독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
뉴스주제 · 英伟达cuBLAS新算法중대2 분 소요DisconfirmAI 편집부

NVIDIA Releases New cuBLAS Algorithms to Accelerate Double-Precision Scientific Computing

NVIDIA's CUDA Toolkit 13.4.2 introduces Ozaki-II and 2M algorithms, leveraging integer Tensor Cores to accelerate FP64 matrix operations with significant gains in quantum chemistry and materials science.

2026-11-14 06:13 GMT+8

NVIDIA has introduced two new algorithms, Ozaki-II and 2M, in the latest release of its CUDA Toolkit 13.4.2 for the cuBLAS math library. These technologies aim to resolve a long-standing contradiction in high-performance computing: how to leverage the faster integer Tensor Cores—designed primarily for low-precision tasks—while maintaining the full accuracy of double-precision floating-point (FP64) calculations. This update allows scientific simulations dependent on high-precision dense general matrix multiplications (GEMM), such as those in molecular chemistry and materials science, to automatically gain speed without requiring any changes to application code.

Traditionally, GPU FP64 units offer high precision but limited throughput, while Tensor Cores provide extreme speed for lower-precision data. The new cuBLAS methods employ "floating-point emulation" techniques, decomposing each double-precision number into multiple integer segments, performing exact integer multiplications on the Tensor Cores, and reassembling the result to meet strict FP64 standards. Specifically, the Ozaki-II algorithm utilizes the Chinese Remainder Theorem (CRT) to significantly reduce the number of required integer multiplications compared to its predecessor, Ozaki-I. Meanwhile, the 2M algorithm optimizes complex matrix multiplication by reducing the operation from four real multiplications to two, thereby lowering computational overhead.

According to NVIDIA's official technical blog, these algorithms demonstrate substantial performance improvements in real-world applications. In quantum chemistry, DGEMM (real double-precision matrix multiplication) using Ozaki-II runs 5.5x faster than native FP64 on NVIDIA VR NVL72 systems and 3.9x faster on GB200. For Coupled Cluster Singles and Doubles (CCSD) calculations—considered the "gold standard" in quantum chemistry—the end-to-end speedup reaches 2x on VR NVL72 and 1.4x on GB200. When employing mixed-precision modes, these gains extend to 5x and 3.5x respectively.

로그인 / 회원가입
독서리서치레이더투자 프레임워크

In materials science, ZGEMM (complex double-precision matrix multiplication) within the BerkeleyGW software package achieves a 5x speedup on GB200 when using the 2M algorithm. Tests on the BerkeleyGW Epsilon module show an end-to-end speedup of 6.6x on VR NVL72 and 3.9x on GB200. Notably, all tests verified that results remain within FP64 accuracy bounds, and users do not need to manually adjust APIs or rewrite kernels, as cuBLAS automatically selects the optimal path.

Furthermore, the upcoming NVIDIA Rubin GPU architecture introduces the new TI16 numeric format and the tcgen05.mma instruction, which will further accelerate Ozaki-series computations. Existing users can enable these features simply by downloading and installing the latest CUDA Toolkit. Developers can refer to the floating-point emulation section of the cuBLAS documentation for detailed API control options, such as cublasSetEmulationStrategy(), to fine-tune the balance between precision and performance for specific workloads.

출처

  • developer.nvidia.comNVIDIA Technical Blog↗
주제

英伟达cuBLAS新算法

英伟达在 CUDA Toolkit 13.4.2 中为 cuBLAS 库引入了 Ozaki-II 和 2M 算法,旨在通过整数 Tensor Core 加速双精度矩阵乘法(DGEMM/ZGEMM),同时保持全 FP64 精度。该方案将浮点数…

3 단계1 건업데이트됨 2026-11-18 13:29

현재까지의 과정

  1. 2025-10-24英伟达发布技术博客,介绍 CUDA Toolkit 13.0 Update 2 中 cuBLAS 的浮点仿真功能,包括基于 Blackwell INT8 Tensor Core 的 FP64 仿真(Ozaki Scheme)。developer.nvidia.com ↗
  2. 2026-11-13英伟达宣布 CUDA Toolkit 13.4.2 引入 Ozaki-II 和 2M 算法,声称在保持全精度下显著加速科学计算中的双精度矩阵乘法。developer.nvidia.com ↗
  3. 2026-11-16TechTimes 报道指出,虽然英伟达宣称最高 5.5 倍加速,但该增益仅适用于矩阵乘法,向量运算无提升,且所有数据均来自英伟达内部测试。techtimes.com ↗

아직 해결되지 않은 문제

  • 独立实验室或第三方 HPC 中心(如 TACC)能否复现英伟达报告的端到端加速比?
  • Ozaki-II 算法在输入数据指数分布较宽时的实际精度损失是否在可接受范围内?
  • 相比原生 FP64 实现,仿真算法带来的额外内存占用(约两倍)是否会限制大规模并行计算的扩展性?
전체 주제 읽기 →
리서치 계속하기 →