Glossary

The words this site uses for ROCm, gfx IDs, HIP, Vulkan, and home-lab inference.

Short definitions. If a word has an official AMD sentence, that is cited. Site jargon is labeled as ours.

Platform

ROCm: AMD’s open-source GPU computing platform: compilers, runtimes, and libraries for AI, HPC, and other workloads, on Linux and Windows, for Instinct, Radeon, and Ryzen AI.^[1]Historically Radeon Open Compute. See overview.
CUDA: NVIDIA’s GPU compute platform. HIP exists in part so CUDA-like code can be ported; HIP is not a drop-in CUDA replacement.^[2]
TheRock: ROCm’s build and release system since 7.14.^[12]^[14]
ROCm CLI: Official rocm binary for install / serve / diagnose / dashboard. 10.1 blog: v0.2.0.^[13]^[15]

gfx ID: LLVM target for the GPU ISA — gfx1201, gfx1100, and so on. The 10.1.0 hardware table lists each supported device with its LLVM target and architecture.^[12]gfx page · directory.
RDNA: Radeon (and many Ryzen iGPU) architecture family. R9700 / RX 9070-class = RDNA 4 /gfx1201 on the 10.1.0 table.^[12]
CDNA: Instinct datacenter architecture family. Same ROCm docs, different SKUs and often different OS footnotes.^[12]
Radeon AI PRO R9700: Workstation / local-AI card. 32 GB. gfx1201. Our test-rig GPU. There is no consumer “RX 9700.” Directory.

amdgpu: In-tree Linux DRM driver for modern AMD GPUs (GCN, RDNA, CDNA).^[4]
KFD / ROCK: Kernel-side compute interface. ROCR docs: user-mode runtime talks to the AMDGPU driver set and the ROCK kernel driver.^[3]On current Linux this is amdkfd / /dev/kfd.
ROCr: AMD’s user-mode HSA runtime.^[3]
HIP: C++ runtime API and kernel language for AMD GPUs.^[2]
comgr: Code Object Manager — compile, link, inspect GPU code objects. HIPRTC uses it.^[5]
rocBLAS: ROCm BLAS library, HIP C++, for AMD GPUs.^[6]
hipBLASLt: Flexible GEMM beyond classical BLAS.^[7]
MIOpen: AMD deep-learning primitives library.^[8]
RCCL: ROCm Communication Collectives Library — multi-GPU / multi-node primitives.^[9]
rocWMMA: Header library for mixed-precision matrix multiply-accumulate.^[10]
Composable Kernel (CK): Programming model for performance-critical ML kernels in HIP C++.^[11]
amd-smi: Successor to rocm-smi after the deprecation period, per the 10.1 blog.^[13]
HSA_OVERRIDE_GFX_VERSION: Unofficial env var used in community threads to load a nearby gfx’s code objects (example:10.3.0 on gfx1031).^[16]Not an official recommendation. Not needed on official gfx1201 + ROCm 10.x.Details.
Vulkan: Khronos graphics/compute API. llama.cpp can use it instead of HIP.^[17]Vulkan vs HIP.

Official (our column): Appears on AMD’s current matrix or matching release notes. Not a speed claim.Policy.
Works / try: Ran here or in a cited community report, not on the official row. Labeled.
Daily driver: Qwen3.6-35B-A3B UD-Q5_K_XL, llama.cpp Vulkan, one R9700. Measured on our rig, 2026-10 — 100.8 / ~88 / ~50 tok/s. Stacks.
Fast agent: Qwen 3.8 Flash UD-Q6_K_XL on Strata v0.1.42. First-party ~32.6 vs ~16 tok/s.
House rules: llama.cpp over vLLM for consumers; Q5 over Q4; never below Q4; NVMe not HDD; one LLM per GPU unless you deliberately combine.

Every version and support claim on this page points here. If a sentence has no mark, treat it as lab opinion or a TODO — not as AMD policy.