op is the candidate dominant multiplication substrate, with a bounded auxiliary set of integer deconstruction/reconstruction work, FP32/Kulisch reductions, data movement and a native-FP64 fallback, organised as a hierarchy from the FP8 op through Ozaki II and the Berkeley dwarfs to applications. The instrument is the Tensor-Memory Equilibrium (TME) model, a Roofline extension with four parameters (compute multiplier $\alpha=3r+1$, bandwidth multiplier $\beta$, reconstruction cost $\gamma$, and the per-input deconstruction cost $c_q$ identified in an NVIDIA review) under which, at its upper bound, the reduction to FP8 costs no performance against an ideal native-FP64 machine of equal bandwidth. On-chip tile fusion drives $\beta \to 1$; the deconstruction term sets a threshold intensity below which emulation is conversion-bound. At the fused, engineered-$c_q$ bound every surveyed class reaches the memory roof, with two priced exceptions: large dense-square DGEMM sits at a deconstruction floor near 0.50 of the FP8 arithmetic roof (about 235 of 473 TFLOPS on the NVIDIA Rubin GPU), a liftable co-design coordinate, and the 3-D FFT is walled by a per-output integer epilogue at $4.9$-$6.7\times$ its roof in software, recoverable with minor hardware and one moderate ask. Ozaki II lifts the emulated FP64 ceiling from $\approx 1.3$ to $\approx 135$ TFLOPS on B300 and $\approx 473$ on Rubin; three deconstruction-path hardware options are given; constants are engine-checked.
Abstract:We argue that on AI-optimised GPUs of the NVIDIA B300 generation and beyond, the FP8 tensor-core matrix operation, composed through CRT-based Ozaki Scheme II, can serve as the dominant matrix-work substrate for the surveyed matrix-dominated FP64 kernel classes at FP64-grade accuracy, with native FP64 recast from a hardware requirement into a derived accuracy guarantee. The claim is conditional: the FP8 op is the candidate dominant multiplication substrate, with a bounded auxiliary set of integer deconstruction/reconstruction work, FP32/Kulisch reductions, data movement and a native-FP64 fallback, organised as a hierarchy from the FP8 op through Ozaki II and the Berkeley dwarfs to applications. The instrument is the Tensor-Memory Equilibrium (TME) model, a Roofline extension with four parameters (compute multiplier $\alpha=3r+1$, bandwidth multiplier $\beta$, reconstruction cost $\gamma$, and the per-input deconstruction cost $c_q$ identified in an NVIDIA review) under which, at its upper bound, the reduction to FP8 costs no performance against an ideal native-FP64 machine of equal bandwidth. On-chip tile fusion drives $\beta \to 1$; the deconstruction term sets a threshold intensity below which emulation is conversion-bound. At the fused, engineered-$c_q$ bound every surveyed class reaches the memory roof, with two priced exceptions: large dense-square DGEMM sits at a deconstruction floor near 0.50 of the FP8 arithmetic roof (about 235 of 473 TFLOPS on the NVIDIA Rubin GPU), a liftable co-design coordinate, and the 3-D FFT is walled by a per-output integer epilogue at $4.9$-$6.7\times$ its roof in software, recoverable with minor hardware and one moderate ask. Ozaki II lifts the emulated FP64 ceiling from $\approx 1.3$ to $\approx 135$ TFLOPS on B300 and $\approx 473$ on Rubin; three deconstruction-path hardware options are given; constants are engine-checked.
This is the 37th revised version (Sep 3). We have made corrections to the TME performance model to account for the deconstruction cost, as well as propose hardware extensions to mostly eliminate the cost
Hardware Architecture (cs.AR); Artificial Intelligence (cs.AI); Distributed, Parallel, and Cluster Computing (cs.DC); Performance (cs.PF)
arXiv:2606.06510 [cs.AR]
(or arXiv:2606.06510v4 [cs.AR] for this version)
Text integral sub licența deschisă a publicației · 331 cuvinte
Subiecte conexe
FP8 is All You Need (Part 1): Debunking Hardware FP64 as the HPC Holy Grail
arXiv · 8 iun. 2026
FP8 is All You Need (Part 1): Debunking Hardware FP64 as the HPC Holy Grail (June 13th version)
arXiv · 16 iun. 2026
Performance Profiling on AMD GPUs – Part 5: Profiling-Driven Kernel Optimization with an AI Code-Assist Tool
AMD · 16 iul. 2026
Quantum computers outperform classical ones , with results you can trust
arstechnica.com · 30 iul. 2026