Partners PRs

The collaboration between 10xEngineers, a services company specializing in AI compilers, and Andes Technology Corporation, a leading provider of high-performance, low-power 32- and 64-bit RISC-V processor IP and a Founding Premier Member of RISC-V International, delivers first-class AI workload compilation for Andes AX46MPV cores using 10xEngineers’ AI graph compiler, Baltoro. This collaboration enables fast, efficient, and scalable deployment of generative AI and computer vision models on systems powered by the AX46MPV, accelerating time-to-market for RISC-V-based AI platforms.

Baltoro is an MLIR-based, RISC-V-first AI compiler designed with a strong focus on performance, extensibility, and modularity. It provides out-of-the-box support for major open-source models – including LLaMA, Qwen, and DeepSeek – across leading frameworks such as PyTorch and ONNX. Advanced capabilities such as custom quantization, Flash Attention, paged attention, and KV caching are natively supported. Baltoro’s MLIR middle-end and LLVM-based code generation enable dynamic tensor handling, constant hoisting, RISC-V VLEN-aware tiling, and rapid prototyping of custom instructions and kernels. A compact runtime with both C and Python bindings allows seamless integration into production-grade inference-serving solutions.

The AX46MPV is designed for compute-intensive applications such as LLM and Transformer-based AI for powerful edge and data center and has already been adopted by several innovated customers. It is the third generation of Andes’ RISC-V vector core, following the NX27V and AX45MPV, and features:

• Architecture: 8-stage superscalar pipeline capable of dual-issuing key combinations of scalar and vector instructions
• Vector Performance: Supports the RISC-V Vector (RVV) extension v1.0, with configurable vector length (VLEN) up to 1024-bit and a datapath (DLEN) up to 1024-bit. The highly parallel Vector Processing Unit (VPU) can execute up to five RVV operations in parallel, with full chaining & pipelining
• Multi-core Scalability: Supports up to 16-core coherent cluster with private L1/L2 caches and a shared L3 cache
• Efficiency: Includes a dual-load/store unit to maximize data throughput for vector operations, and an enhanced High Bandwidth Vector Memory (HVM) subsystem with a crossbar design with up to 64 memory banks shared for up to 16 cores to keep up with the memory demand.

“Baltoro is purpose-built to deliver state-of-the-art large language model (LLM) and vision-language model (VLM) inference on RISC-V. Its modular and extensible architecture is designed to scale alongside the rapidly evolving RISC-V ecosystem, enabling seamless integration with next-generation accelerated systems. We are proud to collaborate with Andes Technology to optimize performance on their AX46MPV vector processor, unlocking new levels of efficiency and throughput for the latest generative AI models and applications.” Said Dr. Bilal Zafar, President and CEO, 10xEngineers.

“The AX46MPV is a powerhouse for generative AI, but hardware is only as good as the software that feeds it. By collaborating with 10xEngineers and their Baltoro compiler, we are removing the complexity of RISC-V vector programming. This partnership ensures that developers can take a model like LLaMA or DeepSeek and run it at the peak efficiency on Andes-powered silicon quickly and seamlessly.” Said Dr. Charlie Su, President and CTO, Andes Technology.