From 5f2072261a568acead4cfb8c626b867d2263d68f Mon Sep 17 00:00:00 2001 From: "C. Regouby" Date: Fri, 19 Jun 2026 19:58:25 +0200 Subject: [PATCH 1/2] add HW cmpatibility matrix --- vignettes/articles/compatibility-matrix.Rmd | 39 ++++++++++++++++----- 1 file changed, 31 insertions(+), 8 deletions(-) diff --git a/vignettes/articles/compatibility-matrix.Rmd b/vignettes/articles/compatibility-matrix.Rmd index 3fbc33a803..02b01ec3ba 100644 --- a/vignettes/articles/compatibility-matrix.Rmd +++ b/vignettes/articles/compatibility-matrix.Rmd @@ -8,6 +8,7 @@ Finding the right combination of libtorch, {torch}, and CUDA can be tricky. This | {torch} | libtorch | Suitable CUDA versions | |:-------:|:--------:|:-----------------------:| +| 0.17.x | 2.8.0 | 12.6, 12.8 | | 0.16.x | 2.7.1 | 12.6, 12.8 | | 0.15.x | 2.5.1 | 11.8, 12.4 | | 0.14.x | 2.5.1 | 11.8, 12.4 | @@ -20,11 +21,33 @@ Finding the right combination of libtorch, {torch}, and CUDA can be tricky. This ## CUDA version compatibility -| CUDA version | Suitable {torch} version | -|:------------:|:---------------------------:| -| 11.6 | 0.10.0 (Linux only), 0.11.0 | -| 11.7 | 0.10.0 to 0.13.0 | -| 11.8 | 0.12.0 to 0.15.x | -| 12.4 | 0.14.x to 0.15.x | -| 12.6 | 0.16.x | -| 12.8 | 0.16.x | +The following table details the suitable `{torch}` versions for each CUDA release, along with the supported GPU architectures and their corresponding compute capability ranges. + +| CUDA version | Suitable {torch} version | Supported architectures | Compute capability | +|:------------:|:------------------------:|:-----------------------:|:------------------:| +| 11.6 | 0.10.0 (Linux only), 0.11.0 | Maxwell, Pascal, Volta, Turing, Ampere | 5.0 - 8.6 | +| 11.7 | 0.10.0 to 0.13.0 | Maxwell, Pascal, Volta, Turing, Ampere | 5.0 - 8.6 | +| 11.8 | 0.12.0 to 0.15.x | Maxwell, Pascal, Volta, Turing, Ampere | 5.0 - 8.6 | +| 12.4 | 0.14.x to 0.15.x | Maxwell, Pascal, Volta, Turing, Ampere, Hopper | 5.0 - 9.0 | +| 12.6 | 0.16.x to 0.17.x | Maxwell, Pascal, Volta, Turing, Ampere, Hopper | 5.0 - 9.0 | +| 12.8 | 0.16.x to 0.17.x | Turing, Ampere, Hopper, Blackwell [^1] | 7.5 - 12.0 | + +: CUDA compatibility matrix with architectures and compute capabilities + +## GPU Architectures and Compute Capabilities + +Here is a mapping overview of the major GPU architectures to their typical models and exact compute capabilities. + +| Architecture | Example GPU Models | Compute Capability | +|:------------:|:------------------:|:------------------:| +| Maxwell | GTX 900 series, Tesla M10/M40 | 5.0 - 5.3 | +| Pascal | GTX 1000 series, Tesla P100/P40 | 6.0 - 6.2 | +| Volta | Titan V, Tesla V100 | 7.0 - 7.2 | +| Turing | RTX 2000 series, Tesla T4 | 7.5 | +| Ampere | RTX 3000 series, A100, A10 | 8.0 - 8.7 | +| Hopper | H100, H200 | 9.0 | +| Blackwell | RTX 5000 series, B100, B200 | 10.0 / 12.0 | + +: GPU architectures reference + +[^1] Previous architecture deprecated according to [pytorch issue 172351] (https://github.com/pytorch/pytorch/issues/172351) \ No newline at end of file From f2d1d8e23c782352aa5e8671e106217f011bd990 Mon Sep 17 00:00:00 2001 From: "C. Regouby" Date: Fri, 19 Jun 2026 20:05:15 +0200 Subject: [PATCH 2/2] fix typo --- vignettes/articles/compatibility-matrix.Rmd | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vignettes/articles/compatibility-matrix.Rmd b/vignettes/articles/compatibility-matrix.Rmd index 40db4b3fad..b89bc676f2 100644 --- a/vignettes/articles/compatibility-matrix.Rmd +++ b/vignettes/articles/compatibility-matrix.Rmd @@ -50,4 +50,4 @@ Here is a mapping overview of the major GPU architectures to their typical model : GPU architectures reference -[^1] Previous architecture deprecated according to [pytorch issue 172351] (https://github.com/pytorch/pytorch/issues/172351) +[^1] Previous architecture deprecated according to [pytorch issue 172351](https://github.com/pytorch/pytorch/issues/172351)