ThinkStation PGX vs Your Setup | Generated by AI
Let me check that page.This is Lenovo’s ThinkStation PGX — their version of the NVIDIA DGX Spark (formerly “Project DIGITS”). It’s a compact desktop AI development box built around the NVIDIA GB10 Grace Blackwell Superchip.
The relevant specs from the page:
- SoC: NVIDIA GB10 Grace Blackwell Superchip — a 20-core Arm Grace CPU fused to a Blackwell GPU via NVLink-C2C
- Memory: 128GB LPDDR5x unified (CPU+GPU share it, coherent)
- Storage: up to 4TB NVMe M.2
- OS: NVIDIA DGX OS (Ubuntu-based)
- Price: ~$5,079 (showing $4,825 promo), SFF form factor
The pitch is “your own little DGX.” The GB10 delivers roughly 1 petaflop of FP4 (sparse) AI compute, and the 128GB coherent pool is the actual selling point — it lets you load models that won’t fit on a single discrete GPU’s VRAM. Two of them can be NVLink-bridged to run ~405B-class models for inference.
How this maps to your situation, concretely:
Where it wins over your RTX 4070 (12GB): the 4070 caps you hard on VRAM. The PGX’s 128GB unified memory means you can run/fine-tune models in the 70B range (quantized) or do LoRA on much bigger bases without the OOM dance. For your DeepSeek-v4 MoE exploration and nanochat-scale work, the memory ceiling is the thing that actually unblocks you.
Where it loses: FP4-petaflop marketing ≠ training throughput. LPDDR5x has far less bandwidth (~273 GB/s) than the MI300X’s HBM3 (~5.3 TB/s) or even a 4090’s GDDR6X. For raw training speed — your GPT-2 760M-from-scratch runs — it’ll be much slower than the MI300X. This is a “fits big models, runs them slowly” device, optimized for inference, prototyping, and fine-tuning rather than from-scratch pretraining.
The honest comparison for you: at ~$5K, this is roughly 2,500 hours of MI300X dev-cloud time at $2/hr. Given you already have MI300X access for heavy training and a 4070 for local iteration, the PGX’s niche is “local box that can hold a 70B model in memory for agent/inference work without renting.” That’s a real gap in your setup, but it’s a convenience/latency purchase, not a capability you lack. If the goal is training competency that compounds, the MI300X hours buy more learning per dollar. If the goal is a quiet always-on local inference/agent host with huge memory, the PGX fits.
References: