Nvidia GB10 Superchip: Early Glimpses of Performance Emerge
Some early glimpses of Nvidia's GB10 Superchip abilities have started to emerge, this time as Geekbench 6 scores. The chip is designed for compact AI workstations, and while its GPU reserved for AI promises much in terms of power, the initial whispers of its CPU compute performance are generating some buzz. Remember that these are pre-release figures, with the system running on a Windows 11 Enterprise Insider Preview.
Under the Hood of the GB10
GB10 Superchip is a high-end system-in-package (SiP) product. It features at the heart a blend of 10 high-performance Arm Cortex-X925 cores, stated to clock frequencies of up to 3.90 GHz, and 10 efficiency-focused Cortex-A725 cores. The CPU is supported by a high-power Blackwell GPU, which Nvidia stated can support 1 PetaFLOPS of FP4 compute throughput – a crucial measure for AI computing.
For the purpose of providing this processing capacity, the SiP has a 256-bit memory interface that can support 128 GB of unified LPDDR5X memory. The configuration provides a large memory bandwidth of up to 273 GB/s, which comes in handy when processing large AI models and datasets.
Early CPU Performance Insights
Looking at the CPU, the Arm Cortex-X925 cores, even in their pre-release and 3.90 GHz clock speed configurations, demonstrate decent single-threaded capability. The single-thread score on Geekbench 6 measured at 2,960. For AI workstations, good single-thread performance can be very useful for some areas of workload handling and responsiveness.
Nvidia GB10 Initial Geekbench 6 Scores
| Measure | Score |
|---|---|
| Single-Core | 2960 |
| Multi-Core | 10682 |
Reminder: These are preliminary benchmark results.
With the 20-core CPU configuration (10 performance + 10 efficiency cores) in mind, high multi-thread scores would be expected. The GB10's first multi-core score on Geekbench 6 was 10,682 points, though. This is a number to track with the product as it matures. There are a few things that could be going on here. The Cortex-A725 efficiency cores are not doing their full part in the score yet, possibly due to pre-release Windows 11 environment scheduling issues or still-emerging microcode. Later revisions should unleash additional multi-threaded performance.
The AI Workstation Context
It is important to keep in mind the GB10's role. In tiny AI workstations, the CPU is often secondary to the super-parallel GPU. Tasks like scheduling, data preparation, and light pre-processing may fall to the CPU, but Blackwell AI computation is left to the heavy-lifting of the Blackwell GPU. In this kind of setup, CPU architecture might prefer a good balance between general-purpose performance but high optimization in power consumption and reduced die area, especially under the tight thermal budget of dense systems like Nvidia's own design for DGX Spark.
Nvidia has signaled its intentions in the client PC processor space, and the GB10 Superchip for AI workstations is an important part of that strategy. Although these preliminary Geekbench 6 results are a news grabber, Geekbench is a synthetic benchmark and is known to be so. The true measure of the GB10 will be its real-world performance in AI workloads and how well its CPU and GPU elements work together within its designed power and thermal budgets.
As development continues and the GB10 nears its official release, we should have a clearer idea of what it can do. In the meantime, focus is on its powerful Blackwell GPU for AI, with CPU performance being an interesting thing to observe.
