GB10 is Not a Baby B200

GB10, also known as DGX Spark, is a powerful mini-PC with 128GB of unified memory and a Blackwell GPU. You may have heard the pitch: write and verify code locally, then deploy to big B200 GPUs in the cloud.

Not all Blackwell GPUs are created equal, though. B200 and co. are sm100. RTX 5090 and lower consumer GPUs are sm120. GB10 is sm121.

Bigger numbers are not necessarily better here. Both sm120 and sm121 lack the crucial tcgen05 instructions. There is no portability story for low-level GPU kernel optimizations.

Dependency Considerations

GB10 requires CUDA 13.0 or newer, which limits how old a PyTorch version you can use. This also creates problems for projects with locked dependency versions. For comparison, RTX 50x0 requires CUDA 12.8 or newer — a bit more flexible.

GB10 uses an aarch64 CPU. I found that certain PyTorch/CUDA version combinations lack wheel packages. Another major project that lacks aarch64 wheels is vLLM (though nightly aarch64 builds are available via vLLM’s package index).

Inference Machine?

With 128GB of unified memory, GB10 can run fairly large local AI models. However, the 273 GB/s memory bandwidth is a bottleneck. For comparison, RTX 3090 has 936 GB/s and RTX 5090 has 1792 GB/s.

The Good Parts

The small form factor and low power consumption mean that GB10 does not pose a living space challenge, unlike a custom-built PC.

The nearly 128GB of VRAM allows training or inference jobs that are infeasible on one or two 3090s.

GB10 includes a 200 Gbps ConnectX-7 NIC with two QSFP ports, though buying multiple GB10s and the networking equipment to link them up can get expensive.

For completeness, GB10 also has a 10GbE Ethernet port and WiFi 7.

Choose Between System Builders

GB10s are available from different builders: Asus, Dell, Gigabyte, HP, Lenovo, MSI, PNY, etc.

I bought an Asus GB10 with 1TB SSD for $3k, which was the cheapest option at the time. The same SKU has since increased to $3.5k.

Features should be exactly the same across builders. Check online reviews if you are concerned about thermals or SSD quality.

Note that the SSD is M.2 2242, not the more common M.2 2280. Getting a larger SSD from the start might not be a bad idea.

Recap

GB10 is a fascinating machine with an interesting set of trade-offs.

It is not a perfect staging ground for B200 or other powerful cloud GPUs due to architectural differences and software ecosystem immaturity.

However, if your primary need is VRAM capacity, its 128GB of unified memory is a game-changer. For those who need to fit massive models locally, GB10 is a compelling if slightly quirky option.