John CarmackVerified account

Expensive to operate, but maybe for something like a Cerebras wafer that already has an exotic support system, adding liquid helium might not be ridiculous.

Fully superconducting circuits haven't panned out, but how would just replacing the metal layers on a conventional die with a L-He SC impact operation? Is quantum tunneling less of a problem with SC? Presumably the Si transistors need to be tweaked for the low temperature.

I like working with notebooks, but still get surprised by this.

Actually doing any processing in Python, as opposed to directing libraries to do things, is *shockingly* slow. I just wanted to build a bit array from a byte array, and it was close to 100x slower than the (tiny) conv net generating the data.

I wonder if open plan office workers have a statistically significant higher level of sick days than private office workers.

\ Did any microcomputers before the Amiga allow the full NTSC frame to be usefully addressed beyond "border color"? TV bezels were an issue, but the flexibility would have found uses.

\ The memory access time was always there, regardless of blanking, and all text/GR/HGR modes could work. The original numbers were a result of @stevewoz hyper-optimizing chip count, but by the time of the //e, adding new video mode addressing might have been justifiable. \

Randomly thinking how it would have been neat if 8 bit Apple // systems had an option to use the full NTSC active area instead of adding additional blanking. That would be around 53 chars * 7 = 371 horizontal and 240 vertical, with the option to interlace for double vertical. \

Struggling though some Expert+ songs in @BeatSaber , then going back to Expert level makes you feel like Quicksilver in the X-Men movies.

This is weird: I wanted to use MEM_LARGE_PAGES, but doing AdjustTokenPrivileges for SeLockMemoryPrivilege (after adding my user account to the security group) makes everything go 40% slower, regardless of whether I actually use the large pages or not. Anyone else seen this?

I just ran the same code on a 96 vCPU google cloud instance, and it took 207 seconds using all 96 vCPU across two groups. I'm not going to wait around for any of the slower counts to finish. That is an $8k/month instance, and it is less than 1/5 the speed.

instead of shared two to a core. That means that if you don't want to add the windows group code, you are better off disabling hyperthreading and having 64 single thread cores in a single group. I expected this code to be memory bound sooner, I'm impressed with the scalability!

AMD 3990 CPU scaling tests: Because of the Windows group limit of 64 CPUs, just firing up a lot of C++ std::threads didn't work great: 128 t = 67 s 64 t = 63 s 32 t = 84 s 16 t = 160 s 8 t = 312 s 32 to 64 threads wasn't a big boost, and 64 to 128 was slower. However! \

I grudgingly acknowledge the value of all the warnings about possible int overflow when calculating a size_t value on a 64 bit build, so it is irritating to get: error C2148: total size of array must not exceed 0x7fffffff bytes

Reading an old USB book, "Key USB Features" included "Single Connector Type".