NVIDIA has started accepting AMD pull requests into its NIXL Inference communication library, ending a 10-year upstreaming freeze. SemiAnalysis brokered the détente over the past 24 months, according to the firm's report.
Key facts
- 10-year block on AMD upstreaming into NVIDIA libraries
- 24 months of SemiAnalysis brokerage for the détente
- NIXL is NVIDIA's inference communication library
- AMD MI300X and MI400 accelerators gain NIXL access
- SemiAnalysis published the report on X (formerly Twitter)
NVIDIA has started accepting AMD pull requests into its NIXL Inference communication library, ending a 10-year upstreaming freeze. SemiAnalysis brokered the détente over the past 24 months, according to the firm's report. According to @SemiAnalysis_, the move is a significant shift in the competitive dynamics of AI inference infrastructure.
Key Takeaways
- NVIDIA accepted AMD PRs into NIXL after 10-year block, brokered by SemiAnalysis over 24 months.
- The move could reshape AI inference infrastructure competition.
What NIXL Actually Does
NIXL is the communication backbone for multi-GPU inference workloads, handling the data transfers between GPUs during model execution. For a decade, NVIDIA maintained a closed-door policy that prevented AMD from contributing code to any of its libraries, forcing AMD to maintain separate, often less optimized communication stacks. The acceptance of AMD PRs into NIXL means AMD hardware can now leverage the same optimized communication primitives as NVIDIA GPUs, potentially closing the performance gap in multi-GPU inference scenarios.
The change is not merely symbolic. AMD's MI300X and upcoming MI400 series accelerators have been competitive on raw compute but have lagged in the software ecosystem. With NIXL access, AMD can now ensure its hardware is first-class citizen in the inference communication layer, rather than relying on workarounds or third-party middleware.
The SemiAnalysis Brokerage
The détente was not accidental. SemiAnalysis, the semiconductor research firm known for its deep technical reporting, spent 24 months acting as an intermediary between the two companies. The firm's role in squashing the beef highlights how technical journalism can influence corporate strategy in the AI chip race.
The specifics of the negotiation are not disclosed, but the outcome is clear: NVIDIA has recognized that a fragmented communication ecosystem hurts the entire AI industry. By accepting AMD code, NVIDIA ensures that NIXL becomes the de facto standard for inference communication, even on competing hardware.
What This Means for the Inference Market
The implications are significant. AMD has struggled to gain enterprise traction despite competitive hardware, largely due to software gaps. With NIXL support, AMD's ROCm stack gains a critical piece of interoperability that could make multi-vendor deployments more practical. This could pressure NVIDIA's dominant position in AI data centers, where its CUDA ecosystem has been a moat.
However, the move is not without risks for NVIDIA. By legitimizing AMD's participation in NIXL, NVIDIA may be ceding some control over the inference stack. But the alternative — maintaining a closed ecosystem — risks pushing customers toward fully open alternatives like the UXL Foundation or other multi-vendor initiatives.
The development also signals a broader trend: as AI inference becomes a commodity workload, hardware vendors are being forced to cooperate on the software layer. The days of proprietary, single-vendor communication libraries may be numbered. [The company did not disclose the specific PRs accepted or the timeline for full integration] (@SemiAnalysis_ has not provided those details in the public thread).
What to watch
Watch for AMD's next ROCm release to include NIXL-based communication primitives, and whether NVIDIA's next NIXL version documents AMD-specific optimizations. Also monitor whether the détente extends to other NVIDIA libraries like NCCL (training) or cuBLAS, which would signal a broader policy shift.
