Blog · Deep dives · Unum

Latest from Unum Changing Our Development Practices with Time

A decade in, AI agents made writing code cheap and keeping it coherent expensive. How that changed our Git branches and releases, moved eight repositories into lockstep, turned pre-commit hooks into design guardrails, made model checking routine, and what we still have not figured out

Read post All notes

2026

  1. Nov 9 Changing Our Development Practices with Time A decade in, AI agents made writing code cheap and keeping it coherent expensive. How that changed our Git branches and releases, moved eight repositories into lockstep, turned pre-commit hooks into design guardrails, made model checking routine, and what we still have not figured out
  2. Apr 20 FP8 Search & KV-Caching in USearch GPUs moved to FP8. CPUs didn't. 200'000 lines of SIMD kernels later, USearch bridges both worlds — native Float8 Vector Search and KV Caching for the upcoming Arm-based NVIDIA Vera racks, Intel Xeons, open-source RISC-V chips, IBM Mainframes, Chinese LoongArch, and even WebAssembly in the browser — for higher throughput at 99% recall and half the memory cost

2023

  1. Nov 7 USearch: 10x - 100x Faster than Meta's FAISS Scaling Retrieval Augmented Generation with the USearch vector search engine on Intel Sapphire Rapids CPUs on Gcore, outperforming Meta's FAISS library by 10x - 100x on a scale of 100M to 1B vectors.
  2. Aug 17 UForm v2: now in 21 languages Releasing UForm v2 with support for 21 languages, delivering state-of-the-art efficiency in AI-based text-to-image search and multimodal embeddings.
  3. Jun 8 Making the Edge Efficient with Gcore and Intel Partnering with Gcore and Intel to deploy the world most efficient data-processing infrastructure at the edge, powered by Unum assembly-optimized engines.
  4. Feb 20 Beating OpenAI CLIP with 100x less data and compute Demonstrating how to train state-of-the-art Vision-Language transformers for semantic search using 100x less data and compute than OpenAI CLIP.

2022

  1. Sep 20 Billion Rows in Pandas? CuDF, Modin, Arrow, Spark Evaluating in-memory tabular processing frameworks including CuDF, Modin, Arrow, and Spark on billion-row taxi ride datasets to find the fastest Pandas alternative.
  2. Sep 13 Beating RocksDB by up to 7x in almost every workload Pushing the limits of UDisk and RocksDB on massive 10TB datasets with 64 threads and 1TB of RAM, achieving up to 7x performance gains in almost every workload.
  3. Mar 22 Unbranding and Extending the Yahoo Cloud Serving Benchmark Introducing the Unum Cloud Serving Benchmark (UCSB), a modernized and extended version of YCSB designed to rigorously evaluate the performance of next-generation storage engines.
  4. Jan 26 50x Faster NumPy on GPUs with CuPy Accelerating NumPy workflows by 50x using CuPy on Nvidia RTX 3090, benchmarking BLAS operations against CPU-based implementations on AMD Threadripper.
  5. Jan 25 CPU vs GPU Benchmarks Series Comparing the number-crunching capabilities of AMD Threadripper PRO 3995WX and Nvidia RTX 3090 across matrix multiplication, convolution, and other compute-intensive tasks.

2021

  1. Dec 31 2021 in Database Startups: Gold Rush! A retrospective on the database industry in 2021, analyzing the technical architecture, funding rounds, and market promises of the hottest new database startups.
  2. Nov 25 UDisk Crushing Benchmarks: YCSB, pt. 2 Benchmarking UDisk against RocksDB, LevelDB, and WiredTiger on 1TB datasets, demonstrating superior performance in read-heavy and write-heavy workloads.
  3. Oct 27 Yahoo Cloud Serving Benchmark for Key-Value Stores A deep dive into the Yahoo Cloud Serving Benchmark to evaluate NoSQL database performance, identifying the strengths and weaknesses of RocksDB, LevelDB, and WiredTiger.
  4. Jan 5 Our Recipe for Fast Computational Storage Exploring the limits of modern hardware through Assembly optimizations, CUDA kernels, and algorithmic improvements to eliminate IO and compute bottlenecks in database engines.
  5. Jan 3 Text Indexing in ElasticSearch, MongoDB and UDisk Comparing indexing performance of ElasticSearch, MongoDB, and UDisk on a 20GB text dataset, revealing the massive overhead of Java-based search engines.
  6. Jan 2 Why Build a Custom DBMS? Analyzing the widening gap between modern NVMe SSD throughput and legacy database software performance, and why we decided to build a new storage engine from scratch.

2020

  1. Nov 12 Networks Processing Performance Report Benchmarking graph traversal performance on persistent storage against NetworkX, evaluating random access patterns on Patent Citations, Gene Regulatory Networks, and Human Brain connectomes.