Ex-Coinbase Staff Engineer. Building apps and games.
Mountain View, CA
Joined May 2021
- Tweets511
- Following468
- Followers265
- Likes1.3K
Tried gpu-unlock on my RTX 3090:
✅ Unlock worked first try: FP16/BF16 matmul ~71 → 112-116 TFLOPS
❌ llama.cpp Q4 LLM: +0-2% (INT8 math, not FP16)
🖼️ Qwen-Image GGUF in ComfyUI: ~1.3x faster (estimated)
Big win for BF16 work, not quantized LLMs.
卧槽，3090要可能要涨价了！！
因为老黄的刀法被人破解了
有人给3090解了个锁，同样限制在 360W，PyTorch 测试从 75 TFLOPS 跑到了 135 TFLOPS，提升约 80%！
你以为是超频，把功耗往上拉？
但他改的是显卡里控制运算速度的寄存器，这些寄存器平时有权限保护，正常改不了
他利用 GPU 启动固件的漏洞，赶在 NVIDIA 驱动加载之前拿到写入权限，把相关运算改成全速，再交给原版驱动使用
解开的主要是 FP16／BF16 输入、FP32 累加这类 AI 运算的速度限制 但测试算力多了 80%！！
4090 也有分支项目做到了：同类运算的指令吞吐，从约 165 TFLOPS 提升到 331 TFLOPS，接近翻倍
项目地址，附上👇
github.com/Zagitta/gpu-unloc…
Coolify users, be careful out there
Just added smart hints to my GobNative engine for the 1991 classic Gobliiins! 🧙
I always loved these games, but some puzzles were ridiculously hard. Now the game tracks your progress and gives you little hints when you're stuck, without spoiling the fun!
games.encrypted-cat.com/devl…
The other half was my own code: • stepped Rapier's pipeline directly (no soft bodies to sync): 1.76 → 0.08 ms • 21 → 5 lights in every shader • debris goes static after landing
Lesson: test phones over https.
I vibe coded a native runtime for Gobliins 2, the 1992 Coktel Vision point-and-click puzzle game 🧩
No emulator: it reimplements the original engine, using ScummVM's code as the reference (huge thanks to them).
Before → after
It also runs Gobliiins (1991) and Goblins Quest 3 (1993), all three played start to finish by automated test scripts.
Native on macOS for now; Windows and a browser version are next.
AI slop is delegating the vision.
AI-assisted work is delegating the execution.
Slop: "Hey Suno, make me a cool electronic track."
The model picks the composition, arrangement, mix, and vocals. Your only decision was "cool."
Assisted: "Here's Strudel, where music is code. 128 BPM, F# minor, four-on-the-floor kick, syncopated bass, open the filter slowly, second synth at bar 32."
Now every choice is mine, and the output is something I can read, change, and iterate on.
Same AI. The difference is whether you showed up with decisions or just a wish.
I built my own engine for a one-hour horror game on top of @threejs.
Probably overkill 😅 But now 70 sequences, 116 interactables, dialogue, scares, puzzles, and enemy encounters all live in data.
The engine knows how to be scary. The story just tells it what to be scared of.
Wrote about how it works 👇
0xkot.itch.io/the-cradle-epi…
#threejs #gamedev #indiedev
Really enjoyed the demo! It has that classic old-school Final Fantasy feeling I’ve missed. The battles, magic, summons… it just hits right. Can’t wait for the full release!
