# Llama.cpp b9663 Advances SYCL Backend Parity, Accelerating the Decoupling of Inference from CUDA > The latest release expands operator support and unit testing for Intel and Huawei hardware, signaling a maturation of alternative inference ecosystems. **Published:** June 16, 2026 **Author:** PSEEDR Editorial **Category:** edge **Content tier:** free **Accessible for free:** true **Editorial format:** analysis **News quality eligible:** true **Source count:** 1 **Word count:** 1089 **Tags:** llama.cpp, SYCL, LLM Inference, Intel GPU, Huawei Ascend, Edge Computing **Canonical URL:** https://pseedr.com/edge/llamacpp-b9663-advances-sycl-backend-parity-accelerating-the-decoupling-of-infer --- The recent release of llama.cpp b9663 introduces critical operator support for the SYCL backend, marking another step in the open-source community's effort to decouple large language model inference from NVIDIA's CUDA ecosystem. By expanding unit test coverage and implementing the EXPM1 operator, this update ensures that alternative hardware platforms can achieve the execution correctness required for diverse enterprise edge deployments. The recent [release of llama.cpp b9663](https://github.com/ggml-org/llama.cpp/releases/tag/b9663) introduces critical operator support for the SYCL backend, marking another step in the open-source community's effort to decouple large language model (LLM) inference from NVIDIA's CUDA ecosystem. By expanding unit test coverage and implementing the EXPM1 operator, this update ensures that alternative hardware platforms-particularly Intel GPUs and Huawei Ascend chips-can achieve the execution correctness required for diverse enterprise edge deployments. ## Maturing the SYCL Backend Through Operator Parity The core technical advancement in release b9663 is the integration of Pull Request #24363, which brings support for the EXPM1 (exponential minus one) operator to the SYCL backend. In neural network inference, operators like EXPM1 are frequently utilized in specific activation functions, custom normalization layers, and handling numerically sensitive computations where standard exponential functions might introduce floating-point inaccuracies near zero. By natively supporting this operator in SYCL, llama.cpp allows models utilizing these mathematical structures to execute directly on Intel hardware without falling back to the CPU, thereby avoiding severe latency penalties. Furthermore, the release notes highlight the completion of unit test (UT) cases for the FLOOR, TRUNC, and ROUND operators under the SYCL backend, alongside new test cases for repeat and concat operations. This emphasis on unit testing is a critical indicator of backend maturity. In the context of LLM inference, particularly with quantized models, rounding and truncation operations are foundational. Discrepancies in how different hardware backends handle floating-point rounding can lead to cumulative errors during token generation, resulting in degraded model perplexity or hallucinated outputs. By ensuring these operators pass rigorous unit tests on SYCL, the maintainers are guaranteeing that Intel GPUs will produce deterministic, mathematically identical outputs to their CUDA or CPU counterparts. ## Expanding the Cross-Platform Hardware Matrix Llama.cpp's build matrix has grown into a comprehensive map of the current AI hardware landscape. Release b9663 maintains dedicated SYCL build targets, including specific pipelines for Ubuntu x64 (SYCL FP32) and (SYCL FP16), as well as Windows x64 (SYCL). This dual-precision support for Linux environments indicates a readiness for both high-precision tasks and memory-constrained, lower-precision inference on Intel discrete and integrated graphics. Equally notable is the continued investment in the openEuler ecosystem and Huawei Ascend hardware. The release includes specialized build targets for openEuler x86 and aarch64 architectures, specifically targeting the Huawei Ascend 910b via the ACL (Ascend Computing Language) Graph framework. The inclusion of these targets demonstrates a strategic push to support sovereign AI infrastructure. As enterprises globally navigate hardware export restrictions and supply chain diversification, the ability to deploy standardized inference engines on domestic silicon like the Ascend 910b becomes a critical operational requirement. Despite this aggressive expansion into alternative backends, the project maintains robust support for the dominant NVIDIA ecosystem, packaging Windows x64 builds with both CUDA 12.4 and CUDA 13.3 DLLs. This dual-track approach-sustaining state-of-the-art CUDA support while aggressively closing the feature gap with SYCL, Vulkan, and ACL-cements llama.cpp as a truly hardware-agnostic middleware layer. ## Implications for Enterprise Edge Deployments For enterprise architecture teams, the significance of llama.cpp b9663 lies in the mitigation of vendor lock-in at the inference layer. Historically, deploying LLMs in production required a strict dependency on NVIDIA hardware and the CUDA software stack to ensure performance and stability. This dependency creates friction for edge deployments, where power constraints, physical space, and hardware availability often necessitate the use of alternative silicon, such as Intel integrated GPUs or specialized NPUs. By filling gaps in operator support and enforcing strict unit test coverage across backends, llama.cpp is effectively commoditizing the execution environment. Enterprises can now design their AI applications against the llama.cpp API with increasing confidence that the underlying hardware-whether an Intel Arc GPU in a local workstation, a Huawei Ascend chip in a regional data center, or an NVIDIA RTX card-will execute the model graph correctly. This parity in execution correctness and stability is the prerequisite for dynamic workload scheduling, where inference tasks are routed to whatever compute is available rather than being bottlenecked by a specific vendor's hardware queue. ## Limitations and Open Technical Questions While the expansion of the SYCL backend is a positive signal, the release notes leave several technical questions unanswered. Most notably, the specific performance impact of the newly introduced EXPM1 operator on Intel GPU architectures remains undocumented. Achieving functional correctness is the first step, but whether the SYCL implementation of EXPM1 achieves memory bandwidth and compute utilization comparable to its CUDA equivalent is critical for high-throughput production environments. Additionally, the release does not specify which particular model architectures or activation functions currently supported by llama.cpp benefit most directly from the EXPM1, FLOOR, TRUNC, and ROUND operator optimizations. Without this context, developers must profile their specific models to determine if migrating to the latest SYCL build will yield tangible performance or accuracy improvements. Finally, the build matrix reveals that the macOS Apple Silicon target with KleidiAI enabled is currently marked as DISABLED. KleidiAI represents Arm's optimized micro-kernels for AI workloads, and its disabled status suggests ongoing integration friction or unresolved stability issues on Apple's ARM architecture. The exact nature of these blockers is not detailed in the release, leaving a gap in the optimization path for macOS-based edge deployments. The trajectory of llama.cpp continues to reflect a broader industry mandate: the democratization of AI inference across all available silicon. Release b9663 is not defined by a singular, disruptive feature, but rather by the meticulous, necessary work of achieving mathematical and operational parity across competing hardware ecosystems. As alternative backends like SYCL and ACL Graph mature through rigorous operator support and unit testing, the open-source community is successfully building an abstraction layer that insulates developers from the complexities of hardware fragmentation. ### Key Takeaways * Llama.cpp b9663 introduces EXPM1 operator support and expands unit testing for FLOOR, TRUNC, and ROUND in the SYCL backend. * The release strengthens cross-platform inference capabilities, notably supporting Intel GPUs via SYCL and Huawei Ascend 910b hardware via ACL Graph on openEuler. * Achieving operator parity across backends mitigates vendor lock-in, allowing enterprises to deploy LLMs on diverse edge hardware without sacrificing execution correctness. * Performance metrics for the new SYCL operators and the reasons behind the disabled macOS KleidiAI build target remain undocumented. --- ## Sources - https://github.com/ggml-org/llama.cpp/releases/tag/b9663