NVIDIA DGX Spark
DGX Spark is a desk-side box built around NVIDIA's GB10 Grace Blackwell superchip, 128 GB of unified memory, and the full CUDA stack, aimed at a single developer prototyping and fine-tuning models locally before moving to real GPU cluster capacity. It is not a cluster node: its 273 GB/s memory bandwidth is well behind a datacenter GPU, and while two units can be linked over ConnectX-7 to handle models up to 405B parameters, that is a two-box desk setup, not a scalable rack architecture. Buy it to develop against CUDA and iterate on model code without renting cloud GPU time; don't buy it expecting DGX B200-class throughput.
- 273 GB/s memory bandwidth is the real ceiling, not the 128 GB capacity. Prefill-heavy and fine-tuning workloads are compute-bound and look great; decode-heavy inference is memory-bound and noticeably slower than the raw PFLOP figure suggests.
- It is not a cluster building block. The dual QSFP ConnectX-7 ports look like 400 Gbps of fabric, but PCIe lane sharing caps actual throughput at 200 Gbps aggregate, and NVIDIA's own clustering story tops out at two linked units.
- The 2242 M.2 storage form factor has been reported to lock up under sustained heavy I/O in third-party testing; treat local SSD as working storage, not the system of record for large datasets.
- For raw token throughput on models that fit in 32 GB, a workstation RTX 5090 is faster and cheaper. DGX Spark's reason to exist is the 128 GB unified memory footprint for models a 5090 cannot hold, not raw speed.
- Buy this for CUDA-native local development and fine-tuning iteration, not as a production inference box. Budget cloud GPU time for anything latency-sensitive or high-concurrency.
How much does DGX Spark cost?
The 4 TB Founders Edition has listed between $3,999 and $4,699, reported. NVIDIA raised its own-store price to $4,699 in early 2026 citing memory supply constraints, while retailers like Micro Center have continued to sell at $3,999.99.
What is DGX Spark used for?
Local AI development: prototyping, full-CUDA-stack coding, and fine-tuning models up to roughly 70B parameters or running inference on models up to roughly 200B parameters, thanks to its 128 GB unified memory. It targets a single developer's desk, not production inference serving.
Can I cluster two DGX Spark units together?
Yes. Two units connected over their ConnectX-7 networking can handle inference on models up to roughly 405B parameters. This is a two-box desk setup for larger model capacity, not a scalable multi-node cluster architecture.
Is DGX Spark good for inference or training?
It handles both, but memory bandwidth (273 GB/s) is the bottleneck for decode-heavy inference. It is strongest for full-parameter fine-tuning and prefill-heavy workloads where compute, not memory bandwidth, is the limiting factor.