[Roadmap]: cuDNN Roadmap Q3 2026 (July – September)

[Anerudhan]

[Roadmap]: cuDNN Roadmap Q3 2026 (July – September)

This issue tracks planned cuDNN development and releases for Q3 2026 (July – September).
Plans are subject to change as the team iterates and receives feedback.
If you have suggestions for features, please open a feature request or comment below.

This roadmap covers both the cuDNN backend library (9.x) and the open-source cuDNN Frontend (1.x). Frontend release notes: https://github.com/NVIDIA/cudnn-frontend/releases. Backend release notes: https://docs.nvidia.com/deeplearning/cudnn/backend/latest/release-notes.html.

---

Recently Released (June – July 2026)

cuDNN Backend — 9.23.0 & 9.24.0

9.24.0 — Highlights

• Expanded causal 1-D convolution support: Causal conv1d is expanded through the SubquadraticOps API with NVRTC runtime compilation, offering multiple depthwise variants.
• Ubuntu 26.04 support: cuDNN 9.24.0 adds support for Ubuntu 26.04.
• CUDA Toolkit discovery via environment variables: CUDA_HOME, CUDA_PATH, and CUDA_ROOT can now be used to locate CUDA Toolkit components at runtime.
• Stability fixes: Convolution forward plans with very large filter tensors are now supported (previously returned CUDNN_STATUS_NOT_SUPPORTED); fixed incorrect depthwise convolution results when multiple operations with different spatial dimensions executed sequentially; fixed GEMM fusions (ENGINE_GLOBAL_INDEX=11) producing incorrect results with unpacked tensors.

9.23.0 — Highlights

• CUDA Toolkit 13.3 support: cuDNN 9.23.0 adds support for NVIDIA CUDA Toolkit 13.3.
• SDPA backward d=256 on Blackwell: Scaled dot-product attention backward now supports embedding dimension d=256 with bfloat16/float16 using the 2-CTA MMA path on Blackwell-architecture GPUs.
• Stability fixes: Large-filter convolution forward plans now return CUDNN_STATUS_NOT_SUPPORTED before launch instead of failing/illegal-memory-access; fixed false cache hits in backward normalization passes producing incorrect gradients; fixed RMSNorm/LayerNorm cache-hit issues with large row counts causing CUDA errors.

Supported architectures (this period): Blackwell (B200, B300; compute capability 10.0/10.3/12.0), Hopper (H200/H20; 9.0), Ada Lovelace (L40S, RTX PRO 6000; 8.9), Ampere (A100; 8.0).

For full details, see the cuDNN Backend Release Notes.

cuDNN Frontend — v1.25.0 (10 Jun) & v1.26.0 (07 Jul)

v1.25.0 is the recommended frontend for cuDNN 9.23.0+; v1.26.0 is the recommended frontend for cuDNN 9.24.0+.

• SDPA backward d=256 native path: SDPA backward with head dimension d=256 now uses the cuDNN-native path on cuDNN 9.23+, bypassing the OSS kernel path.
• Structured plan pinning: get_engine_and_knobs_at_index enables exact plan replay via (engine_id, {knob_values}); new knob types SWAP_AB, INPUT_TMA_ENABLE, and OUTPUT_TMA_ENABLE.
• New BYTE_BOOLEAN data type: Boolean tensors automatically map to BYTE_BOOLEAN when running against cuDNN 9.25 and later.
• Block-sparse attention: New block-sparse attention CuTe-DSL kernels for Hopper and Blackwell.
• DSA enhancements: q causal offsets and SM100F support, an SM90 indexer-forward kernel, and an optimized SM100 backward kernel.
• Grouped GEMM: row-scale quantization support, optional caller-provided output tensor, SReLU in the Hadamard fusion, group_offset support on the reduction node for MoE workloads (cuDNN 9.24+), and CuTe DSL 4.5 compatibility.
• General improvements: static linking of libcudnn; improved libcudart loading with an environment-variable override; per-tensor ragged offset multiplier support (cuDNN 9.24+).
• Bug fixes: grid-dimension overflow in the DSA backward convert kernel on SM100; illegal memory access in indexer_topk_wrapper; variant-pack-template lifecycle fixes; SDPA attribute handling in the flash attention node; MXFP8 test synchronization.
• Benchmarks: SDPA benchmark now computes SOL% using the sampled SM clock and per-architecture MMA throughput; new CuTe-DSL fusion-kernel benchmark suite with initial B200/B300 results.

For full details, see the cuDNN Frontend Releases.

---

Coming Soon: cuDNN 9.25.0 / cuDNN Frontend v1.27.x

General Support of Rubin GPU architecture

✨✨Open source Attention, GEMM, and MoE kernels✨✨

Disclaimer: All early-preview / development items are made available to facilitate collaboration. We do not guarantee functional stability, performance, or compatibility before features are officially released.

---

Q3 Roadmap – July '26 – September '26

✨✨General Support of Rubin GPU architecture✨✨

✨✨Open source Attention, GEMM, and MoE kernels✨✨

✨✨Generalized support for CUDA Green contexts in the library✨✨

Attention / SDPA

• Open-source CUTLASS DSL kernel stack for SDPA to deliver highly optimized
  flash attention (forward and backward) for LLM training and inference
  workloads. Per-architecture feature coverage is tracked in the SDPA support
  matrix boards — contributions welcome:

  • Ampere (SM80): [RFE] ⚡ SDPA Support Matrix: Ampere (SM80) — open-source kernels wanted! #377
  • Hopper (SM90): [RFE] 🦗 SDPA Support Matrix: Hopper (SM90) — open-source kernels wanted! #378
  • Blackwell (SM100): [RFE] 🚀 SDPA Support Matrix: Blackwell (SM100) — open-source kernels wanted! #379
  • Blackwell Ultra (SM103): [RFE] 💫 SDPA Support Matrix: Blackwell Ultra (SM103) — open-source kernels wanted! #380
  • Blackwell GeForce / DGX Spark (SM120/SM121): [RFE] ✨ SDPA Support Matrix: Blackwell GeForce / DGX Spark (SM120/SM121) — open-source kernels wanted! #381

• Support d=512 fprop and bprop attention for Blackwell and newer architectures.

Linear Attention

• Open-source cuTile and CUTLASS DSL kernels for GDN/KDA/GDN-2 for LLM training and inference workloads to support Ampere (SM80), Hopper (SM90), Blackwell (SM100) and Blackwell Ultra (SM103).

Sparse Attention

• Continue improving perf for Video sparse attention (Block sparse attention) and support across architectures.

DSA (DSv4 Attention)

• Continue improving perf for Deepseek sparse attention (used in DSv4) and support across architectures.

MoE / Grouped GEMM

• Open-source CUTLASS DSL kernel stack for GEMM and MoE Grouped GEMM to deliver highly optimized fused GEMM and MoE Grouped GEMM patterns for LLM and MoE workloads.
• Initial support is for Swiglu with FP32, FP16, BF16, FP8, MXFP8, and NVFP4 data types that can be readily adopted into Deepseek style models

MegaMoe - Moe + EP Support

• Open source performant fprop and bprop variant of MegaMoE kernel that does MoE + Expert Parallel communication within a single kernel.

Convolution

• BF16 conv. perf improvements
  We are investigating torch issue for perf degradation in certain workloads for 4090/5090 style GeForce cards.

• Causal Conv1d support: Experimental causal conv1d support lives in the cudnn_extension backend library. It covers forward, backward, and fused variants of causal depthwise 1D convolution, including back-to-back causal conv1d fusion patterns and the common $y = \text{activation}(\mathtt{conv1d_causal}(x, w) + b)$
  form used by BioNeMo-style biological and life-sciences models.

Normalization

• FE fusion graph lowering to CUDA Tile: A new lowering path for memory bound kernels including normalization pattern like LN and RMS norm via Tensor IR. Tensor IR will be open source soon and provide lowering to cuda tile directly.

Tooling / Packaging / Compatibility

• CUDA Toolkit support targets. (13.x and 12.x toolkits)
• Improved PyPI preview-package cadence;
• Windows on ARM coverage.
• Improved diagnostic messages for environment issues.

Disclaimer: Some features above may ship in an update during Q3 (e.g., a 9.25.x / 9.26.x backend or a 1.27.x / 1.28.x frontend) or later.

---

Features Under Consideration

Let us know how to improve or prioritize these features for your deep learning and GPU workloads! Contributions, code feedback, issue submissions, and discussion are welcome.

[yeliu-oss]

We will also add support for FFT causal conv1d.

[Lazarus-931]

Hi, thanks for cuDNN-frontend!

I wanted to say KDA for sm90 support would be massively beneficial. FlashKDA currently supports that and having this would be very nice and appreciated!