The direct answer
Is Ryzen AI Software 1.8.0 a fixed public release? Yes. AMD publishes a non-draft, non-prerelease v1.8.0 release. Its lightweight tag resolves to GitHub-verified commit 73c04572fd7b.
Are the GitHub downloads the Ryzen AI runtime? No. GitHub exposes two generated repository snapshots and zero AMD-uploaded release assets. The fixed README calls the repository a subset of the full release; AMD's Windows EXE, NuGet ZIP and Linux TGZ sit behind separate account/EULA download forms.
Does the release verify robot inference performance? No. AMD reports one numerical Whisper RTF statement, but the public release surfaces do not identify the processor SKU, audio, baseline, timing protocol, repetitions or uncertainty. RCI ran no software, Ryzen AI device or robot workload.
AMD Ryzen AI Software v1.8.0 is a fixed non-prerelease GitHub release whose lightweight tag resolves to a GitHub-verified commit. AMD publicly names supported processor families, model classes, three runtime distribution files, eight new LLM/VLM entries, known limitations and one numerical Whisper RTF statement. The GitHub repository is explicitly a subset of the full release, its generated archives contain eight Git LFS pointers rather than their model payloads, and the runtime packages are outside the GitHub asset set. The audited surfaces do not provide the exact host/model/driver test matrix, protocol, raw logs, power/thermal data or robot workload results required to treat v1.8.0 as robot-computer validation.
Five distribution surfaces, three different boundaries
| Surface | Role | RCI check | Access and integrity boundary |
|---|---|---|---|
| GitHub-generated source tar.gz for v1.8.0 | generated repository snapshot | 41,269,858 bytes · 669 entries · RCI hash | Public GitHub-generated archive; includes eight Git LFS pointer files, not their referenced payloads. |
| GitHub-generated source ZIP for v1.8.0 | generated repository snapshot | 41,580,998 bytes · 669 entries · RCI hash | Public GitHub-generated archive; normalized path set matches the tar.gz, but GitHub publishes no digest for either generated download. |
| ryzen-ai-1.8.0.exe | Windows Ryzen AI Software installer | Not downloaded by RCI | Named by AMD documentation behind an AMD account/EULA download form; RCI did not accept terms or fetch the file. |
| ryzen_ai_nuget_1.8.0.zip | Windows NuGet distribution | Not downloaded by RCI | Named by AMD documentation behind an AMD account/EULA download form; no file byte count or publisher digest appears beside the audited documentation link. |
| ryzen_ai-1.8.0.tgz | Linux Ryzen AI Software package | Not downloaded by RCI | Named by AMD Linux documentation behind an AMD account/EULA download form; RCI did not accept terms or fetch the file. |
The two GitHub source archives expose identical normalized 669-entry path sets and RCI audit-date hashes, but neither is the complete Ryzen AI runtime and neither hash is publisher-supplied. The three documented runtime distributions were not downloaded or independently hashed by RCI.
The two generated archives have identical normalized 669-path sets. Their RCI SHA-256 values identify the bytes retrieved on the audit date; they are not publisher-supplied checksum matches. Eight apparent model files are Git LFS pointers declaring 404,905,134 bytes of payload that the archives do not contain.
Compatibility is a class-level issuer statement
| Processor families | AMD-listed model classes | RCI boundary |
|---|---|---|
| Phoenix / Hawk Point | CNN INT8 | Family-level AMD compatibility statement; no exact board, driver or robot workload is fixed here. |
| Strix / Krackan Point | CNN INT8 · CNN BF16 · NLP BF16 · LLM (OGA) | Class support does not establish that every model artifact, shape or operator passes. |
AMD's 1.8 release notes name Phoenix, Hawk Point, Strix, Strix Halo and Krackan Point across four Ryzen series groupings. The table above preserves the public class-level distinction; it is not a frozen list of exact order codes, boards, firmware or validated robot computers.
Processor-family and model-class tables are issuer compatibility statements. They cannot identify an exact Avnet/Weston CPU, carrier board, driver, robot, sensor stack or validated workload without a fixed configuration and result matrix.
What the release page actually claims
| Release statement | Evidence role | RCI boundary |
|---|---|---|
| Eight new LLM/VLM additions | issuer software-support statement | Names eight models; it does not publish exact model-file hashes, host configurations, pass logs or robot-task results. |
| Two TTS models, one embedding model and one translation model | issuer software-support statement | Support naming is not accuracy, latency, memory, power or production-readiness evidence. |
| Eleven Stable Diffusion model/application entries and dynamic-resolution labels | issuer software-support statement | The 5x and 20x labels count resolution variants; they are not speed multipliers or quality scores. |
| Performance optimizations with -O1 for SD3.x | issuer qualitative performance statement | No exact host, driver, model artifact, workload, before/after value, repetitions or uncertainty accompanies the statement. |
| Support running large models on low-memory machine | issuer qualitative capability statement | The release note gives no memory threshold, exact model, operating system, execution mode or observed failure boundary. |
| Three-second Whisper audio RTF reduced to 13.9% and 33.7% | issuer numerical performance statement | The release note names Whisper small and large-v3-turbo but no processor SKU, driver, build, audio sample, timing protocol, baseline, runs or uncertainty. |
| Full Fusion exclusions, a Windows-only TPS model and no Linux model generation | issuer known-issue statement | These restrictions remain attached to any support summary; Windows-generated model compatibility with Linux is issuer-reported. |
| Additional BF16 compiler option for openai-whisper-medium-encoder | issuer configuration requirement | The workaround identifies a configuration dependency, not a successful build or accuracy/performance result. |
The “5x” and “20x” Stable Diffusion labels count dynamic-resolution variants; they are not speed multipliers. The 13.9% and 33.7% RTF figures remain AMD claims, not RCI measurements.
The 13.9% and 33.7% Whisper RTF values are AMD release-note claims. Without exact hardware, software, audio, timing, baseline, repetition and uncertainty disclosures they cannot support a robot inference ranking or an RCI-reproduced result.
Public source improves inspectability; it does not become a result bundle
RCI's fixed-tree heuristic found 25 test-related paths, 3 benchmark-named scripts and 76 result-like files. Some example logs and images are present. None of the documented benchmark JSON output names is fixed in the tree.
GitHub exposes one successful update-pip-graph check and no publicly named build/test check at the tagged commit. That finding describes only public GitHub metadata; it does not assert that AMD ran no private or external validation.
Public examples, benchmark scripts, logs and images improve inspectability. They do not become release-level pass evidence without exact commands, complete dependencies, fixed model payloads and generated result files.
Twenty disclosure checks
| Field | Status | Public evidence | Why it matters |
|---|---|---|---|
| versioned release and date | disclosed | GitHub and AMD documentation identify Ryzen AI Software 1.8 and the release API fixes v1.8.0 publication time. | The software release can be cited and time-bounded. |
| tag and commit identity | disclosed-and-github-verified | The lightweight v1.8.0 ref resolves to commit 73c04572fd7b; GitHub reports the commit signature valid. | The repository snapshot is fixed while tag type and commit verification remain distinct. |
| publisher uploaded github assets | none | The GitHub release API exposes zero separately uploaded assets; the UI supplies two generated source downloads. | No AMD-uploaded installer, driver, model or result bundle is available at the GitHub release level. |
| github generated archive digests | rci-hash-only | RCI computed SHA-256 for both generated archives; GitHub publishes no digest for them. | The hashes identify audit-date bytes but cannot be called publisher digest matches. |
| runtime distribution identifiers | disclosed-metadata-only | AMD documentation names the Windows EXE, NuGet ZIP and Linux TGZ and links them through AMD account/EULA forms. | Users can identify intended package names, but the public GitHub release does not contain those runtime distributions. |
| runtime distribution bytes and digests | not-disclosed-on-audited-pages | No file byte count or publisher SHA-256 appears beside the three audited documentation download entries; RCI did not accept EULA terms or fetch the files. | Package identity and integrity cannot be independently matched from the audited public pages. |
| repository completeness | partial-explicit-subset | The fixed README explicitly calls the repository a subset of the Ryzen AI Software release. | Repository inspection cannot stand in for auditing the complete installed runtime. |
| git lfs payload availability in archives | pointer-only | Eight fixed-tree blobs are Git LFS pointers totaling 1,057 bytes and declaring 404,905,134 bytes of referenced payloads. | The generated archives do not carry the eight model payloads needed for a complete source-example checkout. |
| root source license | disclosed | The v1.8.0 root LICENSE.txt contains the MIT License. | The repository root source license is clear. |
| binary model driver and third party rights | item-specific | Runtime distributions use AMD EULA forms, while the repository references third-party models, datasets, dependencies and Git LFS content. | A blanket MIT statement would erase material item-level rights boundaries. |
| supported processor families | disclosed-family-level | AMD names five processor codenames and four Ryzen series groupings. | High-level hardware scope is visible. |
| exact processor board and robot matrix | not-frozen-in-release | The release notes point to AMD's living processor filter but do not attach a fixed order-code, board, firmware and robot-computer matrix to v1.8.0. | The release cannot identify an exact Avnet/Weston platform or prove support for every purchasable Ryzen AI system. |
| operating system driver and dependency matrix | partial | Windows and Linux pages name prerequisites, driver versions and Linux package names, but no complete exact-host outcome matrix is attached. | Installation instructions do not establish tested deployability across all exact configurations. |
| model type hardware matrix | disclosed-class-level | AMD maps CNN INT8/BF16, NLP BF16 and LLM OGA classes across PHX/HPT and STX/KRK. | The table supports class-level triage but not individual model compatibility conclusions. |
| exact supported model artifacts and hashes | not-released-with-notes | The release names models but attaches no complete model URLs, hashes, quantization configs, shapes and pass logs. | Another lab cannot reconstruct the exact tested model set from the release note alone. |
| general performance test protocol | not-disclosed | The -O1 and low-memory statements have no fixed host, driver, model, workload, baseline, repetitions or uncertainty. | The qualitative claims cannot support a cross-platform speed or memory ranking. |
| whisper rtf reproducibility | not-disclosed | The release reports two 3-second-audio RTF percentages but no processor SKU, audio identity, timing boundary, baseline, run count or raw output. | The only numerical performance claim cannot be independently reproduced from the release surfaces. |
| public benchmark source and fixed outputs | source-disclosed-results-not-fixed | The repository includes benchmark scripts, example logs and result images; documented benchmark JSON output names are absent from the fixed tree. | Runnable examples improve inspectability but do not prove a v1.8.0 release pass or a specific benchmark result. |
| public ci execution results | dependency-check-only | GitHub exposes one successful update-pip-graph check and zero publicly named build/test checks at the tag commit. | Public GitHub metadata does not supply a release build/test result matrix; private or external CI remains unknown. |
| independent robot system reproduction | not-performed | RCI audited metadata, generated archives, source structure and documentation only; it ran no software and accessed no Ryzen AI or robot hardware. | All application performance, compatibility, power, thermal and stability conclusions remain outside independent RCI verification. |
17 of 20 checks are narrower than fully disclosed. The missing core is an exact processor/board/driver/model/workload matrix with commands and raw numerical outputs—not another feature list.
MIT source, item-specific runtime and model terms
The fixed root LICENSE.txt contains the MIT License. AMD's runtime downloads use account/EULA forms; drivers, Git LFS objects, third-party models, datasets, dependencies, documentation and trademarks keep their own item-level terms.
MIT applies to the fixed root repository file; EULA distributions, NPU drivers, Git LFS model objects, third-party models and datasets retain their own item-level terms.
RCI links to the upstream sources and publishes original counts, audit-date hashes and disclosure analysis. It does not mirror AMD's archives, runtime packages, drivers or model payloads.
What RCI independently checked
- Resolved the official release/API, lightweight tag, verified target commit, recursive source tree and fixed root license.
- Downloaded both GitHub-generated source archives, computed audit-date SHA-256 values and compared their normalized 669-entry path sets.
- Inspected the README scope statement and Git LFS rules; counted eight pointer files and summed their declared payload sizes without fetching them.
- Read AMD's current release notes plus Windows and Linux installation pages; recorded package names, prerequisites, support statements and known issues.
- Counted test-, benchmark- and result-like paths and public GitHub checks without executing code or treating source as a passed result.
- Separated twenty release observations from twenty disclosure checks; all runtime, hardware and robot behavior remains outside independent verification.
Limits that stay attached
- RCI did not accept an AMD account/EULA download form or download the Windows EXE, NuGet ZIP, Linux TGZ or NPU driver packages.
- RCI downloaded the two GitHub-generated source archives only; their hashes are RCI audit-date identifiers because GitHub publishes no generated-archive digest.
- The fixed repository describes itself as a subset of the complete Ryzen AI Software release, so source-tree findings cannot be generalized to every installed runtime file.
- Eight model paths in the generated archives are Git LFS pointers; RCI did not pull or hash the 404,905,134 bytes of referenced payloads.
- RCI did not install, import, compile or execute Ryzen AI Software, its quicktest, examples, tests or benchmark scripts.
- RCI did not access a Ryzen AI processor, NPU driver environment, Avnet/Weston platform, sensor stack or robot workload.
- The Whisper RTF percentages, -O1 optimization and low-memory statements remain issuer-reported because the release surfaces do not disclose complete protocols or raw results.
- Test-, benchmark- and result-like path counts are explicit filename/path heuristics; they do not equal tests passed, benchmarks run or results valid.
- The root MIT license does not automatically govern EULA packages, drivers, third-party models, datasets, dependencies, documentation or trademarks.
- AMD's latest documentation and living processor list can change after the 2026-08-10 audit; the fixed Git tag and RCI data release preserve the audited boundary.
Download the release audit
Download the immutable JSON release and record-level CSV. Stable aliases are current JSON and current CSV.
Suggested citation: Robot Component Index. “Ryzen AI Software 1.8 Evidence: 8 LLM/VLM Additions, One RTF Claim, Zero Robot Benchmarks.” RCI 033, version 0.1.0, 2026-08-10. https://robotcomponentindex.com/research/ryzen-ai-1-8-release-evidence-audit/