docs: flesh out some glossary entries

pkgs/development/cuda-modules/doc/glossary.md

@@ -4,8 +4,58 @@
 
 ## Terminology
 
-- compute capability:
+- Compute capability (CC):
+
+    This is the architecture of, and thus features supported by, a given NVIDIA GPU's hardware.
+    Refer to [NVIDIA's CUDA GPU Compute Capability table](https://developer.nvidia.com/cuda-gpus)
+    or [this table on Wikipedia](https://en.wikipedia.org/wiki/CUDA#GPUs_supported) to find the CC of a given GPU.
+
+    Generally, each major CC version corresponds to a major GPU architecture revision (e.g. Maxwell = CC 5.x, Pascal = CC 6.x, Volta = CC 7.x).
+
+    Also known as a device's "SM version".
+    Not to be confused with the CUDA software platform version.
+
+    Reference: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#compute-capability
+
+- (CUDA) binary code:
+
+    This is the actual binary code that is executed by the GPU.
+
+    It is compute capability (CC) specific.
+    Binary code generated for a given CC may run on later minor CC versions.
+    It may not run on previous minor CC versions or any other major CC version.
+
+    Note: this term may be overloaded.
+    You may be familiar with the concept of a non-CUDA compilers producing what is traditionally known as a "binary object".
+    NVCC can produce a traditional "binary object" which does not contain CUDA binary code (and instead only contains CUDA PTX code, see below).
+    Outside of this note, this FAQ will only use the term "binary" to refer to "CUDA binary code", and will refer to a traditional "binary object" as simply an "object".
+
+    Reference: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#binary-compatibility
+
+- PTX code:
+
+    This is an intermediate language produced by NVCC and embedded into its produced output object.
+    It is compiled at runtime by the CUDA driver when loaded, allowing for targeting
+    of multiple CUDA architectures at the expense of some additional startup cost.
+
+    PTX code generated for a given compute capability (CC) can not run on previous CC versions.
+    However, unlike CUDA binary code, it can run on all later CC versions, including versions not yet released.
+    (Note that the PTX code will not be able to utilize features from later CC versions.)
+
+    Thus, to generate code for a variety of CC targets, with forward and backward compatibility,
+    a packager can choose to generate binary code to support all previous major CC versions,
+    along with PTX code for the current (and implicitly future) major CC versions.
+    In CMake 3.23+, this can be done simply by setting `-DCMAKE_CUDA_ARCHITECTURES=all-major`.
+
+    Reference: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#ptx-compatibility
+
 - real architecture:
+
+    When CUDA code is being compiled to binary, this is the compute capability it is targeting.
+
 - virtual architecture:
+
+    When CUDA code is being compiled to PTX, this is the compute capability it is targeting.
+
 - CUDA package set:
 - extension: