NVIDIA DGX Spark

Consumer GPU · Blackwell · GB10

Summary

The DGX Spark is NVIDIA's compact desktop AI computer, built around the GB10 chip. It pairs a Blackwell GPU with an Arm CPU and 128GB of unified LPDDR5x memory.

The DGX Spark's unified memory is signficantly slower than most modern discrete GPUs' HBM. It also has lower processing power than most data centre GPUs and the top gaming GPUs. See our Rent vs Buy article for the DGX Spark's prefill/decode performance on Qwen3.8-27B and Qwen3.8-Flash-Next.

It is currently only available for short-term rental from peer-to-peer marketplaces such as Vast.ai.

Launched
Q4 2025
VRAM
128 GB
Mem. bandwidth
273 GB/s
On-demand from
Compare prices from 1 providers →

Tech specs

VRAM 128 GB LPDDR5x
Memory bandwidth 273 GB/s
CUDA cores 6,144
Tensor cores 192 (Gen 5)
TDP 140 W
Supported data types
FP64FP32TF32FP16BF16FP8FP6FP4INT8

Cloud rental prices

1 provider · available in 2 countries · lowest price per GPU, per hour

Prices updated

On-demand
$0.60 – $0.67 /GPU/h
Spot
$0.54 /GPU/h
Provider On-demand Reserved Spot
Vast.ai ✓ $0.60 — $0.54
Compare all NVIDIA DGX Spark cloud providers & configurations →

Models that fit in VRAM

Open-weights models that fit in VRAM. Estimated using 🤗 accelerate, plus approximation for up to 8K context.

Peak theoretical performance

FP4 Tensor Core 500 TFLOPS
FP6 Tensor Core 425 TFLOPS
FP8 Tensor Core 425 TFLOPS
INT8 Tensor Core 425 TOPS
BF16 Tensor Core 213 TFLOPS
FP16 Tensor Core 213 TFLOPS
TF32 Tensor Core 53 TFLOPS
FP32 31 TFLOPS
FP64 0.5 TFLOPS

Performance figures assume no sparsity; in cases where only sparse performance figures are published by the manufacturer, these are halved to give approximate dense performance.

Resources

Detailed documentation from the manufacturer.