REDDIT · REDDIT// 4h agoINFRASTRUCTURE

DGX Spark users hit thermal throttling

A LocalLLaMA user reports DGX Spark temperatures reaching 90-95°C during sustained local inference, with token/sec dropping as thermal throttling kicks in. The thread is a call for real-world cooling fixes around NVIDIA’s compact GB10-based desktop AI system.

// ANALYSIS

DGX Spark’s pitch is desktop-class local AI infrastructure, but sustained thermals are becoming part of the product story, not just a user setup issue.

–NVIDIA positions Spark for local inference, fine-tuning, and prototyping with up to 1 PFLOP FP4 performance and 128GB unified memory
–Official docs specify a compact integrated cooling design and a 5-30°C operating environment, which leaves little margin for warm rooms or long saturation workloads
–Community reports around power caps, high temperatures, and firmware behavior suggest buyers should benchmark sustained token/sec, not just peak specs
–External fans and cooling cages may help stability, but hardware mods risk support and warranty issues unless NVIDIA or OEM partners provide validated options

// TAGS

dgx-sparknvidiagpuinferenceself-hostededge-ai

DISCOVERED

4h ago

2026-04-22

PUBLISHED

5h ago

2026-04-22

RELEVANCE

7/ 10

AUTHOR

recepayddogdu