GPU LZ77 decode on H100 does not focus where one usually looks. In this analysis, the main key is GPU LZ77 decode, and the question is simple: what exactly makes decoding sequential, and what can actually be removed.
The problem does not arise at the copy level, as is often assumed, but at the parse level. In four bodies, parse accounts for 63.7–71.5% of device-resident decode time, so acceleration needs to be sought precisely there. This is an important shift in the model: if the wrong layer is optimized, throughput changes little, even if the copy itself appears costly.
The authors compare three decoder architectures because each highlights its own bottleneck. Dense full-pipe, wavefront, and v7-RA seek show that different serialization mechanisms exist at different levels: parse, depth, token count. Such an analysis is particularly useful for architects because it separates the systemic bottleneck from the familiar but incorrect hypothesis of “complex copy” as the main cause.
There is also not just one solution here. First, back-reference chain depth can be limited at the encode stage, and this almost does not change the compression ratio: the cost is 0.006% compared to the baseline. Second, self-overlapping matches turned out not to be a chain of dependencies, but periodic fills, meaning they can be parallelized without losing bit-perfect correctness. This changes the execution scheme itself: instead of artificial serialization, a parallel match layer emerges.
The practical trade-off here is fair. The depth cap provides a gain only where depth truly forms latency and has almost no effect on the spike cluster: byte-level comparison of all 15,499 blocks showed that the cap changes 16 blocks, while in a cluster of 181 blocks, it changes nothing. Thus, for this section, the latency cap is useless. This is a good example of how measurement breaks intuition: not all depth in parse translates into latency.
The implementation is based on encode-time absolute offset parse. ACEAPEX stores back-references as absolute positions rather than relative offsets in a sliding window. Because of this, block decode begins as soon as the source blocks are ready. The format stores four streams per block: literals, lengths, absolute offsets, and commands. This provides block independence but retains another sequential element—a four-entry distance history, which can already be removed on the encoder side.
It is here that the toughest compromise arises. Removing distance history eliminates the last genuinely sequential element but costs 0.540% in ratio. However, the dependency-free parse run increases from 4 commands to 706. For a systems engineer, this is a clear trade-off: a slight reduction in compression efficiency in exchange for a simpler dependency graph and a more predictable decode path.
The memory floor is also important. Even after eliminating sequential parts, the granularity issue remains. On chr1, the median match is 7 bytes with a 128-byte cache line, resulting in a bus efficiency of 4.4%. A coalesced write of the same volume was measured as 39× faster. This shows that part of the limitation is no longer algorithmic but physical: too small writes do not fit well on the bus.
In the random access picture, this is particularly clear. For a 50 GB archive, the position within the file hardly changes the time: decoding one 16 KB tile fluctuates between 292.6–386.6 µs, or about ±14% around 344 µs. This is not a bit-perfect comparison because the original is not on disk, but positional stability has been measured. In other words, the format provides addressability but does not eliminate the cost of each small block.
The conclusion of the study is pragmatic. The main sequence in GPU LZ77 decode lies not in copy but in parse. The depth cap is almost free, self-overlap can be parallelized, and the last sequential element can be removed at the cost of a small loss in ratio. However, after this, there remains another limit—memory traffic, and it is this that defines the lower boundary of acceleration.
Information source
arXiv is the largest open preprint repository (since 1991, under the auspices of Cornell), where researchers quickly post working versions of papers; the materials are publicly accessible but do not undergo full peer review, so results should be considered preliminary and, where possible, checked against updated versions or peer‑reviewed journals. arxiv.org