#RocBLAS
Selecting a BLAS Backend: cuBLAS vs rocBLAS vs Vendor BLAS
beefed.ai
October 9, 2026 at 3:55 AM
I don't think it would take a whole lot to get same key batching working with rocBLAS
September 10, 2026 at 10:43 AM
The evaluation results on NVIDIA A100, H800, and Hygon DCU Z100 show that HSMA-TRSM achieves peak speedups of 2.05× over cuBLAS and 2.06× over rocBLAS.
August 29, 2026 at 12:10 AM
And on GPUs, you have CUTLASS / cuBLAS/ cuDNN, rocBLAS, MIOpen, FlashAttention, Triton has its own abstracted assembly, etc etc.
June 14, 2026 at 12:56 AM
Seb wrote a fantastic article exploring how to make matrix multiplications go fast on AMD GPUs. This is a deep dive into both the algorithm and hardware, with great visualizations. He is also looking for work and I only have good things to say about him, so hire him!
seb-v.github.io/optimization...
Optimizing Matrix Multiplication on RDNA3: 50 TFlops and 60% Faster Than rocBLAS
Introduction
seb-v.github.io
February 13, 2025 at 5:16 PM
libraries such as cuBLAS and rocBLAS, while providing these routines on Apple Silicon for the first time. The entire implementation is only a few hundred lines of code, showing that unified Julia programs can deliver near-vendor performance across [4/5 of https://arxiv.org/abs/2504.13821v1]
April 21, 2025 at 5:58 AM
Optimizing Matrix Multiplication on RDNA3 Discussion
Optimizing Matrix Multiplication on RDNA3: 50 TFlops and 60% Faster Than rocBLAS
Introduction
seb-v.github.io
March 28, 2025 at 10:40 PM
AMD MI300x GPUs with GEMM tuning improves throughput and latency by up to 7.2x Discussion
Nscale Benchmarks: AMD MI300x GPUs with GEMM tuning improves throughput and latency by up to 7.2x
Optimising AI model performance: vLLM throughput and latency benchmarks and GEMM Tuning with rocBLAS and hipBLASlt
www.nscale.com
June 29, 2024 at 9:00 AM
The larger packages PyTorch requires:

clang18 and llvm18 - 700+ MiB (what are you going to compile?)
magma - 2.3 GiB
miopen - 1.4 GiB
rocblas - 10.0 GiB (!!!)
rocrand - 422.0 MiB
rocsolver - 2.3 GiB
rocsparse - 5.2 GiB
suitesparse - 139.9 MiB
The entire TeXLive (will not count).
And of course […]
Original post on mastodon.online
mastodon.online
March 9, 2025 at 5:20 PM
"The implementation matches the performance of state-of-the-art implementations (cuBLAS/rocBLAS) and for the first makes linear algebra implementations available for Apple Silicon."
April 21, 2025 at 6:32 AM
rocblas-gfx1103
Next generation BLAS implementation for ROCm platform
aur.archlinux.org
June 14, 2025 at 11:38 AM
rocblas 7.1.0-2 x86_64 Next generation BLAS implementation for ROCm platform

#Extra-Testing #x86_64

Origin | Interest | Match
Arch Linux - rocblas 7.1.0-2 (x86_64)
archlinux.org
November 24, 2025 at 12:38 PM
rocblas 6.4.3-3 x86_64 Next generation BLAS implementation for ROCm platform

#Extra #x86_64

Origin | Interest | Match
Arch Linux - rocblas 6.4.3-3 (x86_64)
archlinux.org
August 29, 2025 at 3:29 PM
rocblas 6.4.2-1 x86_64 Next generation BLAS implementation for ROCm platform

#Extra-Testing #x86_64

Origin | Interest | Match
Arch Linux - rocblas 6.4.2-1 (x86_64)
archlinux.org
August 3, 2025 at 9:54 AM