The intriguing startup NextSilicon, which has generated attention in HPC and other advanced computing markets since its founding in 2017, today disclosed architecture and performance details for Maverick-2, its new processor that the company says enables order-of-magnitude compute improvements for HPC and AI.
Here’s a run-down of what NextSilicon said about Maverick-2:
- Built on the Intelligent Compute Architecture, the dataflow-enabled hardware adapts to provide peak performance regardless of software branching or parallelism.
- In benchmarks released today, Maverick-2 achieves up to 10x performance over leading GPUs at 60 percent less power in algorithmically complex workloads.
- Maverick-2’s architecture allows users to bring their own unmodified code – CUDA, Fortran, etc., to run out-of-the-box with no specific optimizations necessary.
- The company also unveiling Arbel, an enterprise-grade high-performance RISC-V core built on TSMC’s 5nm process.
NextSilicon said Maverick-2 is running at dozens of customer sites, including Sandia National Laboratories (see previous coverage) in its Vanguard-II supercomputer.
NextSilicon Maverick-2
“Over the past three years, our collaboration with NextSilicon has focused on evaluating emerging architectures for next-generation HPC workloads critical to national security,” said James H. Laros III, Senior Scientist and Vanguard program lead at Sandia National Laboratories. “The out-of-the-box HPCG performance results are impressive, showing real promise for advancing our computational capabilities without the overhead of extensive code modifications. Maverick-2’s dataflow approach demonstrates strong potential in real-world scenarios, and we’re continuing to explore its full performance envelope.”
At a pre-announcement briefing, NextSilicon founder and CEO Elad Raz said Maverick-2 focuses on high-precision computing, which is associated with HPC. This garnered a salute from at least one industry observer, Addison Snell, co-founder and CEO of industry analyst firm Intersect360, who said that “the dataflow architecture of Maverick-2 is something new, filling the gap in the market by targeting segments of the market that aren’t well served by low-precision LLMs.”
Snell said that “HPC is a stable growth segment that is getting ignored in the fervor for AI, to the potential detriment of scientific applications and their practitioners. The needs of the technical computing community are as compelling as ever. For this community, the Maverick-2 accelerator from NextSilicon might as well fly in wearing a cape, here to save the day.”
He added that the company’s focus on high-precision performance “harkens back to the Cray days, and it similarly led them to a new approach to fast, scalable performance with the dataflow architecture. Sure, the graph performance means hyperscalers will be interested, but I respect that their home market is technical computing.”
Elad Raz, NextSilicon
Based in Israel, NextSilicon has more than 350 employees globally. The company has raised more than $300 million in funding from investors Amiti, Aleph, Liberty Technology VC, Playground Global, Standard Investments, StepStone and Third Point Ventures. The company said its partner ecosystem includes native support for commercial HPC libraries with partners such as NAG and Bio Team, as well as hosting and integration services with partners like Dell Technologies, Penguin Solutions, Databank, Vibrint, E4 and Partec.
“For eight decades, rigid hardware designs have forced software to adapt, with modern processors dedicating roughly 98 percent of silicon to overhead and just 2 percent to actual computation,” said Raz. “Maverick-2 flips this paradigm on its head by devoting the majority of hardware real estate to compute, moving runtime overhead management to intelligent algorithms and software in real-time through our dataflow architecture.”
Designed for data-intensive and algorithmically complex workloads, Maverick-2 introduces capabilities that conventional GPUs cannot achieve, Raz said. Enabling out-of-the-box acceleration for both legacy and emerging applications simplifies deployment and future-proofs investments by ensuring flexibility as workloads evolve.
By removing the need for time-consuming code rewrites or proprietary software stacks, Maverick-2 accelerates time-to-results, reduces operational costs, and unlocks new possibilities for applications ranging from agentic AI real-time decision making or advanced reasoning models processing complex queries to modeling irregular data patterns of complex proteins for new medical treatments.
NextSilicon also released performance data for Maverick-2 across HPC benchmarks that the company said match or exceed popular accelerators and were achieved using unmodified, out-of-the-box application code, “with no specialized programming, vendor-specific optimizations, or lengthy porting cycles.”
Benchmarking results are available at www.nextsilicon.com/maverick, with highlights including:
- PageRank: 10x higher graph analytics performance than leading GPUs, enabling workloads on 25GB+ graph sizes that GPUs cannot process.
- GUPS (Giga-Updates Per Second): 32.6 GUPS at 460 watts, 22x faster than CPUs and nearly 6x faster than GPUs, for high-throughput databases and AI inference.
- HPCG (High-Performance Conjugate Gradients): 600 GFLOPS at 750 watts, matching leading GPUs at half the power and delivering performance-per-watt comparable to the world’s Top 500 systems..
- STREAM: Validated full performance HBM3e for your most demanding applications.
Looking ahead to a potential future architecture, NextSilicon today also unveiled Arbel, an enterprise-grade RISC-V core using patented intellectual property on TSMC’s 5nm process and designed to compete with offerings from AMD and Intel, the company said. “The test chip demonstrates NextSilicon’s ability to build world-class IP and explore multiple strategic paths from vertically integrated solutions to enabling a broader RISC-V ecosystem in datacenter environments,” NextSilicon said.
Maverick-2 is now available in product volumes. The company will be at booth #3824 during the SC25 Supercomputing Conference, held from November 16-21, 2025, in St. Louis.
