AMD's Software Updates Boost AI Chip Performance Without New Hardware

AMD has reported that its AI processors are delivering up to 38% better performance on large language models, despite using the exact same physical hardware as previous tests. This improvement comes entirely from updates to the company's software stack, known as ROCm. The results, published in the MLPerf Inference 6.1 benchmark, mark a significant shift in how AI infrastructure evolves, prioritizing rapid software iteration over waiting for new silicon.

For two years, the main criticism of AMD's Instinct chips has been that their software was not as mature or reliable as competitors'. By committing to a strict six-week release schedule, AMD is turning that criticism into a measurable metric. As reported by GN technics/hardware (en-US), this change allows customers and analysts to track performance gains with precision, making the software's progress a tangible asset for procurement decisions.

Software updates drive major performance gains

The core of this news is that speed is now coming from code, not circuitry. In the latest test round, AMD's MI355X chips processed the GPT-OSS-120B model 38% faster than they did in the previous cycle. Another model, Wan 2.2, saw a 70% increase in single-stream performance. These improvements happened without changing a single component in the machine; only the instructions telling the chips how to work were updated.

This approach offers a practical benefit for businesses already using AMD hardware. If a company bought these chips for earlier workloads, they now have significantly more effective capacity without spending money on new equipment. The extra power arrived simply through software package updates, a model that is increasingly common in the high-performance computing world but is particularly notable here because of the scale of the gain.

Six-week cycle creates predictable upgrade path

To make these gains consistent, AMD has moved from a roughly quarterly release pattern to a fixed six-week cadence. This shift was enabled by a new automated build system called TheRock, which reached production in July. By standardizing the release timeline, AMD provides a clear schedule for when new features and optimizations will become available. This predictability is crucial for IT planners who need to budget and schedule system upgrades.

The first release under this new system, ROCm 10.0, introduced tools for automated coding and workload optimization. While the company claims an average 3.3x inference improvement over older versions, the real value lies in the reliability of the delivery. Users can now expect regular, incremental improvements rather than large, unpredictable jumps. This reduces the risk associated with adopting new AI infrastructure, as the path to better performance is now documented and scheduled.

Caveats remain for broad adoption

Despite the strong results, there are important limitations to consider. The comparisons are based on specific models and configurations, and not all AI workloads will see the same degree of improvement. Additionally, AMD did not submit results for its newest MI455X chips in this round, positioning that hardware for a different benchmark format called MLPerf Endpoints. This means the full picture of AMD's current hardware capabilities is not yet fully captured in these specific numbers.

Furthermore, the benchmark comparisons against NVIDIA hardware are sensitive to the specific version of the software used by the competitor. The gap between AMD and NVIDIA can vary depending on how each company optimizes their stacks for the test. While AMD's progress is clear, the trade-off is that users must still carefully evaluate whether the specific software updates align with their unique application needs. The speed of delivery is now a strength, but it does not eliminate the need for detailed technical validation before widespread deployment.