Top all-time · #6491
Unum
Company blog · unum.cloud/blog · on HN as ashvardaniandatasystemsai
Unum is the company blog at unum.cloud, writing about data, systems and ai. Since 2021, 12 different Hacker News readers have linked it 33 times, most often “Beating OpenAI CLIP with 100x less data and compute”. No feed was found, so read it on the site itself.
Too few sampled comments (5) to say how often HN recommends it.?
Why HN links it
HN readers link Unum's blog and project pages for its systems work: UStore/UDisk storage, FastAPI speedups with simdjson and io_uring, and a multimodal model compared against OpenAI CLIP.
Summarised by Claude from a sample of 5 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
Unum builds the Search, Storage, and Networking engines powering the modern data stack. Deployed on 100M+ devices, our technology serves DBMS unicorns and frontier AI labs.
Most-linked posts?
- Beating OpenAI CLIP with 100x less data and compute4 people · 4 mentions · first linked 2023-02-24
- Apple to Apple Comparison: M1 Max vs. Intel2 people · 2 mentions · first linked 2021-12-21
What Hacker News says?
USearch author here :) The engine supports arbitrary predicates for C, C++, and Rust users. In higher level languages it’s hard to combine callbacks and concurrent state management. In terms of scalability and efficiency, the only tool I’ve seen coming close is Nvidia’s cuVS if you have GPUs available. FAISS HNSW implementation can easily be 10x slower and most commercial & venture-backed alternatives are even slowe…
Comment linking /blog/2023-11-07-scaling-vector-search-with-intel
on “The best way to use text embeddings portably is with Parquet and Polars”Submitted Apple to Apple Comparison: M1 Max vs. Intel as a story · 243 points
HN discussionIts configured with the fastest memory possible on the Threadripper. The exact specs and RAM stick models match the ones mentioned in the preceding article, I have linked below. We didnt include those, as we don’t know the exact specs of RAM modules in MacBooks. https://unum.cloud/post/2021-11-25-ycsb/#our-toys
Comment linking /post/2021-11-25-ycsb/
on “Apple to Apple Comparison: M1 Max vs. Intel”Submitted Failing to reach DDR4 bandwidth as a story · 101 points
HN discussionLiterally anything. There is so much to do better. Faster I/O, kernel bypass and async filesystem, new persistent data-structures, alternative lock-free concurrency resolution schemes… Disclaimer: I am highly biased, as I am funding/researching/developing a DBMS myself. Out of necessity though, as we constantly hit bottlenecks in the persistent I/O layer. We are not selling or offering anything, but will soon share…
Comment linking /post/2021-12-31-dbms-startups/
on “Timescale raises $110M Series C”Submitted 2021 in Database Startups: Gold Rush as a story · 86 points
HN discussionI wish I'd had a short answer :) For years, I've had a hope to build it in the form of an open-core project: open-source SotA solutions for Storage, Compute, and AI Modeling built bottom up. You can imagine the financial & time burden of building something like that with all the weird optimizations and coding practices listed above. A few years in, with millions spent out of my pocket, without any venture support or…
Comment linking /blog/2023-11-07-scaling-vector-search-with-intel
on “Less Slow C++”Submitted Server Hardware Super-Cycle 2022 as a story · 81 points
HN discussionOh MSVC, bless. Mingw setup might be a pain, but the dividends accrue over time. -- This was a good reminder that I need to pay more attention to Unum's projects. I noticed this older blog article, https://www.unum.cloud/blog/2021-01-03-texts , and that brings up some questions. First, in 2025, is UStore a wholesale replacement for UDisk or are the two complementary? Second, what is the current Unum approach for rep…
Comment linking /blog/2021-01-03-texts
on “Less Slow C++”Submitted Beating OpenAI CLIP with 100x less data and compute as a story · 20 points
HN discussionFive years ago I was absolutely frustrated with the state of Graph databases and libraries and tried putting several non-Graph DBMSs behind a NetworkX-like Python interface < https://github.com/unum-cloud/NetworkXum >. When benchmarked, Neo4J crashed on every graph I’ve tried < https://www.unum.cloud/blog/2020-11-12-graphs >, making SQLite and Postgres much more viable options even for network-processing workloads.…
Comment linking /blog/2020-11-12-graphs
on “Postgres as a Graph Database: (Ab)Using PgRouting”Submitted Unum Rocks: 10 TBs, 64 Cores and CUDA as a story · 12 points
HN discussionIf I remember correctly, for dot product it uses BLAS libraries and for Cosine distance and L2 have very minimal SIMD support. In both cases SimSIMD should be faster for distance computations. Multiply it by a more efficient HNSW implementation and you might be looking at a 100x performance difference on a scale of a billion vectors: https://www.unum.cloud/blog/2023-11-07-scaling-vector-search...
Comment linking /blog/2023-11-07-scaling-vector-search-with-intel
on “SimSIMD: Hardware-accelerated similarity metrics and distance functions”Submitted Apple to Apple RAM Comparison: M1 Max vs. Intel MacBooks vs. $50K Server as a story · 12 points
HN discussionUSearch author here :) Not sure if it's fair to compare USearch and pgvector. One is an efficient indexing structure, the other is more like a pure database plugin. Not that they can't be used in a similar fashion. If you are looking for pure indexing benchmarks, you might be interested in USearch vs FAISS HNSW implementation [1]. We run them ourselves (and a couple of other tech companies), so take them with a grai…
Comment linking /blog/2023-11-07-scaling-vector-search-with-intel
on “90x Faster Than Pgvector – Lantern's HNSW Index Creation Time”Submitted UCSB: Extending the Ultimate Yahoo NoSQL Benchmark as a story · 11 points
HN discussioncuDF is the most impressive DataFrame implementation I've seen and have been recommending for years. The API is exceptionally close to Pandas (just a couple of different function arguments here and there), much more so than PyArrow or Modin. Throughput and energy efficiency were often 10x that of PyArrow running on a comparable TDP SotA CPU 2 years ago [1]. [1]: https://www.unum.cloud/blog/2022-09-20-pandas
Comment linking /blog/2022-09-20-pandas
on “cuDF – GPU DataFrame Library”Submitted 10x Faster Than Meta's Faiss as a story · 8 points
HN discussionThere are also some other blank pages like https://www.unum.cloud/ukv/ and https://www.unum.cloud/ukv/details
Comment linking /ukv/details
on “Show HN: Up to 100x Faster FastAPI with simdjson and io_uring on Linux 5.19”Submitted UKV: Open Binary Interface for NoSQL Database Management as a story · 8 points
HN discussionhttps://www.unum.cloud/ujrpc Is this supposed to be just a blank page, which is what I see?
Comment linking /ujrpc
on “Show HN: Up to 100x Faster FastAPI with simdjson and io_uring on Linux 5.19”Submitted You don't need a thousand GPUs to train a Vision-Language Transformer for Search as a story · 7 points
HN discussionIt may just be that OpenAI corrected its 2020-era understanding of model size using 2022 insights from DeepMind.. https://towardsdatascience.com/a-new-ai-trend-chinchilla-70b... Seems model sizing, compression and quantization are still an art form, see also https://www.unum.cloud/blog/2023-02-20-efficient-multimodali...
Comment linking /blog/2023-02-20-efficient-multimodality
on “Introducing ChatGPT and Whisper APIs”Submitted UnumDB Crushing Benchmarks: YCSB, pt. 2 as a story · 7 points
HN discussion- nl · 2023-02-28
This looks interesting for image retrieval. I don't love the way their tables[1] report performance though. My understanding is that the "Dataset" column in the table represents the size of the training dataset, not the size of the dataset they are evaluating on. Note that this undersells their performance though, so it isn't like they are trying to hide something here! Also I'd love to see someone do a similar benc…
Comment linking /blog/2023-02-20-efficient-multimodality
on “Beating OpenAI CLIP with 100x less data and compute” Submitted UForm v2: Most Efficient AI-based Text-to-Image Search, now in 21 languages as a story · 6 points
HN discussion8K QPS is probably quite trivial on their setup and a 10M dataset. I rarely use comparably small instances & datasets in my benchmarks, but on 100M-1B datasets on a larger dual-socket server, 100K QPS was easily achievable in 2023: https://www.unum.cloud/blog/2023-11-07-scaling-vector-search... ;) Typically, the recipe is to keep the hot parts of the data structure in SRAM in CPU caches and a lot of SIMD. At the tim…
Comment linking /blog/2023-11-07-scaling-vector-search-with-intel
on “zvec: embedded vector database”Submitted High-Precision Software Benchmarking and Tricks for Beginners as a story · 4 points
HN discussionI am really stunned by this story. It made me check the MemGraph benchmarks section. Don't get me wrong, it may be 10-100x faster than Neo4J in even the most basic operations. Moreover, given the quality of Neo4J, it is hard not to be that much quicker. Even Postgres and MySQL are better at storing graphs than Neo4J. --- Disclosure: I have worked on Graph Algorithms, Graph Databases, and Database Engines for years,…
Comment linking /post/2022-03-22-ucsb
on “Bullshit Graph Database Performance Benchmarks”Submitted Best CPU vs. GPU: 879 GB/S Reductions in C++ and CUDA as a story · 4 points
HN discussion
