CarbonForge

@carbonforgeai

Using AI to discover more efficient ways to run AI. Optimizing inference across software and hardware for more performance per watt.

Montréal, QC carbonforge.ai Joined October 2026

39 Following 10 Followers 3 Posts

Latest posts 3

CarbonForge @carbonforgeai Oct 23

In offloaded MoE serving, every expert miss is a PCIe copy. Routing tricks trade quality for speed; the cheap win is the eviction rule. A clairvoyant evictor cuts transfers 2.1x with identical outputs. NeurIPS 2026 AXIOM workshop, by @pierrelux

carbonforge.ai/blog/paging-ex…

carbonforge.ai Paging experts: in offloaded MoE serving, the cheap win is eviction, not routing When MoE experts live in CPU memory, every miss is a PCIe copy. What copies cost, why sticky routing trades quality for speed, and Belady's rule for experts.

CarbonForge @carbonforgeai Oct 14

An agent choosing its own experiments can confirm false findings by peeking, testing many claims, or retrying. When nothing is true, peeking alone gives 43% false supports; EvidenceTree keeps it under 2%. NeurIPS 2026 e-values workshop, by @pierrelux

carbonforge.ai/blog/testing-b…

carbonforge.ai Testing by betting: how we keep discovery agents from fooling themselves Adaptive agents can confirm false findings by peeking, by testing many claims, or by retrying. E-values and a harness that enforces their rules fix all three.

CarbonForge @carbonforgeai Oct 14

LLM decode is often memory-bound, so extra SM clock becomes power, not tokens. A serving-aware clock governor on 2xA100 cut GPU power 12 to 13% at matched capacity, and about 30% on saturated decode. Model and engine untouched. By @pierrelux

carbonforge.ai/blog/adaptive-…

carbonforge.ai Adaptive SM clocking for energy-efficient LLM serving How CarbonForge can help lower GPU energy without touching your inference stack.