The intelligence is in the classification. What comes out is raw bytes.
The server you don't have to buy.
Your integer column is 8 bytes wide because nobody asked how big it really is. Smart2Raw asks. It measures the real range and stores the column in the smallest native class that range needs — 8, 16, 32 or 64 bits, signed or not. No dictionary, no bit-packing, no decode step: the stored bytes are native integers, so every operation runs on them directly.
Up to 8× fewer bytes on the same machine, when the column's range allows it — and the same machine then holds 8× the working set. The demonstration below measures that on your data, in your own browser, in about 30 seconds.
Why this matters now
These three numbers are not ours. They come from third parties, and each one carries its source. The numbers we measure are above, and you reproduce every one of them right here — we keep the two apart on purpose.
3.0× against 1.6×
The bottleneck stopped being arithmetic
Over 20 years, peak hardware compute grew 3.0× every 2 years. DRAM bandwidth grew 1.6×; interconnect bandwidth, 1.4×. The gap widens every year.
IEEE Micro · arXiv:2403.14123
→ Adding FLOPs stopped fixing it. Reading fewer bytes, with no decode step to pay for, attacks exactly the side that got narrow.
+58% to 63%
In a single quarter
That is how much DRAM contract prices rose in Q2 2026, in the worst supply shortage in nearly 15 years. A single AI server uses 8 to 10 times the memory of a conventional one.
TrendForce, April 2026
→ A byte you never store is money you never spend — and memory left over for something else.
945 TWh
Data centres in 2030
Global data centre electricity demand is projected to more than double by 2030, reaching around 945 TWh, with AI as the main driver.
International Energy Agency
→ Moving fewer bytes is spending less energy per query. That is not a slogan: it is the same arithmetic, seen from the other side.
A library that asks you to trust a measurement should be the first to say which numbers are its own. Ours are above, and reproducible. These are other people's, and linked.
Where it fits
Every one of these already has an integer column in it. That is the whole addressable surface.
Databases and columnar engines
Ids, codes, dates, counters. The stored bytes are already native integers — no decode step, no materialisation, no dictionary resident in memory.
Operating systems and Linux
/proc counters, eBPF metrics, PIDs, inodes, uids, log timestamps. One header,
C11, zero dependencies: it fits inside a daemon or an agent without dragging a library in.
Observability, IoT, telemetry
Fixed-interval time series, sensors, counters. This is where the common stride and the per-block frame of reference work together.
AI and machine learning
Token ids, vocabulary indices, feature ids, dataset offsets, neighbour lists from a vector index. Inference is bound by memory bandwidth, not FLOPs — reading fewer bytes with no decode cost is the currency that is missing.
Embedded, MCU, edge, automotive
Any reading buffer. The lean mode is already tested: no stdio, no mmap, no SIMD.
Financial market data
Prices in cents (a stride!), nanosecond timestamps, instrument ids. Tick data is the picture of the ideal case.
Developer tooling
Symbol tables, offsets, indices in compilers, linkers and binary formats. A single header drops into any build.
Why you can check every claim on this page
The demonstration above is not a JavaScript imitation of Smart2Raw. It is include/smart2raw.h itself, compiled to WebAssembly and running in your browser. When it reports a size, s2r_pool_bytes() answered. When it offers a .s2r to download, s2r_blocked_save() wrote those bytes — the page only swapped the disk for a block of memory.
And every number it can check against a naive loop, it checks: every value is read back and compared to the original, the block metadata's sum is compared to a plain sum, and the file is written, reopened, CRC-verified and compared again before the download button appears. A disagreement raises a red badge instead of printing a pretty number.
Deposited and citable
Three versions with their own DOI on Zenodo, with a concept DOI covering all of them. How to cite.
Open, under AGPL-3.0
The published version is free software. Building it into a closed product needs a commercial licence — which is also how you reach Smart2Raw Premium, the version that is not published.
Portable by construction
The same file runs on x86, ARM with NEON and SVE2, RISC-V with RVV, big-endian machines and microcontrollers — and now in WebAssembly and in a Windows executable with no CRT.