Published · Phase 1
NVIDIA DGX Spark LLM Benchmarks
DGX Spark is a capacity-first local AI system, not the fastest box for every model. Choose it when a model or context budget needs more than 32–96 GB of accelerator memory; choose a desktop GPU when the model already fits and low latency matters more.
Editorial review: complete · Updated 2026-08-30
What this hardware is
What this hardware is guidance: The capacity, bandwidth, power, and architecture figures come from NVIDIA; performance examples are configuration-specific public or StackBench rows.
Who should consider it
Who should consider it guidance: DGX Spark suits developers who routinely exceed consumer-GPU VRAM, want a quiet compact system, or need CUDA on a 128 GB shared-memory platform. It is a poor value for 7B–32B models that already fit comfortably on a high-bandwidth desktop GPU.
Current market pricing
Current market pricing guidance: Compare complete-system cost with the price of a 5090 workstation and a 96 GB professional workstation. A launch price or reseller listing is a shopping lead, not a stable market value; capacity per dollar is the useful comparison.
Memory / capacity
Memory / capacity guidance: Its 128 GB coherent unified memory makes 70B-class 4-bit models, gpt-oss-120b MXFP4, Mistral Small 4 NVFP4, and many 100B–200B-class optimized checkpoints realistic. Keep 15–25% free for runtime workspaces, KV cache, the OS, and concurrent requests.
Observed model performance
Observed model performance guidance: Public examples span roughly 14 tok/s for very large DeepSeek V4 configurations, about 30 tok/s for dense 27B Qwen, and 55+ tok/s for gpt-oss-120b under llama.cpp. Treat those as workload anchors, not a single platform score.
Software / runtime support
Software / runtime support guidance: Start with llama.cpp for GGUF models, simple local chat, and flexible quantization. Use vLLM or an NVIDIA NIM/TensorRT-LLM image for an OpenAI-compatible multi-user service, but expect ARM64 image and kernel compatibility to matter.
Strengths
Strengths guidance: The standout advantage is 128 GB of coherent memory in a compact CUDA system, with no PCIe transfer boundary between CPU and GPU allocations.
Weaknesses
Weaknesses guidance: The main constraint is memory bandwidth: 273 GB/s is dramatically below top desktop GDDR7 GPUs, so memory-bound token generation can be slower even when the model fits easily.
Alternatives
Alternatives guidance: Use an RTX 5090 for the best single-user speed/value below 32 GB, RTX PRO 6000 for 96 GB plus much higher memory bandwidth, or two DGX Sparks only when one 128 GB node cannot hold the target configuration.
Where to buy / check price
Where to buy / check price guidance: Compare complete-system cost with the price of a 5090 workstation and a 96 GB professional workstation. A launch price or reseller listing is a shopping lead, not a stable market value; capacity per dollar is the useful comparison. Compare the exact product, memory configuration, warranty, seller, and complete-system cost before ordering.
Affiliate disclosure
Affiliate disclosure guidance: Any referral option is secondary to this recommendation: Buy for capacity and appliance convenience. Do not buy it expecting RTX 5090-class decode speed on models that fit within 32 GB. Merchant availability or commission never changes the technical ranking.
Evidence and limitations
- Performance varies strongly with quantization, prompt length, runtime build, speculative decoding, and concurrency; compare a nearby configuration rather than a brand-wide average.
- Unified memory is not the same as 128 GB of free model memory. Reserve capacity for the OS, runtime, vision encoder, KV cache, and request bursts.
Planner CTA
Compare this recommendation against your model, context, concurrency, latency, and budget in the ComputeSage Planner.
Sources / methodology
Recommendations combine official specifications, public model/runtime documentation, adjacent benchmark observations, and clearly labeled engineering estimates. Review the ComputeSage evidence methodology.