3 ms·

It’s a very cool achievement (a NN in 200KB, to compress? Wow) but the CMIX compression program performs better in the 1st benchmark, and is competitive in the

It’s a very cool achievement (a NN in 200KB, to compress? Wow) but the CMIX compression program performs better in the 1st benchmark, and is competitive in the 2nd. So he does not beat the SoTA “by a substantial margin”. Your comment is overstating the case.

rndgermandude 5y ago
True. Aside from the compression efficiency, I'd like to see other metrics too, like the work it takes to compute the compression (in units of energy, time, or whatever), to get more of a sense on what kind of workloads using something like that would make sense. These super-efficient compressors often only make sense in long term archival and compress-once-transmit-often scenarios. It appears http://www.mattmahoney.net/dc/text.html http://www.mattmahoney.net/dc/text.html (that Bellard links to on the NNCP page) has some information: - NNCP (w/GPU) used 161812 nanoseconds/byte to compress, 158982 nanoseconds/byte to decompress, on an Intel Xeon E3-1230 v6, 3.5 GHz, RTX 3090 GPU, using about 6GB of RAM, to compress enwik9 down to 110,034,293 bytes - CMIX used 602867 nanoseconds/byte to compress, 601569 nanoseconds/byte to decompress, on an Intel Core i7-7700K, 32 GB DDR4 using about 25GB of RAM, to compress enwik9 down to 115,714,367 bytes - gzip -9 used 101 nanoseconds/byte to compress, 17 nanoseconds/byte to decompress, on an... Athlon-64 3500+ in 32-bit mode running XP, to compress enwik9 down to 322,591,995 bytes - 7zip 9.20 used 1031 nanoseconds/byte to compress, 42 nanoseconds/byte to decompress, on a Gateway M-7301U laptop with 2.0 GHz dual core Pentium T3200 (1MB L2 cache), 3 GB RAM, Vista SP1, 32 bit, to compress enwik9 down to 227,905,645 bytes These benchmarks are not directly comparable, as vastly different machines were used, but it still gives a trend at the very least. As expected, these super-efficient algorithms are slow, really slow. Honestly, I'd think at least today too slow to make sense even for long term archival, and far too slow for compress-once-transmit-often scenarios, as the decompression rate is almost as slow as the compression rate. Still, it's impressive work, don't get me wrong.

rndgermandude 5y ago

True. Aside from the compression efficiency, I'd like to see other metrics too, like the work it takes to compute the compression (in units of energy, time, or whatever), to get more of a sense on what kind of workloads using something like that would make sense. These super-efficient compressors often only make sense in long term archival and compress-once-transmit-often scenarios. It appears http://www.mattmahoney.net/dc/text.html http://www.mattmahoney.net/dc/text.html (that Bellard links to on the NNCP page) has some information: - NNCP (w/GPU) used 161812 nanoseconds/byte to compress, 158982 nanoseconds/byte to decompress, on an Intel Xeon E3-1230 v6, 3.5 GHz, RTX 3090 GPU, using about 6GB of RAM, to compress enwik9 down to 110,034,293 bytes - CMIX used 602867 nanoseconds/byte to compress, 601569 nanoseconds/byte to decompress, on an Intel Core i7-7700K, 32 GB DDR4 using about 25GB of RAM, to compress enwik9 down to 115,714,367 bytes - gzip -9 used 101 nanoseconds/byte to compress, 17 nanoseconds/byte to decompress, on an... Athlon-64 3500+ in 32-bit mode running XP, to compress enwik9 down to 322,591,995 bytes - 7zip 9.20 used 1031 nanoseconds/byte to compress, 42 nanoseconds/byte to decompress, on a Gateway M-7301U laptop with 2.0 GHz dual core Pentium T3200 (1MB L2 cache), 3 GB RAM, Vista SP1, 32 bit, to compress enwik9 down to 227,905,645 bytes These benchmarks are not directly comparable, as vastly different machines were used, but it still gives a trend at the very least. As expected, these super-efficient algorithms are slow, really slow. Honestly, I'd think at least today too slow to make sense even for long term archival, and far too slow for compress-once-transmit-often scenarios, as the decompression rate is almost as slow as the compression rate. Still, it's impressive work, don't get me wrong.