Comprising of four members, the first device in the series is the AX46MPV, a 64-bit multicore superscalar vector processor IP, the third generation of the Andes Vector core.
While it maintains the same 8-stage dual-issue pipeline as its predecessor AX45MPV, it incorporates numerous new features such as dual load/store units, private L2 cache, an extended VLEN, Andes Matrix Multiply Extension, and RVA22 profile support.
Except vector and matrix support, the second member, AX46MP, has the exact same features as the AX46MPV while the third member, A46MP, and fourth member, A46MPV, are the 32-bit versions of AX46MP and AX46MPV.
This makes the 46-series suitable for network processors running Linux, high-performance embedded controllers requiring high memory throughput, and large-scale AI/ML applications.
The AX46MPV enhances SpecInt2006 performance by over 15% compared to the previous generation AX45MPV, thanks to its newly designed memory subsystem. The private L2 cache significantly reduces memory latency while the dual load/store engine eliminates memory bottlenecks in memory-bound computation kernels.
It can support up to 16 cores in the same cluster, featuring cache coherence and a shared L3 cache with doubled capacity of the AX45MPV. Additionally, the AX46MPV provides TrustZone-level security with ePMP and IOPMP, accompanied by a pre-integrated software solution.
The AX46MPV also see the introduction of several enhancements for AI/ML, including longer 2048-bit vectors, new Andes Matrix Extensions to efficiently speed up GEMM performance, and greatly improved High-Bandwidth Vector Memory (HVM) interface supporting outstanding requests and out-of-order responses.
A high-performance HVM controller serving up to 16 cores and DMA accesses with 64 memory banks is optional for licensing. The AX46MPV now also includes BF16 full arithmetic mode as a standard feature.
Additionally, the Andes Automated Custom Extension (ACE) is available in the 46-series, featuring customised vector instructions (ACE_RVV) with enhanced pipeline and Streaming Port (ASP).
Commenting Dr. Charlie Su, President and CTO of Andes Technology, said, "Our NX27V and AX45MPV have been highly successful in high-performance AI SoC applications. The AX46MPV, boosted with both compute performance and memory bandwidth, is intended to bring the new-generation AI SoCs to the next level of compute density. Additionally, with enhanced capabilities in Linux, security, and AIoT, the highly configurable 46-series is a well-rounded solution with balanced performance, power, and area"
The AndesCore 46-series processor IP’s, 64-bit AX64MP(V) and 32-bit A64MP(V), can be configured from one core up to 16 cores.
It will be available for lead customers in Q1 2025 through the early access program and for general customers in Q2 2025.
