EdgeCortix Announces 3.36 PFLOPS Chip, Surpassing NVIDIA

Summary

Japanese chipmaker EdgeCortix has announced its 4-chip RAIDEN X4 module. The module delivers 3.36 PFLOPS of performance at FP4 precision. The company claims it outperforms NVIDIA's Jetson AGX Thor (2,070 TFLOPS at FP4) in edge AI workloads. The design utilizes a DNA-X neural accelerator, RISC-V cores, 256 GB of LPDDR5X memory, and UCIe chiplet connectivity.

Why it matters

This development signals the emergence of a reliable non-NVIDIA option in the edge AI chip sector. In this field, power budgets and TOPS-per-watt are more decisive than raw FLOPS. For OEMs developing robots, industrial control units, or in-vehicle AI systems, a 60% performance gap at the same precision level could influence procurement decisions. It also highlights Japan's continued efforts in the high-performance domestic AI chip space.

Highlights

• The 3.36 PFLOPS figure is not a single-chip result but belongs to the 4-chip RAIDEN X4 module. The per-chip figure is approximately 840 TFLOPS at FP4 precision.
• The headline comparison is made at FP4 (4-bit) precision. At higher precisions, the gap with NVIDIA may narrow or reverse.
• For edge AI buyers, TOPS/W is the deciding factor. EdgeCortix's SAKURA-II targets 60 TOPS with 8 W power consumption. This is a more relevant feature for devices with battery life or thermal constraints.

Critical take

The phrase 'outperforms NVIDIA' is a marketing simplification. A 4-chip module surpassing a single SoC in peak FLOPS says little about real-world efficiency, software maturity, or cost per inference. Additionally, comparing only at FP4 makes the numbers look better than they are. Until independent and sustained workload comparisons at mixed precision (FP8/FP16/INT8) are published, the 1.6x claim should be viewed as a headline in a spec sheet rather than a verified performance gain.