Israeli startup NextSilicon has launched its Maverick-2 dataflow processor, positioning the chip as a high-performance computing alternative to offerings from industry leaders Nvidia, Intel, and AMD. The announcement comes alongside details of a new RISC-V CPU called Arbel, signaling the company’s ambition to compete across multiple segments of the semiconductor market.
After eight years of development and $303 million in funding, NextSilicon is bringing its 64-bit dataflow engine to market. The company’s technology is already being evaluated by Sandia National Laboratories, which assisted in developing an earlier proof-of-concept processor and is expected to be among the first to deploy production systems using the new chips.
Revolutionary Architecture for HPC Workloads
The Maverick-2 processor represents a fundamental departure from traditional computing architectures. Rather than following the conventional Von Neumann design used in most CPUs, NextSilicon has implemented a dataflow engine that dynamically maps hardware resources to software requirements.
“This architecture allows us to achieve significantly lower overhead compared to traditional CPUs and GPUs,” said Ilan Tayari, NextSilicon’s co-founder and VP of Architecture. “We dedicate the majority of resources to actual computation rather than control overhead.”
The chip features 54 billion transistors manufactured using Taiwan Semiconductor Manufacturing Co’s 5-nanometer process. Its unique design includes thousands of arithmetic logic units organized into compute blocks that can be dynamically reconfigured to match application requirements.
Software Compatibility and Automated Optimization
A key innovation lies in NextSilicon’s compiler technology, which automatically optimizes existing C, C++, and Fortran code for the dataflow architecture without requiring developers to learn new programming models. The system continuously analyzes and improves performance during execution, potentially achieving higher utilization rates than traditional processors.
“NextSilicon’s performance results are impressive, showing real promise for advancing our computational capabilities without the overhead of extensive code modifications,” said James H. Laros III, senior scientist at Sandia National Laboratories.
Competitive Positioning and Performance Claims
While peak performance numbers for the Maverick-2 appear modest compared to recent Nvidia GPUs—approximately a third of the H100’s FP64 performance—NextSilicon emphasizes real-world efficiency over theoretical peaks. The company claims significant advantages in performance-per-watt across various benchmarks, including:
- 22x faster than CPUs on memory subsystem tests (GUPS benchmark)
- 6x faster than GPUs on the same memory tests
- Matching leading GPU performance on HPCG benchmark at half the power consumption
- 10x better performance on graph analytics benchmarks compared to leading GPUs
Expanding the Portfolio with RISC-V CPU
In a strategic move mirroring Nvidia’s approach, NextSilicon also revealed Arbel, a homegrown RISC-V processor currently in the test chip phase. The company claims the Arbel core can compete directly with Intel’s Lion Cove and AMD’s Zen 5 architectures while avoiding Arm licensing fees.
The dual-chip strategy positions NextSilicon to offer complete computing solutions as it seeks to capitalize on what it sees as an opening in the high-performance computing market created by Nvidia’s increased focus on AI applications.
With production Maverick-2 systems now available and customer evaluations underway, NextSilicon aims to establish itself as a viable alternative for scientific computing and simulation workloads where precision and efficiency remain paramount.
