ComfyUI Workflows Crawl on 128GB DGX Spark vs RTX 4090
jungseungoh97 · reddit · 2026-08-12
A developer shared a severe performance regression after upgrading hardware. A ComfyUI workflow running smoothly on an RTX 4090 (24GB) crawled on a 128GB unified memory DGX Spark (Blackwell architecture).
Using a default template to run the MiniMax H3 INT8 SafeTensor model for a 60s 2MP video, the DGX Spark barely made progress. The user acknowledged the system's unique architecture and Linux requirements, suspecting a fundamental misconfiguration in their setup and asking the community for troubleshooting advice.
More from Infra
- Foxconn: Vera Rubin Racks Mass Production to Double AI Server Shipments This Year — firstadopter · 2026-08-12
- Benchmark: ComfyUI on Linux is 30% Faster than Windows with Dual GPUs — Icy_Restaurant_8900 · 2026-08-12
- Australian AI Data Center Firmus Raises $2B from Nvidia — Beth_Kindig · 2026-08-12
- AMD Execs Discuss CPU Architectural Approaches for the Agentic Era — BenBajarin · 2026-08-12
- Local Video Generation on RTX 3060 32GB: Performance and Workflow Discussion — Nakidka · 2026-08-12
- Quantized Krea-2-Turbo Runs on 6GB VRAM, Humming Kernel Hits 1.6x Speedup — ali_byteshape · 2026-08-12