docs: add datatype and hardware support matrix

docs/index.rst

@@ -30,6 +30,7 @@ Transformer Engine documentation
 
    installation
    getting_started/index
+   support_matrix
    faq
 
 .. toctree::

docs/support_matrix.rst

@@ -0,0 +1,139 @@
+..
+    Copyright (c) 2022-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+
+    See LICENSE for license information.
+
+Datatype and hardware support matrix
+====================================
+
+This page summarizes which low-precision quantization formats Transformer Engine
+supports on which NVIDIA GPU architectures. Support is keyed by CUDA *compute
+capability*, since that is what Transformer Engine checks at runtime to decide
+whether a format is available.
+
+.. note::
+
+   The support conditions below mirror the runtime capability checks in
+   ``transformer_engine.pytorch.quantization``
+   (``_compute_fp8_support``, ``_compute_fp8_block_scaling_support``,
+   ``_compute_mxfp8_support``, ``_compute_nvfp4_support``) and the BF16 check in
+   ``transformer_engine.pytorch.utils``. If those checks change, update this
+   table to match.
+
+Compute capability reference
+----------------------------
+
+.. list-table::
+   :header-rows: 1
+   :widths: 20 25 55
+
+   * - Compute capability
+     - Architecture
+     - Representative GPUs
+   * - 8.0, 8.6
+     - Ampere
+     - A100, A10, A40, RTX 30 series
+   * - 8.9
+     - Ada Lovelace
+     - L4, L40S, RTX 40 series
+   * - 9.0
+     - Hopper
+     - H100, H200
+   * - 10.0, 10.3
+     - Blackwell (data center)
+     - B200, B300, GB300
+   * - 12.0
+     - Blackwell (workstation / consumer)
+     - RTX PRO 6000, RTX 50 series
+
+The architecture and GPU columns are indicative; the compute capability column is
+the value Transformer Engine uses for its support decisions.
+
+Format support by compute capability
+-------------------------------------
+
+.. list-table::
+   :header-rows: 1
+   :widths: 22 13 15 17 17 16
+
+   * - Compute capability
+     - BF16
+     - FP8 (per tensor)
+     - FP8 block scaling
+     - MXFP8
+     - NVFP4
+   * - 8.0, 8.6 (Ampere)
+     - Yes
+     - No
+     - No
+     - No
+     - No
+   * - 8.9 (Ada)
+     - Yes
+     - Yes [1]_
+     - No
+     - No
+     - No
+   * - 9.0 (Hopper)
+     - Yes
+     - Yes
+     - Yes [2]_
+     - No
+     - No
+   * - 10.0, 10.3 (Blackwell DC)
+     - Yes
+     - Yes
+     - Yes [2]_
+     - Yes
+     - Yes
+   * - 12.0 (Blackwell workstation)
+     - Yes
+     - Yes
+     - Yes [2]_
+     - No [3]_
+     - Yes [4]_
+
+* **BF16** requires compute capability 8.0 or higher.
+* **FP8 (per tensor)** covers the :class:`~transformer_engine.common.recipe.DelayedScaling`
+  and :class:`~transformer_engine.common.recipe.Float8CurrentScaling` recipes.
+  It requires compute capability 8.9 or higher.
+* **FP8 block scaling** is the :class:`~transformer_engine.common.recipe.Float8BlockScaling`
+  recipe.
+* **MXFP8** is the :class:`~transformer_engine.common.recipe.MXFP8BlockScaling` recipe.
+* **NVFP4** is the :class:`~transformer_engine.common.recipe.NVFP4BlockScaling` recipe.
+
+.. [1] On Ada (compute capability 8.9), FP8 additionally requires cuBLASLt
+   version 12.1.3.x or higher and CUDA 12.1 or higher.
+
+.. [2] FP8 block scaling additionally requires CUDA 12.9 or higher.
+
+.. [3] MXFP8 is not yet supported on compute capability 12.0 and higher
+   (support is currently limited to compute capability 10.0 through 10.x).
+
+.. [4] The capability check reports NVFP4 as available on compute capability
+   10.0 and higher. The default NVFP4 recipe additionally uses a random Hadamard
+   transform and stochastic rounding, whose FP4 conversion instructions are
+   architecture specific to compute capability 10.0 and 10.3. On compute
+   capability 12.0, running the default recipe currently raises an
+   architecture-specific error, and NVFP4 there requires round-to-nearest
+   (stochastic rounding disabled).
+
+Default recipe by architecture
+-------------------------------
+
+When no recipe is passed explicitly, Transformer Engine selects a default based on
+the device (see ``get_default_fp8_recipe`` in
+``transformer_engine.pytorch.quantization``):
+
+.. list-table::
+   :header-rows: 1
+   :widths: 35 65
+
+   * - Compute capability
+     - Default recipe
+   * - 8.9, 9.0 (Ada, Hopper)
+     - ``DelayedScaling``
+   * - 10.0, 10.3 (Blackwell DC)
+     - ``MXFP8BlockScaling``
+   * - 12.0 (Blackwell workstation)
+     - ``Float8CurrentScaling`` (temporary, until MXFP8 supports all GEMM layouts)