Enable CUDA 13.0 binaries

[tinglvv]

🚀 The feature, motivation and pitch

CUDA 13.0 is released on 8/4, creating issue tracker for CUDA 13.0 binaries enablement.

CUDA 13.0 is a major upgrade over CUDA 12, benefits from upgrading in the nightlies binaries are mainly:

1. CUDA 13.0 supports all NVIDIA architectures from Turing through Blackwell. Featuring sm_110 (Jetson Thor), sm_103 (B300/GB300), sm_121(DGX SPARK/DIGITS). Note: CUDA 12.9 first introduced sm_12x and sm_10x arch support, which will be compatible with sm_120 and sm_100 binaries respectively. Thor and SPARK now both on SBSA platforms.
2. Binary size reduction from compression method change: Supporting --compress-mode flag for nvcc across all drivers of CUDA 13.X toolkits, enabling the possibility to use --compress-mode=size for significant size reduction (~71% less for CUDA Math APIs for example). Default will be --compress-mode=default which gives smaller binaries than LZ4 speed mode used in previous CUDA versions.
3. New features to Math libraries: cuBLAS improvements, including implicit autotuning mode
4. Arm platforms are now unified in CUDA Toolkit, enabling single-install and consistent builds across all Arm architectures.
5. General CUDA code compilation performance https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#id2

See https://developer.nvidia.com/blog/whats-new-and-important-in-cuda-toolkit-13-0/ for more comprehensive notes.

Update Plan

Feature	Target Date
Linux and Linux aarch64 CD Nightly	Aug 22, 2025
Enablement of CUDA 13.0 in PyTorch CI	Aug 29, 2025
Windows CD Nightly	Aug 29, 2025
Enable CUDA 13.0 in Ecosystem Libraries	Aug 29, 2025
Removal of CUDA 12.9 Builds	Aug 29, 2025

PRs

Linux builds CD nightly (x86, sbsa, libtorch)

1. Almalinux docker --> Magma build

• Add almalinux docker for magma and CUDA installation for CUDA 13 #160201
• Add Magma build for CUDA 13.0 #160770

2. Build addition

• x86: [CD] Add CUDA 13.0 x86 nightly builds #160956
• sbsa: [CD] [aarch64] Add CUDA 13.0 sbsa nightly build #161257
• torchvision/torchaudio x86: Add x86 domain builds for CUDA 13.0 test-infra#7046
• torchvision/torchaudio sbsa: Add sbsa CUDA 13.0 domain builds test-infra#7059
• libtorch build: [CD] Add cuda 13.0 libtorch builds, remove CUDA 12.9 builds #161916

3. Build size reduction

• Use -compress-mode=size for CUDA 13 build for binary size reduction #161316

4. 13.0.U2 upgrade

• [CD] Upgrade to CUDA 13.0.2 for nightly binaries #165470

CUDA 13 CI

• [CI] Add basic CUDA 13.0 periodic test #161013

Windows CD nightly

1. Add Windows AMI

• Add Windows CUDA 13 AMI test-infra#7003
• Update CUDA 13 CUDNN to 9.12.0.46 test-infra#7020
• Fix CUDA 13.0 Windows AMI building test-infra#7035

2. Add Windows build script

• Add Windows CUDA 13 build and magma script #161073
• [CD] Add CUDA 13.0 Windows build #161663

Alternatives

No response

Additional context

No response

cc @seemethere @malfet @atalman @ptrblck @msaroufim @eqy @jerryzh168 @nWEIdia

[malfet]

It would be nice to have a short description why updating to CUDA-13 is helpful. I.e. it might be that it improves perf, adds new features, new HW capabilities etc

[seemethere]

It would be nice to have a short description why updating to CUDA-13 is helpful. I.e. it might be that it improves perf, adds new features, new HW capabilities etc

I'd agree. Currently as it sits we have 3 very expensive binary build workflows for CUDA binaries that seems to increase in both memory requirements as well as build time. If we add a 4th to this including new hardware architectures (which probably means even longer build times) I feel like we're approaching crazy levels of requirements for producing binaries.

I'd suggest if we add CUDA 13 we should remove at least one of the CUDA 12 builds and maybe even one more (since CUDA 12.8 and 12.9 seem very similar).

[tinglvv]

Hi @malfet @seemethere , thanks for the suggestion. I have updated the description with CUDA 13.0 benefits.
Agree that we should keep one stable CUDA 12 builds, 12.8 would be good, since 13.0 will cover the added functionalities for 12.9, and there's the build issue of Windows OOM with CUDA 12.9 as well. #156181

[malfet]

Supporting new HW capabilities: sm_103 (B300/GB300), sm_110 (Jetson Thor), sm_120 (Geforce 5000 series), sm_121(DGX SPARK). Notes: sm_121 will be compatible with sm_120 binaries, sm_10x will be compatible with sm_100 binaries.

@tinglvv Aren't sm_120 already supported by CUDA-12.8 and rest is already supported by CUDA-12.9, isn't it? (Or can you clarify what HW were added as part of CUDA-12.9?)

[tinglvv]

Supporting new HW capabilities: sm_103 (B300/GB300), sm_110 (Jetson Thor), sm_120 (Geforce 5000 series), sm_121(DGX SPARK). Notes: sm_121 will be compatible with sm_120 binaries, sm_10x will be compatible with sm_100 binaries.

@tinglvv Aren't sm_120 already supported by CUDA-12.8 and rest is already supported by CUDA-12.9, isn't it? (Or can you clarify what HW were added as part of CUDA-12.9?)

Thanks for catching it, yes sm_120 is already supported in CUDA 12.8. Modified the description.

[nWEIdia]

https://developer.nvidia.com/blog/whats-new-and-important-in-cuda-toolkit-13-0/#wheel_package_changes_to_cuda_130
it is also time to evaluate this.