Saar research · systems, fabric and evidence

Compression is moving
into the fabric.

Saar starts with byte-exact recovery and complete encoded size, then asks what happens when compression leaves a single process and enters streams, shards, memory fabrics and accelerators. The fabric is a research target, not a throughput claim.

Deterministic shardsTransport-neutralByte-exact recovery
future compression stack5 steps
5layers between today's measured engine and a future fabric
Single processmeasured floor: 483.2 MB/s per core
Parallel streamsordered worker batches, byte-identical output
Shard protocoltransport-neutral task and result, order-independent assembly
Fabric runtimescheduling, backpressure and failure recovery
AcceleratorsDPU, GPU, ASIC, CXL and photonic paths
Byte-exactOrder-independentMeasured, not assumed
Each layer must preserve the same recovery contract. A faster path that changes the bytes is not a Saar result.

Reproducible by default

Every study records the corpus, revision, environment, settings and complete encoded sizes so another qualified reviewer can repeat the comparison.

Recovery before size

A candidate is accepted only after it decodes back to the source bytes. Byte-for-byte recovery, cross-decoding and malformed-input rejection come first.

Fresh, bounded holdouts

Development inputs are separated from evaluation workloads, each set is frozen before it is measured, and negative results are kept rather than discarded.

Controlled disclosure

Only approved black-box findings are published. Encoder mechanisms, tuning rules and patent-sensitive material stay under review.

Future compression stack

Where compression goes next.

The single-process engine is the measured floor. The interesting research is what happens when the work is distributed, offloaded or embedded in the path the data already takes.

Fabric-native execution

Saar's shard protocol is transport-neutral and assembles in completion order. The research question is scheduling, backpressure and failure recovery across a real fabric.

Accelerator offload

DPUs, GPUs, ASICs and in-storage compression units change the break-even point. Saar measures the whole path, not just codec throughput.

Memory fabric

CXL pooling and near-memory compression can turn saved bytes into effective bandwidth. The open question is whether the saved traffic pays for added latency and area.

AI data movement

KV-cache transfer and collective communication are becoming compression problems. Saar's byte-exact contract is a candidate for the transport layer, not a claim about model quality.

Learned and adaptive modes

Learned models can choose transforms or predict entropy. Saar's constraint is that every accepted path stays deterministic, reversible and auditable.

Photonic and neuromorphic

Photonic accelerators use task-relevant subspace compression, not file compression. The adjacent question is whether a byte-exact container can carry their inputs and results without loss.

Quantum hardware and the fabric

Fabric Research explores real quantum hardware for the fabric, including compression and decompression workloads. No quantum hardware is integrated or claimed; QNSI post-quantum signing is live in the product and is separate from this exploration.

Open hypotheses

What we would test before claiming it works.

A hypothesis is useful only if it can fail. These are the questions we would put to a fabric, accelerator or memory tier before calling the result a Saar result.

Break-even

Compression is a system property

Compression wins only when encode + transfer + decode + verify is cheaper than moving the original bytes. Measure the whole path, not one component.

Granularity

Small chunks change the answer

At DPU, ASIC and CXL granularity, per-chunk overhead can erase the gain. Find the minimum viable chunk for each device before claiming a speed-up.

Failure

Determinism must survive failure

Retries, shard migration and partial completion must not change the decoded bytes. Test failure paths, not just the happy path.

CXL

Memory compression is not free bandwidth

CXL compression adds area, power and load-to-use latency. Compare effective bandwidth and end-to-end latency, not compression ratio alone.

AI streams

Exactness and speed together

KV and tensor transfer need bitwise recovery and GPU-friendly kernels. A codec that costs more than the transfer it replaces is not a win.

Auditability

Learning must remain inspectable

A learned selector is acceptable only if the chosen path is deterministic and the output is verified. A model is not a licence to skip recovery.

External research signals

The field is moving in public.

These are external papers and engineering notes we track. They are not Saar measurements, endorsements or proof that a technique will work for our engine. They show where the problem is moving.

2023 · ISCA / Google Research

CDPU: Co-designing compression and decompression processing units for hyperscale systems

Hyperscale systems already treat compression and decompression as a datacentre tax. Hardware co-design is an active response.

Read the source

28 September 2025 · arXiv 2509.23693

DPZip: ASIC-based compression accelerators for storage systems

In-storage compression and decompression units are being designed and profiled as first-class device hardware.

Read the source

3 September 2025 · arXiv 2509.03377

TRACE: Unlocking effective CXL bandwidth via lossless compression and precision scaling

CXL memory bandwidth, not just capacity, is a compression problem. Layout and precision scaling matter as much as the codec.

Read the source

30 April 2026 · arXiv 2604.27844

ZipCCL: Efficient lossless data compression of communication collectives

Lossless compression can help collective communication, but only when kernel overhead does not erase the bandwidth saving.

Read the source

3 May 2026 · arXiv 2605.01708

SplitZip: Ultra fast lossless KV compression for disaggregated LLM serving

KV-cache transfer between prefill and decode workers is a concrete, GPU-bound compression target.

Read the source

24 August 2026 · Meta Engineering

MetaRoCE: A new RDMA transport built for AI-scale Ethernet

The fabric itself is being redesigned around AI traffic. Compression research has to meet the transport where it actually runs.

Read the source

2026 · Opto-Electronic Advances

A compression-decompression framework for universal acceleration of photonic neuromorphic computing

Photonic computing uses task-relevant subspace compression. It is adjacent to, not identical with, byte-exact file compression.

Read the source

Scaling to fabric

One machine today. A fabric on purpose.

The throughput research recorded on 21 October 2026 measured the fast tier at 483.2 MB/s per core and 2674.7 MB/s at 8 threads on Apple M4 Max, 36 GB (macOS arm64), with scaling capped near 5.6× by serialised I/O. The conclusion is explicit: TB/PB is a fabric and accelerator question, not a single-process tweak. QNSI post-quantum signing is live in the licence and attestation paths; exploring real quantum hardware for the fabric belongs to Fabric Research, not the core integrity claim.

Built · parallel streams

Deterministic ordered batches

Native worker batches via --threads N are byte-identical across tested thread counts, and native stream I/O now overlaps compute, delivering higher ratio-preserving throughput on the same machine.

Built · distribution

Shard protocol and worker

A transport-neutral binary shard task/result protocol with bounded worker orchestration and completion-order-independent SAST v3 assembly: saar shard-worker. Any fabric runs it; order never decides correctness.

Built · large delivery

Resumable verified parts

Streams checkpoint every verified part and resume across sessions. The hosted workspace admits files up to 5 TiB, and the hourly maintenance task continues interrupted work.

Next · single process

I/O overlap, then SIMD hashing

The record names the next levers in order: I/O/compute overlap to recover ratio-preserving throughput (partially landed), then SIMD batch hashing per core. Each is a measured, byte-identical change.

Single machine, single workload, development evidence. Not a cross-hardware, production or competitor-leadership claim. The only baseline that counts is your current codec on your machine.

Newest measured audit

Domain workload audit: 13 workloads

Thirteen public and synthetic workload samples measured against the best recorded output of gzip -9, bzip2 -9, zstd -19 --ultra, brotli -q11 and xz -9 on one host. The 87.13% figure is the reduction from original bytes to Saar bytes; the 27.48% figure is the margin over the best competitor per file. Development evidence: not a fresh holdout, not an independent certification, and not a claim about files that were not measured. Comparative speed was not measured for this study.

Byte-exact round-trip verified for every file. The complete Saar object is compared per file.

  1. Files

    13 workloads

    model.stl · rdb.sqlite · gdsii.gds · graph.bin · cloud.pcd · las.bin · fits.bin · seismic.sgy · netcdf.nc · dicom.dcm · fastq.bin · hdf5.h5 · pcap.bin

  2. Aggregate

    87.13% smaller than original · 27.48% vs best · 13 of 13 won

    7.77× compression from original bytes. The structured & scientific subset of 7 files runs 88.05% smaller than original (30.1% vs best).

  3. Recovery

    Byte-exact

    Byte-exact round-trip verified for every file. The complete Saar object is compared per file.

  4. Recorded

    21 October 2026

    Public and synthetic samples; not a fresh holdout and not independently curated.

Size-only study · 13 workloads

Every workload in the domain audit.

These rows are the 13 files behind the 87.13% reduction from original bytes (7.77× compression) and the 27.48% margin over the best competitor per file. Comparative speed was not measured for this study. The speed suites below are a separate dataset.

Each bar is one workload; height is the saving against the best recorded competitor for that file.

Workload fileInput bytesSaar bytesBest otherBest codecResult
cloud.pcd35,6506,8568,244brotli -q11Won · 16.8% vs best
dicom.dcm1,890336409brotli -q11Won · 17.8% vs best
fastq.bin3,041280333brotli -q11Won · 15.9% vs best
fits.bin57,6006,0836,506brotli -q11Won · 6.5% vs best
gdsii.gds92,27835,23538,272xz -9Won · 7.9% vs best
graph.bin109,26199,495107,692brotli -q11Won · 7.6% vs best
hdf5.h52,216286291brotli -q11Won · 1.7% vs best
las.bin36,43719,09621,740xz -9Won · 12.2% vs best
model.stl186,04914,23618,528brotli -q11Won · 23.2% vs best
netcdf.nc7,646612645brotli -q11Won · 5.1% vs best
pcap.bin1,400264265zstd -19Won · 0.4% vs best
rdb.sqlite2,899,968260,216407,540xz -9Won · 36.1% vs best
seismic.sgy14,6007461,460xz -9Won · 48.9% vs best
13 files3,448,036443,741611,925Best per file87.13% smaller than original · 27.48% vs best · 13 won

Development evidence: not a fresh holdout, not an independent certification and not a claim about files that were not measured. Full method and limitations are in the audit record.

Recorded evidence

Every report, with its limits attached.

Open any entry to see the revision, environment, trial method and the limitations that bound the result. Measurement-only extracts. No codec source, models, private scripts or corpus payloads. No independent benchmark certification.

All Saar timing is measured on a single local development reference machine (currently Apple M4 Max, 36 GB, macOS arm64). It is never a world baseline and is not a cross-hardware comparison; the only meaningful baseline is your own current codec on your own representative environment.

Native fast profile · 9 workloadsNine previously acquired public binary/media/model files, each at most 16 MiB; not freshly curated and not an independ…17 Oct 2026 · 9 workloads
Revision
Current native revision re-measured on 18 October 2026 over the frozen round31 file set. Whole-command CLI method for every codec; output frame sizes are deterministic and match earlier revisions.
Environment
macOS-27.0.1-arm64-arm-64bit-Mach-O
Method
One warmup, then seven rotated serial whole-command trials per codec and file. Charts sum the per-file medians.
Limitations
Nine previously acquired public binary/media/model files, each at most 16 MiB; not freshly curated and not an independent leaderboard. Process launch and pipe transfer are included and dominate small files, so aggregate speeds understate in-process codec throughput. The large-stream tier is published separately on /product. Peak memory was not measured. Results support this workload set only.
Comparison versions
Brotli brotli 1.2.0 · Zstandard Zstandard CLI 1.5.7 · LZ4 LZ4 CLI 1.10.0
Native fast profile · text & config · 4 workloadsFour previously acquired public text/config files; not freshly curated. Process launch and pipe transfer are included…17 Oct 2026 · 4 workloads
Revision
Current native revision re-measured on 18 October 2026 over the frozen round23b file set. Whole-command CLI method for every codec.
Environment
macOS-27.0.1-arm64-arm-64bit-Mach-O
Method
One warmup, then seven rotated serial whole-command trials per codec and file. Charts sum the per-file medians.
Limitations
Four previously acquired public text/config files; not freshly curated. Process launch and pipe transfer are included and dominate small files. Peak memory was not measured. Results support this workload set only.
Comparison versions
Brotli brotli 1.2.0 · Zstandard Zstandard CLI 1.5.7 · LZ4 LZ4 CLI 1.10.0
17-file archive · 7 codecsExisting local corpus, not a blind holdout. Process startup is included for CLI baselines; Saar ran in-process. Timing…11 Oct 2026 · 17 workloads
Revision
Historical TypeScript experiment, 11 October 2026. Not the default encoder or current native release.
Environment
macOS arm64 · Node v24.21.0 · hardware model not recorded
Method
One timing trial per codec and file. No confidence intervals.
Limitations
Existing local corpus, not a blind holdout. Process startup is included for CLI baselines; Saar ran in-process. Timing is not an isolated algorithm comparison. Per-codec RSS was not measured. Historical results are not current-release performance.
Comparison versions
gzip -9 Apple gzip 487.0.1 (macOS) · zstd -19 Zstandard 1.5.7 · brotli -q11 Brotli 1.2.0 · xz -9 XZ Utils 5.8.4 · lz4 -9 LZ4 1.10.0
Round22 unique holdout · 3 codecsSix acquired holdout files; two byte-identical duplicates excluded, four unique. The full profile searches the active…17 Oct 2026 · 4 workloads
Revision
Current native revision re-measured on 18 October 2026 with the full compression profile over the frozen round22 holdout; two duplicate payloads remain excluded.
Environment
macOS-27.0.1-arm64-arm-64bit-Mach-O
Method
One warmup, then seven rotated serial whole-command trials per codec and file. Medians per file.
Limitations
Six acquired holdout files; two byte-identical duplicates excluded, four unique. The full profile searches the active native encoder families and is magnitude slower by design (tens of KB/s class on this host). It is for archived, rarely-read data, not interactive paths. Not a fresh holdout or independent certification.
Comparison versions
Brotli brotli 1.2.0 · bzip2 bzip2 1.0.8
Multimodal holdout · 4 codecsFive previously acquired multimodal files (protocol text, PNG, MP4, Ogg, JPEG); not freshly curated. Near-entropy medi…17 Oct 2026 · 5 workloads
Revision
Current native revision re-measured on 18 October 2026 with the full compression profile over the frozen multimodal set.
Environment
macOS-27.0.1-arm64-arm-64bit-Mach-O
Method
One warmup, then seven rotated serial whole-command trials per codec and file. Medians per file.
Limitations
Five previously acquired multimodal files (protocol text, PNG, MP4, Ogg, JPEG); not freshly curated. Near-entropy media cannot shrink losslessly; stored fallback bounds expansion. Not a fresh holdout or independent certification.
Comparison versions
gzip Apple gzip 487.0.1 · xz XZ Utils 5.8.4 · bzip2 bzip2 1.0.8
Parallel stream · Silesia 12 filesHeterogeneous public benchmark set (text, binaries, executables, images); not freshly curated and not an independent l…17 Oct 2026 · 12 workloads
Revision
Current native revision re-measured on 18 October 2026 through the checked SAST v3 stream CLI. 4 MiB chunks, --threads 8; zstd/lz4 run -T8.
Environment
macOS-27.0.1-arm64-arm-64bit-Mach-O
Method
Per file: three timed runs after one warmup, whole-command, medians. Charts sum the per-file medians.
Limitations
Heterogeneous public benchmark set (text, binaries, executables, images); not freshly curated and not an independent leaderboard. Per-file runs include process launch for every codec equally. The same record also measures the full 211.9 MB set as one combined stream: Saar 686 MB/s encode / 310 MB/s decode, vs Zstandard -3 1411/423 and LZ4 -1 1933/379. This is the tier comparable to the fast-throughput research. Byte-exact recovery verified on every trial.
Comparison versions
Zstandard Zstandard CLI 1.5.7 · LZ4 LZ4 CLI 1.10.0

Development records

Dated, reproducible, including the retraction.

Beyond the published suites, the research log itself is public: each entry names its evidence file under evidence/, including what was retracted and why.

Workload class matrix: all 31 classesEvery public workload class has a dated measurement against the strongest available standard codec: 31 wins shown in place.18 October 2026
Evidence file
evidence/benchmarks/workload-class-matrix-2026-10-18.json
Result
Every public workload class has a dated measurement against the strongest available standard codec: 31 wins shown in place.
Real-sample audit: gap classesPublic samples acquired from package registries and official images (NIST KATs, Bitcoin vectors, DKIM mail, OCI inventory, server configs, a genuine core-dump prefix, LiDAR): the fixed-profile audit is 8 wins, 2 losses on real files. The targeted configuration scan closes the OCI inventory loss at the strongest verified configuration (584 vs 588), leaving nginx-configs.tar.gz as the single raw-file loss.18 October 2026
Evidence file
evidence/benchmarks/workload-audit-2026-10-18-real-samples.json
Result
Public samples acquired from package registries and official images (NIST KATs, Bitcoin vectors, DKIM mail, OCI inventory, server configs, a genuine core-dump prefix, LiDAR): the fixed-profile audit is 8 wins, 2 losses on real files. The targeted configuration scan closes the OCI inventory loss at the strongest verified configuration (584 vs 588), leaving nginx-configs.tar.gz as the single raw-file loss.
Domain workload audit: 13 workloads13 public/synthetic samples vs the best of gzip -9, bzip2 -9, zstd -19 --ultra, brotli -q11, xz -9: 87.13% smaller than original, 27.48% smaller than the best competitor per file, 13/13 wins, byte-exact round-trip.15 October 2026
Evidence file
evidence/benchmarks/workload-audit-2026-10-15.json
Result
13 public/synthetic samples vs the best of gzip -9, bzip2 -9, zstd -19 --ultra, brotli -q11, xz -9: 87.13% smaller than original, 27.48% smaller than the best competitor per file, 13/13 wins, byte-exact round-trip.
Fast-throughput research249 MiB, 4 MiB chunks: 483.2 MB/s per core, 2674.7 MB/s at 8 threads (three-run median), 5.5× scaling. The earlier 2982.8 MB/s result is historical. The path to TB/PB is a fabric and accelerator question.17 October 2026
Evidence file
evidence/benchmarks/fast-throughput-research-2026-10-17.json
Result
249 MiB, 4 MiB chunks: 483.2 MB/s per core, 2674.7 MB/s at 8 threads (three-run median), 5.5× scaling. The earlier 2982.8 MB/s result is historical. The path to TB/PB is a fabric and accelerator question.
Record-engine exact transforms: with a retractionExact reversible transforms measured; earlier lossy JSON prototype figures (−20% to −33%) retracted because they were not byte-reversible. The real payoff: ~6% on tabular CSV.17 October 2026
Evidence file
evidence/benchmarks/record-engine-exact-2026-10-17.json
Result
Exact reversible transforms measured; earlier lossy JSON prototype figures (−20% to −33%) retracted because they were not byte-reversible. The real payoff: ~6% on tabular CSV.
Planned-domain headroom scanThe generic context-mixing model already wins on the planned-domain data held; build order is classification coverage and speed, not a new engine per format.17 October 2026
Evidence file
evidence/benchmarks/planned-domain-headroom-2026-10-17.json
Result
The generic context-mixing model already wins on the planned-domain data held; build order is classification coverage and speed, not a new engine per format.
Domain modes 27-34Reversible, byte-exact specialist modes recorded per domain; mode 34 (sparse-residual context mixing) closes the sparse-image virtualization gap with TS-encode/native-decode cross-verification.15-17 October 2026
Evidence file
evidence/benchmarks/{png-mode27,wav-mode30,zip-mode29,signal-mode28,raster-mode32,pdf-mode33,cm-sparse-residual-mode34}-*.json
Result
Reversible, byte-exact specialist modes recorded per domain; mode 34 (sparse-residual context mixing) closes the sparse-image virtualization gap with TS-encode/native-decode cross-verification.
Web samples coverageVolumetric medical (NIfTI), gaming (glTF), model weights (ONNX/TFLite) and virtualization images measured against the best of five standard codecs.16 October 2026
Evidence file
evidence/benchmarks/web-samples-coverage-2026-10-16.json
Result
Volumetric medical (NIfTI), gaming (glTF), model weights (ONNX/TFLite) and virtualization images measured against the best of five standard codecs.

Intellectual honesty

Where Saar does not lead.

A vendor that only shows winning workloads has told you nothing about risk. These are the cases where we are the wrong tool, or where the result does not support a strong claim.

Encrypted or already-compressed data

Data at its entropy limit cannot be made smaller losslessly. Saar bounds expansion in that case instead of promising a saving that is physically impossible.

Ultra-fast, low-ratio codecs

On near-incompressible binary, media and model files, LZ4 -1 reaches memory-bound speed with almost no size gain (1.06×). That is a different class of tool, not a like-for-like compression result.

Single measurements

The current stream-throughput figure is one local Silesia run with no confidence interval. It is not a fresh holdout and not a universal leadership claim.

Workload dependence

Results move with the data. Text and source compress very differently from binary, media and models. Two suites are shown side by side instead of one headline number.

Public evidence

What buyers can evaluate

  • Recorded black-box results and the limitations attached to each
  • Format, compatibility and resource-bound facts
  • Approved papers and reports after technical and IP review

Protected research

What remains confidential

  • Encoder and decoder mechanisms, tuning and model rules
  • Exact ablations, unsuccessful experiments and source fingerprints
  • Patent-sensitive material, private reports and partner data

Publication status

No Saar research paper, patentability opinion, independent certification or public algorithm disclosure is currently approved. Future publication will follow technical review, reproducibility review and owner-approved IP disclosure controls.

Research with us

Bring an independent question.

We welcome rigorous benchmarking, systems research and domain-specific workload studies with clear disclosure boundaries. This includes adversarial studies that try to falsify a Saar claim.