NVIDIA · Grace Blackwell · Layer 2
NVIDIA DGX Spark
Developers who want a coherent 128 GB memory space on a desk, at wall-socket power.
buy You can purchase this outright.
Specifications
- Memory
- 128 GB unified LPDDR5X
- Bandwidth
- 273 GB/s
- Compute
- GB10 Grace Blackwell superchip
- Power
- Roughly 170 W
- Interconnect
- ConnectX for pairing two units
- Software
- CUDA
- Form factor
- desktop
- Workload
- inference
Verified 2026-09-06. Fields reading “not published” are exactly that — we do not estimate a figure a vendor withholds.
The catch
273 GB/s is an order of magnitude below HBM. It loads big models; it does not serve them fast. Judge it as a development machine, not a server.
The economics
The cheapest legitimate route to running a 70B-class model locally with no cloud bill and no data leaving the building.
Cost per million tokens is the metric that decides most real purchases, and there is no neutral benchmark for it — every published figure comes from a company selling one side of the comparison. Read all of them, ours included, as claims rather than measurements.
The power question underneath this
A rack of frontier accelerators draws 120–200 kW against a 2026 average of about 27 kW, and the US grid interconnection queue exceeds 2,600 GW with waits approaching five years. Whether you can energise this part is now a harder question than whether you can buy it.
Layer 1 — Energy · The interconnection queue · Rack power density · Tokens per watt
Compared against
- NVIDIA RTX PRO 6000 Blackwell vs NVIDIA DGX Spark — Both buyable, both local. Bandwidth against unified memory capacity.Compare with NVIDIA RTX PRO 6000 Blackwell
Related silicon
- NVIDIA RTX PRO 6000 Blackwell — 96 GB GDDR7, buy
- NVIDIA B200 — 192 GB HBM3E, rent
- NVIDIA GB200 NVL72 — 13.5 TB HBM3E across the rack, rent
- NVIDIA GB300 NVL72 — Higher HBM3E capacity than GB200; per-rack figure not consistently published, rent
- NVIDIA H100 — 80 GB HBM3, rent
- NVIDIA H200 — 141 GB HBM3E, rent