Sweat Capital: Ripping Out TEI for a 27x Embedding Speedup with Go and Custom CUDA
Why I ripped out Hugging Face TEI and built a Go + custom-CUDA embedding engine that runs fast on sub-H100 hardware.
READ ARTICLE
ENGINEERING // FIELD NOTES
7 DISPATCHES
From the Kernel.
Technical deep-dives on GPU memory, sorting, and deterministic compute.
-
GPU Compute -
Model Ingot Poured: Our MTEB Submission Model
Voxell's MTEB(eng, v2) submission: architecture, training methodology, contamination defense, and API access.
READ ARTICLE -
Retrieval Why Your RAG Pipeline Doesn't Learn
Your retrieval system reruns the same expensive searches and never tracks which results helped. The retrieval N+1, and how closing the loop changes things.
READ ARTICLE -
Caching The Cache Feedback Gap: Why Your Prefetcher Doesn't Learn
Your retrieval cache predicts what to preload and never learns if it was right. The open-loop problem, and what closing the feedback loop unlocks.
READ ARTICLE -
Determinism Why Your Vector Search Lies
Same query, different results. The floating-point problem that quietly breaks reproducible embeddings, and what bit-exact determinism takes in production.
// THE PROBLEM WITH PARALLEL REDUCTIONREAD ARTICLETHREAD A: (a + b) + c = 0.30000000000000004THREAD B: a + (b + c) = 0.3// SAME INPUTS. DIFFERENT OUTPUTS. NONDETERMINISM. -
Memory Architecture Why Vector Search is Memory Abuse (And How to Fix It)
Your vector index burns GPU bandwidth chasing pointers through random memory, and a brute-force linear scan often beats HNSW on the GPU.
READ ARTICLERANDOM ACCESS PATTERN Cache misses, stalled coresSEQUENTIAL ACCESS PATTERN Coalesced reads, saturated bus -
GPU Compute Starved Cores: Listening to the Silence Between Epochs
Your H100 sits idle 40% of the time, starved by a data supply chain that can't keep up. Here is the physics, and why the real fix is architectural.
READ ARTICLE