Interesting #discussion about maximizing #ram #bandwidth https://unum.cloud/post/2022-01-29-ddr4/ As the #hackernews #comments point out, with a #amd #threadripper #cpu you need to have your code #numa aware: https://news.ycombinator.com/item?id=301.. This means delegating a task to a specific #core and #ram to minimize #throttling the #infinityfabric .
Very cool work on #audio #compression https://ai.googleblog.com/2021/08/sounds.. I had an idea a while back about training an #audio #profile per #user and then pre-sending this to each client, which essentially then just performs #tts on the other end. Using this method, you could then communicate over something like #lora with ultra-low #bandwidth ...