An LZ4 cache pitched at np.save speed writes at about 1.5x np.save, by its own numbers

Start with the arithmetic the post declines to do. Its own fast-path baseline is under 1 ms per megabyte for np.save ^[3], and the demo array works out to exactly 100,000,000 bytes, the same size in decimal megabytes as the array used for the zlib baseline ^[13]. On that array, zlib is at least thirty times slower than the raw dump ^[14]. Being twenty times faster than zlib therefore lands at roughly 150 ms, about 1.5 ms per megabyte, or half again the cost of np.save rather than parity with it ^[15]. Halving a file for half again the write time is a trade most caching loops would take. It is a different sentence from the one the feature list is written to suggest.

The size claim is the easier of the two to check, because the format hands you the instrument. uncompressed_size sits at byte 0 and compressed_size at byte 8 ^[20], so the ratio actually achieved on your data is a 16-byte read and one division, with no decompression and no timing harness. Worth noting what the example compresses: float32 output of np.random.randn ^[12]. Noise-like input is the least friendly case for a byte-level matcher, so the measured figure on quantised or repetitive arrays could go either way from 2x. Ratio is a property of your bytes, not of the library.

Then the header. The fields shown come to 88 bytes: two 8-byte sizes, 64 bytes of shape slots, and two 4-byte words ^[16]. The declared size is 96 ^[8] and the excerpt stops after the version field ^[17], so eight bytes carry whatever encodes dtype and rank, which load() has to recover across int, uint, float and bool ^[9][10]. Tight, and plausible. The shape block is less thrifty: a 2-D array pays for eight slots and leaves 48 bytes of zeros in every file ^[21]. Immaterial at 100 MB, visible if the cache holds a million small arrays.

The claim I would read the C for is the lock. The post says offloading to C avoids interpreter overhead and the GIL during serialization ^[7], with a direct pointer from PyArray_DATA and no copy where the layout allows ^[6]. Whether the lock is actually dropped around the compression call decides whether a training loop can overlap its cache write with the next batch, and the published excerpt ends before that code ^[17]. The same gap covers what happens to an array that is not contiguous, where the zero-copy path does not apply by the post's own wording ^[6].

None of this needs the author's cooperation to settle. The measurement is three writes of your own arrays at both ends of the acceleration range ^[5], plus the two sizes read out of each header. That is a morning's work, and it produces the number the post left as a bullet point.