Hacker News1d ago

NumKong: 2'000 Mixed Precision Kernels for All

https://ashvardanian.com/posts/numkong/

#HackerNews #NumKong #Mixed #Precision #Kernels #AI #Performance #Optimization

NumKong: 2'000 Mixed Precision Kernels For All ๐Ÿฆ

Over 2'000 SIMD kernels for mixed-precision BLAS-like numerics across 7 languages โ€” from Float6 to Float118, on RISC-V, Intel AMX, and Apple SME, in 5 MB.

Ash's Blog
Show threadMidgePhoto3d ago

@epistatacadam @ChrisMayLA6
On the first clause, yes (but it measures what it measures).
On the second, it is a basic principle of #measurement that #accuracy and #precision should each be known and reckoned with, but that large departures from perfection in either or both do not prevent useful measurements being made.

(When the fiction is reduced the instrument needs recalibrating.)

HGPU group3d ago

True 4-Bit Quantized Convolutional Neural Network Training on CPU: Achieving Full-Precision Parity

#Precision #CNN #Package

https://hgpu.org/?p=30680

True 4-Bit Quantized Convolutional Neural Network Training on CPU: Achieving Full-Precision Parity

Low-precision neural network training has emerged as a promising direction for reducing computational costs and democratizing access to deep learning research. However, existing 4-bit quantization โ€ฆ

hgpu.org
N-gated Hacker News4d ago
Iranian citizens are now Yelp reviewers for air strikes, apparently giving five-star ratings to the IDF's #precision โœˆ๏ธ๐Ÿคฏ. Who needs spy satellites when Tehran residents are dialing in coordinates like a pizza order? ๐Ÿคทโ€โ™‚๏ธ๐Ÿ“ž
https://www.iranintl.com/en/202603179685 #IranianCitizens #YelpReviews #IDF #AirStrikes #Warfare #Humor #HackerNews #ngated
Israeli official says tip from Tehran residents helped enable Larijani strike

Israeli official says tip from Tehran residents helped enable Larijani strike

Iran International
๐Ÿ’ง๐ŸŒ Greg Cocks5d ago
๐ŸŒŸ ๐ˆ๐ง๐ญ๐ซ๐จ๐๐ฎ๐œ๐ข๐ง๐  [Google Researchโ€™s] ๐†๐ซ๐จ๐ฎ๐ง๐๐ฌ๐จ๐ฎ๐ซ๐œ๐ž - ๐€๐ง ๐จ๐ฉ๐ž๐ง ๐ฌ๐จ๐ฎ๐ซ๐œ๐ž ๐๐š๐ญ๐š๐ฌ๐ž๐ญ ๐จ๐Ÿ ๐ก๐ข๐ฌ๐ญ๐จ๐ซ๐ข๐œ ๐Ÿ๐ฅ๐จ๐จ๐ ๐ž๐ฏ๐ž๐ง๐ญ๐ฌ ๐Ÿ๐ซ๐จ๐ฆ ๐ง๐ž๐ฐ๐ฌ ๐š๐ซ๐ญ๐ข๐œ๐ฅ๐ž๐ฌ.
--
https://doi.org/10.31223/X5RR2K / https://eartharxiv.org/repository/view/12083/ <-- shared paper
--
https://zenodo.org/records/18647054 <-- shared link to associated dataset
--
https://sites.research.google/gr/floodforecasting/ <-- shared link to Google Research flood forecasting effort entry page
--
#GoogleResearch #Google #Gemini #AI #ClimateTech #MachineLearning #DataScience #FloodForecasting #Sustainability #TechForGood #aggregation #curated #newsarticles #news #media #article #harvesting #reports #reporting #global #world #historic #naturalhazards #naturaldisaster #floods #flooding #flashflood #water #hydrology #extremeweather #climatechange #GDACS #Groundsource #GIS #spatial #mapping #spatialanalysis #spatiotemporal #geographic #openaccess #openscience #opendata #floodevents #LLM #gemini #largelanguagemodel #deeplearning #AI #precision #metrics #historicresource #model #modeling #forecasting #opensource
WIST Quotations5d ago

A quotation from Hyman Rickover

Nature is not as forgiving as Christ.

Hyman Rickover (1900-1986) Polish-American naval engineer, admiral [b. Chaim Gdala Rykower]
(Attributed)

More about this quote: wist.info/rickover-hyman/6585/

#quote #quotes #quotation #qotd #hymanrickover #engineering #forgiveness #JesusChrist #nature #precision #tolerance #marginoferror

HGPU group6d ago

Diagnosing FP4 inference: a layer-wise and block-wise sensitivity analysis of NVFP4 and MXFP4

#LLM #FP4 #NVFP4 #MXFP4 #Precision #AMD #NVIDIA

https://hgpu.org/?p=30661

Diagnosing FP4 inference: a layer-wise and block-wise sensitivity analysis of NVFP4 and MXFP4

Quantization addresses the high resource demand for large language models (LLMs) by alleviating memory pressure and bandwidth congestion and providing significantly scaled compute power with a toleโ€ฆ

hgpu.org
Hacker NewsMar 11

Lego's 0.002 mm Specification and Its Implications for Manufacturing

https://www.thewave.engineer/articles.html/productivity/legos-0002mm-specification-and-its-implications-for-manufacturing-r120/

#HackerNews #Lego #Manufacturing #3DPrinting #Precision #Engineering #Innovation

HGPU groupMar 8

Practical FP4 Training for Large-Scale MoE Models on Hopper GPUs

#CUDA #LLM #Hopper #FP4 #Precision #Package

https://hgpu.org/?p=30640

Practical FP4 Training for Large-Scale MoE Models on Hopper GPUs

Training large-scale Mixture-of-Experts (MoE) models is bottlenecked by activation memory and expert-parallel communication, yet FP4 training remains impractical on Hopper-class GPUs without nativeโ€ฆ

hgpu.org
Bruce Simpson, Ph.D.Mar 7
Currently being bitten by the lack of #nanosecond #precision in the standard #Python 3 #datetime module, even though time has it: https://github.com/python/cpython/issues/59648
datetime module has no support for nanoseconds ยท Issue #59648 ยท python/cpython

BPO 15443 Nosy @malemburg, @tim-one, @mdickinson, @abalkin, @giampaolo, @bitdancer, @andyclegg, @gareth-rees, @eli-b, @serhiy-storchaka, @pganssle, @shlomoa PRs #21987 Files datetime.nanosecond.pat...

GitHub