NextSilicon Maverick-2: Redefining AI and HPC With New ICA
While the broader tech ecosystem remains fixated on traditional GPU cores and rigid ASIC logic for computational acceleration, emerging innovators continue exploring alternative computing paths. NextSilicon has introduced a radically distinct model through its Maverick-2 Intelligent Compute Architecture (ICA). By blending real-time adaptability with high-performance metrics via reconfigurable, software-defined hardware acceleration, the system functions essentially like an advanced, highly approachable FPGA where the software shoulders the heavy lifting.
As a foundational baseline, the engineering team integrated a high-speed dual-core CPU utilizing the open-source RISC-V framework, further enhanced with 256-bit vector processing and out-of-order execution. This design choice positions the setup advantageously for future high-performance computing (HPC) environments as the broader RISC-V ecosystem matures. These execution cores feed directly into the Maverick-2 engine, managing scalar operations and control protocols while effectively offloading the primary host processor.
Maverick 2: A Reconfigurable Accelerator for HPC and AI
The Maverick-2 chip introduces an innovative computing paradigm driven by intelligent software-defined hardware acceleration, yielding superior real-time adaptability and robust performance. Initially tailored for demanding HPC sectors—such as national research laboratories, quantitative hedge funds, and geoscience enterprises—the technology targets vector databases and artificial intelligence applications next. Featuring native 64-bit math support, NextSilicon’s exhibition space at Supercomputing consistently drew dense crowds of engineers and scientists analyzing both the Mav2 hardware and the accompanying RISC-V demonstrations.
Operating as a reconfigurable dataflow architecture, Maverick-2 utilizes software frameworks to map complex algorithms dynamically across an array of arithmetic execution units (ALUs). Historically, dataflow structures have faced adoption hurdles due to the extensive associative structures and extra clock cycles required to monitor operand availability, tag data tokens, and pair them with correct instructions. For code lacking deep parallelism, this operational overhead has traditionally degraded efficiency—though such workloads fall outside NextSilicon's core targets.
To mitigate these structural bottlenecks, the integration of RISC-V cores plays a crucial role. Rather than forcing developers to adopt custom dataflow programming languages, the ICA compiler ingests standard development tools including C++, Fortran, Python, OpenMP, Kokkos, and CUDA. It automatically translates these inputs into an intermediate representation to extract dataflow graphs seamlessly. Furthermore, compatibility layers for ROCm and HIP are currently under active development to accommodate AMD codebases.
While industry analysts await comprehensive benchmark evaluations, early figures appear promising. NextSilicon reports achieving up to four times greater FP64 performance-per-watt metrics relative to the Nvidia HGX B200, alongside greater than twenty times the energy efficiency demonstrated by Intel Xeon Sapphire Rapids processors on selected scientific workloads.
Shifting toward a dataflow instruction set typically mandates sweeping overhauls across compilers, languages, and existing code repositories. NextSilicon’s software-centric strategy provides a compelling workaround for the structural pitfalls that previously derailed ventures like Esperanto.
The physical hardware presentation reveals a dual 5-nanometer die layout, accompanied by eight HBM3e memory stacks and embedded RISC-V cores positioned strategically between them. An alternative air-cooled configuration operates on a single die while consuming half the thermal envelope, with subsequent generations slated to incorporate four cores.
By computing natively in FP64—which demands substantially more silicon acreage than compact 8-bit or 16-bit units—Maverick-2 maintains competitive AI processing capabilities through weight conversion techniques. Although an NVFP4 floating-point unit occupies a footprint roughly 256 times smaller than an FP64 block, it was omitted because modern inference bottlenecks are predominantly constrained by input/output and memory bandwidth rather than raw compute capacity. Expanding into artificial intelligence represents a secondary strategic phase following market establishment within HPC.
NextSilicon Software Strategy
A central pillar of NextSilicon’s approach is running legacy software unmodified, allowing decades-old applications and standard C programs to execute without manual code refactoring. The underlying software stack functions autonomously via telemetry-driven optimization loops. By continuously gathering hardware telemetry regarding memory utilization and runtime behavior, the system iteratively recompiles and fine-tunes execution performance.
Whereas prevailing AI ASICs championed by major cloud providers are meticulously engineered to execute transformer-based large language models (LLMs), NextSilicon asserts its architecture remains resilient against post-transformer paradigms—including Mamba-style frameworks and non-standard attention kernels. This flexibility permits varied kernel execution without demanding costly model retraining or fine-tuning investments that can reach tens of billions of dollars.
Market Positioning and Business Development
NextSilicon initially concentrated its deployment efforts on HPC environments through the Maverick-1 generation, building strong customer co-design familiarity prior to venturing into wider AI workloads.
Company executives explicitly differentiate their strategy from competitors like Groq and SambaNova. While rival architectures optimize for maximum token throughput per user via dense 3D DRAM or SRAM stacking, NextSilicon prioritizes concurrent user density—supporting between 50 and 150 simultaneous users across stacked units—to handle massive models efficiently without necessitating cumbersome setups involving dozens of server racks.
Evaluating the Need for a New Computing Architecture
Both high-performance computing and artificial intelligence share common demands, most notably the requirement for energy-efficient, high-throughput semiconductors. Cloud hyperscalers and foundational AI developers alike are actively designing proprietary chips to mitigate reliance on Nvidia, while HPC facilities continually test experimental architectures in pursuit of enhanced speed and efficiency.
Illustrating this trend, NextSilicon partnered with Sandia National Laboratories to deploy Spectra, a 64-node supercomputer developed under Sandia's Vanguard initiative. The system leverages NextSilicon’s reconfigurable Maverick-2 accelerators to evaluate novel computing models for HPC, AI, and nuclear deterrence simulations. Additionally, NextSilicon delivered four Maverick-2 accelerators to Los Alamos National Laboratory during SC26, integrating them directly into the institution's Darwin testbed environment.
Should the platform continue delivering on its technical promises, broader adoption across the HPC community seems assured, paving a clear pathway into the expanding landscape of heterogeneous artificial intelligence.