Nvidia’s much-anticipated GB10 Grace Blackwell ARM superchip has surfaced in early benchmark results just weeks before its expected showcase at Computex 2025.
The GB10, designed for compact AI workstations, combines 10 high-performance Cortex-X925 cores (up to 3.9 GHz) and 10 efficiency-focused Cortex-A725 cores with a powerful Blackwell GPU.
This system-in-package (SiP) is aimed at AI professionals and developers looking for a blend of CPU and GPU performance. According to Tom’s Hardware, the GB10 is currently found in Nvidia’s DGX Spark platforms, with prices starting at $3,000.
Nvidia GB10 Early Benchmarks
The leaked Geekbench 6 scores reveal a single-core score of 2,960 and a multi-core score of 10,682.
These results put the GB10’s CPU performance close to Qualcomm’s Snapdragon X Elite (single-core: 2,939) and Apple’s M3 (single-core: 3,076), but it falls behind in multi-core performance. For example, Apple’s M4 and M3 Max chips, as well as AMD’s Ryzen 9 9950X, score significantly higher in multi-core tests, with some exceeding 20,000 points.
The GB10’s real strength lies in AI workloads. Its Blackwell GPU can deliver up to 1 PetaFLOPS of FP4 compute performance, making it ideal for AI model training and inference. The chip supports up to 128 GB of unified LPDDR5X memory with a massive 273 GB/s bandwidth, matching Apple’s M4 Pro in memory capabilities.
However, experts note that these are pre-release results, and the GB10 was tested on a Windows 11 Enterprise Insider Preview system. Some believe the relatively low multi-core score may be due to early software scheduling issues or unoptimized microcode, especially for the Cortex-A725 efficiency cores. As the product matures, improvements are expected.
Nvidia is also rumored to unveil N1 and N1X ARM-based processors-developed in partnership with MediaTek-at Computex 2025, targeting desktops and laptops and potentially expanding their reach beyond AI workstations.