Freelance writer/editor and half of @ScrewflyStudios, dev of Zafehouse, Deadnaut, Fear Equation. Formerly Weekend Editor, @GizmodoAU, @KotakuAU, @LifehackerAU
I recently had the honor of interviewing Zen’s chief architect Mike Clark! I tried to fit in as many microarchitecture questions as I could, including x64 vs ARM ISA power efficiency, 4k vs larger page sizes, 64-byte cache lines, scatter/gather and more:
https://t.co/PXQuw8ZzRh
Something that's definitely been lost is accepting its OK to show fewer or no results when, you know, there aren't any. Filling space, infinite scroll, etc. it's all cursed.
The Google Shopping experience.
Me: "4090 laptop"
Google Shopping: "Did you say 13-inch Ultrabooks with integrated graphics?"
How can a company the built its business on search be so terrible at search.
First look at RDNA 4 by @ChipsandCheese9 doesn't disappoint. This is really a big uarch change, the highlight for me was something I tried to analyse in a previous thread but I failed to realize what it meant! (Too ashamed to link to my own thread.) 🧵
https://t.co/lWk5KmFgEz
Tentative implementation in prodigy-plus-schedule-free. Applied to the update itself (rather than the first moment
before denom as per code), but seems to work well in very limited testing.
TLDR: 1⃣ line modification, satisfaction (theoretically and empirically) guaranteed 😀😀😀
Core idea: 🚨Do not update if you are not sure
👨����https://t.co/hrbtWQO53X
🤗https://t.co/xCLOEXSVX5
📚https://t.co/7JdekGRMOO
@cranialxix @lqiang67 @Tim38463182
@s4schoener Experienced similar pain with the Mono runtime in 2019.4 when working on Deadnaut: Signal Lost. I ended up just optimising the runtime instead. https://t.co/nkRRBksYtn
This included hand-optimising so the JIT would replace calls like Mathf.Round with direct SSE.
My current side project -- a modified Mono JIT (2019.4) with native, optimised SIMD support for @unitygames Vector and float types. Not Burst, just regular C#.
Currently working on adding support for fused backward pass into https://t.co/KnH423jDei. Should help cut memory usage even more (if I can make it work)!
I'm sure there are better options available, but I've found https://t.co/423EpNg6MY to be super-useful for quickly plotting curves. Process is to get o1 to generate the general plot, and then I can just copy/paste and tweak.
One of the things I've enjoyed most about machine learning is how much games dev / shader knowledge has transferred across. At the end of the road, there's always something matrix-shaped.
Is it overwhelming? Very much so. But at least I brought a snorkel.
@divideconcept I've only tested with SDXL LoRA training, where it's behaviour has been similar to AdamWSF as long as Prodigy predicts the LR early and stabilises. And yes, the idea is no parameters tweaking (hopefully). Recently added Adam-atan2, which removes epsilon too.
Had a crack at combining Prodigy with ScheduleFree, along with a bunch of optimisation pull requests on the Prodigy repo. Also includes several tweaks and improvements (atan2 updates, EMA for d): https://t.co/KnH423jDei
Seems to work well, but only really tested with SDXL.
I wasn't in the mood these days, but I brought a little update on the combat system, tomorrow I'll finish the first enemy AI
#doom#doomhand#modding#gamedev
Missing ThrottleStop on my new Ryzen, so decided to cook up my own homage. Obviously has less features and not everything is 1:1 (no per-core temps), but it's nice to see power/temps at a glance.