GPU adds three new notable observations · V2

GPU Behavior Re-Observation

Recompute the observed value (Fit × Efficiency) scores for the Creative Observations Behavior Map — but this time with GPU-driven throughput instead of CPU throughput. Three models that were merely "good" on CPU became notable Swarm Workers on a single GTX 1070, exceeding the conventional 10.0 observed-value ceiling.
new GPU notables
3
granite4:1b GPU OpValue
12.6
granite4:1b CPU → GPU
+4.9

Where the speedup creates a dimension transition

For each model: CPU-side observed value (as observed in Creative Observations), GPU-side recomputed observed value, the delta, and the resulting dimension classification. A model exits the conventional ceiling when its observed value passes 10.0.
CPU OpValue
llama3.2:1b
8.4
granite4:1b
7.7
cogito:latest
7.2
dolphin-mistral:latest
6.9
smollm2:360m
5.3

5 entries · ordered by CPU OpValue as observed in this run · not a ranking

Observed dimension transitions

granite4:1b · Notable

CPU 7.7 → GPU 12.6

IBM's 1-billion-parameter Granite 4. On CPU it was a strong worker. On GPU at ~120 tok/s it crossed into the notable band. For latency-sensitive swarm patterns, this is an observed candidate.

dolphin-mistral · Notable

CPU 6.9 → GPU 11.5

Mistral 7B Dolphin variant. The bigger jump happens at single-slot GPU throughput. Observed for code-review and synthesis with mild-thinking models.

cogito:latest · Notable

CPU 7.2 → GPU 12.3

The DeepCogito 8B (or whichever tag :latest pointed to that day). Largest single CPU → GPU OpValue lift in this set (+5.1).

Publication note

Public summary updated

The Observatory's public Behavior Map widget records notable badges. Three new notable entries are recorded on the public summary.

Models

Families

Companies

Related observations

AI-generated & AI-scored. Every observation on this site is produced and judged by AI, and AI makes mistakes. Treat each result as a recorded observation under a declared protocol, as of its date. Method and marker legend