-
CUTLASS Tutorial: Hardware-supported Block-scaling with Nvidia Blackwell GPUs
Welcome to part 4 of our series investigating GEMM on the NVIDIA Blackwell architecture. So far we have discussed the capabilities of the new Blackwell Tensor Core UMMA instructions, including handling sub-byte data types, and how to work with them in CUTLASS. In this part, we will continue our exploration of low-precision computation by discussing
-
FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling
Modern accelerators like Blackwell GPUs continue the trend of asymmetric hardware scaling, where tensor core throughput grows far faster than other resources such as shared memory bandwidth, special function units (SFUs) for transcendental operations like exponential, and general-purpose integer and floating-point ALUs. From the Hopper H100 to the Blackwell B200, for instance, BF16 tensor core…
-
A User’s Guide to FlexAttention in FlashAttention CuTe DSL
|
Many variants of attention (Vaswani et al., 2017) have become popular in recent years, for reasons related to performance and model quality. These include: The PyTorch team at Meta recognized that most of these variants (including all of the above) can be unified under one elegant framework, dubbed FlexAttention (Guessous et al., 2024). This simple
-
Categorical Foundations for CuTe Layouts
|
In GPU programming, performance depends critically on how data is stored and accessed in memory. While the data we care about is typically multi-dimensional, the GPU’s memory is fundamentally one-dimensional. This means that when we want to load, store, or otherwise manipulate data, we need to map its multi-dimensional logical coordinates to one-dimensional physical coordinates.
-
CUTLASS 3.x APIs: Orthogonal, Reusable, and Composable Abstractions for GEMM Kernel Design (External)
|
In this blog post presented on the NVIDIA technical blog, we give a concise introduction to the CUTLASS 3.x APIs, focusing on the collective, kernel, and device layers and the functionality of the collective builders. This post was authored in conjunction with members of the CUTLASS team.
-
CUTLASS Tutorial: Sub-byte GEMM on NVIDIA® Blackwell GPUs
Welcome to part 3 of our series investigating GEMM on the NVIDIA Blackwell architecture. In parts 1 and 2, we looked at the Tensory Memory and 2 SM capabilities of the new Blackwell Tensor Core UMMA instructions and how to work with them in CUTLASS. In this part, we introduce low-precision computation and then discuss
2805 Bowers Ave, Santa Clara, CA 95051 | 408-730-2275
research@colfax-intl.com