3 ms·

The post with egregious errors was also put up on a Sunday afternoon. And while we're all acting conciliatory now, it's pretty remarkable how biased the post wa

The post with egregious errors was also put up on a Sunday afternoon. And while we're all acting conciliatory now, it's pretty remarkable how biased the post was, the author using some clearly erroneous numbers to prove their prior, baseless claim that the "M1 chip is far inferior" in some respects, when those respects were specifically SIMD. Then becoming strangely defensive when some people rightly pointed out that ARM64 has 128-bit NEON and a number of other advantages. Far inferior becomes....actually superior in many cases, even at SIMD.

jacobolus 6y ago
Let’s try to be charitable, shall we? Everyone makes mistakes sometimes, even leading experts in low-level algorithm optimization. Lemire was upfront about making a mistake, and not at all defensive about it; if you are reading it that way, it’s just you. It is clearly the case that the M1 CPU/SoC has a significant performance advantage in typical branchy single-core code, but much less advantage if any for certain kinds of heavily optimized numerics. Beyond that high-level summary, it’s good to dive into the details, and spark discussions. Everyone is just now getting their hands on these chips, learning how to work with them, and trying to figure out how to best optimize for them.
- bfgoodrich 6y ago
  After the original post, several people pointed out some pretty significant, breaking mistakes. For instance the claim that the Apple chip has "nothing of the sort" compared to the SIMD on the Intel chip, making it "far inferior [in some respects]". It seems dubious that someone cognizant of the 128-bit NEON instructions would say such a thing. The follow-up post then defensively used pejorative language to describe those people, such as that they were "rubbed the wrong way", among other dismissive lines, pointing to some prior entry to prove that surely he isn't unaware of such a thing, then going into the demonstration with the intention of proving his original point. That's...not useful. He made an enormous error because he was certain that his original wrong claims must be true. I'm not surprised that he recanted given that it was a pretty embarrassing, egregious error seen by thousands.

jacobolus 6y ago

Let’s try to be charitable, shall we? Everyone makes mistakes sometimes, even leading experts in low-level algorithm optimization. Lemire was upfront about making a mistake, and not at all defensive about it; if you are reading it that way, it’s just you. It is clearly the case that the M1 CPU/SoC has a significant performance advantage in typical branchy single-core code, but much less advantage if any for certain kinds of heavily optimized numerics. Beyond that high-level summary, it’s good to dive into the details, and spark discussions. Everyone is just now getting their hands on these chips, learning how to work with them, and trying to figure out how to best optimize for them.

bfgoodrich 6y ago
After the original post, several people pointed out some pretty significant, breaking mistakes. For instance the claim that the Apple chip has "nothing of the sort" compared to the SIMD on the Intel chip, making it "far inferior [in some respects]". It seems dubious that someone cognizant of the 128-bit NEON instructions would say such a thing. The follow-up post then defensively used pejorative language to describe those people, such as that they were "rubbed the wrong way", among other dismissive lines, pointing to some prior entry to prove that surely he isn't unaware of such a thing, then going into the demonstration with the intention of proving his original point. That's...not useful. He made an enormous error because he was certain that his original wrong claims must be true. I'm not surprised that he recanted given that it was a pretty embarrassing, egregious error seen by thousands.