DGX Spark (GB10)
TDP 170W · 273 GB/s bandwidth
The DGX Spark (GB10) has a 170W TDP, 273 GB/s memory bandwidth.
128 GB unified LPDDR5x memory shared by Grace CPU and Blackwell GPU via NVLink-C2C. 273 GB/s bandwidth. 1 PFLOP FP4 sparse. Not a datacenter GPU — personal AI supercomputer. Arm64 architecture.
DGX Spark (GB10) has 0GB of memory. Compare its specs and listed cloud rental prices below. The workload must fit in VRAM with room for runtime memory; a listed price does not confirm stock.
DGX Spark (GB10) Overview
The NVIDIA DGX Spark (GB10) is a personal AI supercomputer pairing a Grace CPU with a Blackwell GPU via NVLink-C2C. Its 128 GB of unified LPDDR5x memory is shared between CPU and GPU at 273 GB/s — enough to run a 70B parameter model at INT4 on a single node.
Unlike datacenter GPUs with discrete HBM VRAM, the DGX Spark uses unified LPDDR5x. This means the full 128 GB is accessible to both CPU and GPU workloads, but bandwidth (273 GB/s) is lower than HBM3e (3,350 GB/s on H100). For memory-bound inference the bandwidth gap matters; for batch-tolerant workloads the large memory pool is the key advantage.
The DGX Spark uses an Arm64 (Grace) CPU architecture. Container images must target linux/arm64 — verify compatibility before provisioning.
GPUwerk offers single-node DGX Spark at $0.79/hr and a linked 2-node cluster (256 GB total across two nodes, not one pool) at $1.79/hr. Availability is low — waitlist applies. Billing is per-minute from prepaid credit; stopped instances are billed at 75%.
Memory
Compute Performance
Hardware Specifications
Chip
Memory
Power
Relative Performance
Relative to highest-spec GPU in database
Limitations
Live Cloud PricingOn-demand hourly rates
Compare DGX Spark (GB10) vs…
Use Case Guidance
LLM Model Size Guidance
Estimates only. Actual capacity depends on context length, KV cache, and framework overhead.
Related Guides
LLM APIs Running on This GPU Class
Providers that serve frontier LLM inference on Blackwell-class hardware.
DGX Spark (GB10) vs Alternatives — Spec Comparison
| Spec | DGX Spark (GB10) this | Quadro M4000 8GB | Quadro P4000 8GB | Quadro P5000 16GB |
|---|---|---|---|---|
| VRAM | 0GB LPDDR5x (unified) | 8GB GDDR5 | 8GB GDDR5 | 16GB GDDR5X▲ |
| Memory Bandwidth | 273 GB/s | 192 GB/s | 243 GB/s | 288 GB/s▲ |
| FP16 TFLOPS | 0 | 2.6 | 5.3 | 8.9▲ |
| BF16 TFLOPS | 0 | 2.6 | 5.3 | 8.9▲ |
| FP8 TFLOPS | — | — | — | — |
| INT8 TOPS | 0 | 10.4 | 21.2 | 35.6▲ |
| TDP | 170W | 120W | 105W | 180W▲ |
| Process Node | TSMC 4NP (GPU) + 4N (CPU)▲ | TSMC 28nm | TSMC 16nm | TSMC 16nm |
| Architecture | GB10 Grace Blackwell | GM204▲ | GP104 | GP104 |
| Release Year | 2025▲ | 2015 | 2017 | 2016 |
| Max model (FP16) | ~0B params | ~4B params | ~4B params | ~8B params▲ |
| Max model (INT4) | ~0B params | ~16B params | ~16B params | ~32B params▲ |
Related GPUs
DGX Spark (GB10) Frequently Asked Questions
What is the DGX Spark (GB10) TDP?
The DGX Spark (GB10) has a TDP (Thermal Design Power) of 170W.
What is the DGX Spark (GB10)'s memory bandwidth?
The DGX Spark (GB10) has a memory bandwidth of 273 GB/s.
DGX Spark (GB10) User Reviews
DGX Spark (GB10) User Reviews
Ready to rent?
Compare DGX Spark (GB10) prices across 130+ providers
Live on-demand & spot rates · monthly cost estimates · availability status