A Data Center on Your Desk: Hands-On Review of the Dell Pro Max GB10 AI Workstation

The Dell Pro Max GB10 AI Workstation is built on the Grace Blackwell architecture, combining a 20-core ARM CPU and a Blackwell GPU in a single package, with 128GB LPDDR5x unified memory and 273GB/s of bandwidth reducing the CPU-GPU bottleneck. Thanks to the unified design based on NVLink-C2C, the focus is squarely on loading and running large models locally as-is.

Initial setup is simpler than expected. Powering it on immediately boots into NVIDIA DGX OS (based on Ubuntu 24.04 LTS), and uname -m confirms aarch64. Docker and the NVIDIA Container Toolkit, along with drivers and CUDA, come pre-configured, so you can launch containers right away and start experimenting.

Connectivity also goes beyond typical workstation levels.

• Wi-Fi 7 and Bluetooth 5.4 support
• Three USB-C 3.2 Gen 2 ports, HDMI 2.1b, and a 10GbE port
• Two 200Gbps ConnectX-7 SmartNICs enabling clustering of up to 2 units, forming a 256GB unified memory configuration

The benchmark was conducted by quantizing Meta-Llama-3.3-70B-Instruct to NVFP4 and serving it with TensorRT-LLM 1.2.0rc6.post3. The 70B model, which originally requires about 140GB at FP16, dropped to roughly 60GB VRAM when serving, recording output throughput of 198.31 tokens/sec, TTFT 661.42ms, ITL 240.41ms, and p99 903ms.

Accuracy loss was limited.

• BF16: MMLU 83.3, GSM8K_COT 95.3, ARC Challenge 93.7
• FP4: MMLU 81.1, GSM8K_COT 92.6, ARC Challenge 93.3

The strengths are 128GB of unified memory, power consumption under 100W, and near-silent operation. On the flip side, there remain ARM aarch64 ecosystem compatibility issues and the structural limitation of not being able to individually upgrade memory or the GPU.

In conclusion, this device is closer to a personal AI development node than a general-purpose workstation. It lets you load 70B-class models as-is without the cloud, making it possible to experiment with RAG and multi-model pipelines right on your desk.