Summary
According to a post on X by @ivanfioravanti, he ran Qwen3.8-Flash with eight simultaneous users to compare a Mac M3 Ultra with DGX Spark systems. Once all eight users were generating, the M3 Ultra reached 191 tokens per second, compared with 140 for two DGX Sparks and 81 for one. However, the Mac processed prompts one at a time, so the eighth user waited about nine minutes for a first word, versus about 71 seconds on one Spark. The post attributes this delay to the inference engine rather than the hardware.
AI-assisted summary based on the listed source.
What happened
Ivan Fioravanti ran Qwen3.8-Flash for eight simultaneous users and found the M3 Ultra fastest once all were generating, at 191 tokens per second versus 140 for two DGX Sparks and 81 for one. But the Mac processed prompts one at a time, so the eighth user waited about nine minutes for a first word, versus about 71...
Why it matters
Teams planning local LLM serving for several users may find that peak generation speed and time to first response point in different directions. The post suggests the software running the model may drive the queuing delay, not only the hardware, so results depend on the inference engine used. Readers should watch for follow-up tests with other engines, since the snippet does not show whether the delay can be fixed.
What this means for you
Hardware and robotics watchers may want to track whether this becomes a product, benchmark, or deployment signal.
Signal Intelligence
Signal Strength 83%
Technical label WATCH
Public Interest 26
Category ROBOTS & HARDWARE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 0
Novelty Interest Score 94
Consequence Score 10
Curiosity Score 16
Shareability Score 45
Why this is here
VQV surfaced this signal because it is recent, relevant to LLM Inference, connected to Scobleizer News Pulse (X).
