Random-Access Compressed VRAM Layer

Seek any tile in compressed data — with a provable bound on decode depth.

Don't decompress 50 GB to read 16 KB. Any tile decodes in 0.27 ms — from the start, middle, or end of the archive. One file, both CPU and GPU. Back-references are absolute output positions, so every block stands alone.

The class, as one picture — where ACEAPEX is the only full column

Property	lz4	zstd	Oodle	GDeflate	nvCOMP+DE	DietGPU	DFloat11	ACEAPEX
O(1) seek into compressed	—	frame	—	—	chunk	—	—	✓
One format, CPU + GPU decode	CPU	CPU	—	✓	GPU	GPU	GPU	✓
Bounded tail (depth M ≤ L)	—	—	—	—	—	—	—	✓
Position-invariant access	—	—	—	—	—	—	seq	✓
Open source	✓	✓	—	✓	—	✓	✓	✓
Peak decode throughput	fast	fast	fast	~600 GB/s	~600 GB/s	592 GB/s	—	200 GB/s
Compression ratio	low	high	high	mid	mid	1.3×	1.4×	4.0×*

Read the top rows, not the bottom. On peak throughput and ratio the checkmarks belong to others — specialized decoders and fixed-function hardware are faster in isolation. The point is the shape of the ACEAPEX column: it is the only one filled across seek + one-format + bounded-tail + position-invariance at once. That combination — random access into GPU-resident compressed data without decompressing the whole file — is the class. *ratio is data-dependent; 4.0× on genome, competitive on text, lower on vectors — see Research.

0.27 ms

One tile, any position. 1 block ≈ 100 blocks (launch-bound → O(1)).

ERR194147 · FNV-verified

4.53×

Self-overlapping matches are periodic fills, not dependency chains — fully parallel. 2.75–8.42× across four corpora.

chr1 match layer · bit-perfect · Paper 5

0.081 ms

chr1:5000000-5016000 out of a 254 MB archive — a genomic coordinate, not a byte offset. Against bgzip with a fai index, the standard for compressed references: 1.73× faster on median, 2.03× at p99, and 11.8% denser. Verified against samtools faidx.

byte-verified against the original · aceapex faidx + r --range

First row of the lzbench 2.3 table. Neutral third-party tool.

aceapex 1.0.1

Where ACEAPEX loses — stated first, in our own words

A comparison built only from wins invites the reviewer to fill the gaps in public. These are ours to name.

zstd-19 packs tighter — when it is allowed a whole-file window

On the same genome: ratio 4.04 (ours) vs 4.93 (zstd-19). Ratio is not our axis. But the comparison only means something once the constraint is equalised: under independent 16 KB blocks, where every codec is equally seekable, ACEAPEX is first on 3 of 5 corpora against zstd-19 (enwik8 2.638, enwik9 2.981, silesia 3.005) and ahead of zstd-3 everywhere by 4.9–16.1%. Blocking itself costs 6.7% on genome and 24.2% on text; the match layer returns about three quarters of it. See Paper 5.

A single cold tile over PCIe is faster

~10 µs for a 16 KB PCIe fetch vs 272 µs for one seek. The win is in aggregate and in residency — never staging the 50 GB at all.

Peak throughput belongs to hardware

Fixed-function decompression engines and DietGPU reach ~600 GB/s. Our 200 GB/s buys seekability, not raw rate.