FINIK | GPU・グラボ

AI hardware, GPUs & datacenters. The machines behind your ChatGPT prompt. Sometimes it works.
グラボ・AI・自作PCの話題を英語で発信しています。

[FINIK | GPU・グラボ @Ruslan_BJ]

TWO RTX 5090s. BOTH MELTED.

Both ran AI models 24/7 at full load.

The connector wasn't even the hottest part.
The thermal cam found 92°C right next to it.

The fix cost less than a pizza.

[FINIK | GPU・グラボ @Ruslan_BJ]

@rsensui Nice. If you're running GGUF, try the MLX builds in LM Studio. On Apple Silicon they're usually noticeably faster.

[FINIK | GPU・グラボ @Ruslan_BJ]

@taki_pc_1115 Six years on AM4 is a solid run. If you're on AM5 now, a CPU drop-in around year 3–4 could stretch it even longer.

[FINIK | GPU・グラボ @Ruslan_BJ]

@bunkaich Fun fact: the original Game Boy cost ¥12,500 in 1989. Now ¥6,000 gets you one that plays the whole library.

[FINIK | GPU・グラボ @Ruslan_BJ]

@7GHz And the GPU side of it: 4K is 4× the pixels of FHD. Most people on Full HD aren't behind, they just didn't want to pay for a GPU that can push that.

[FINIK | GPU・グラボ @Ruslan_BJ]

@W_Yoshizaki ありがとうございます！

[FINIK | GPU・グラボ @Ruslan_BJ]

@epi_x Make it a curtain. Speakers you can't see, and the whole window plays music. Sound comes from fabric (4th time).

[FINIK | GPU・グラボ @Ruslan_BJ]

@MacroWorld21 MX23J12808 is a custom Macronix mask ROM made for Nintendo, so no public datasheet. The protection isn't in the chip, it's in the cartridge protocol (KEY1/KEY2 encryption). GBATEK documents all of it.

[FINIK | GPU・グラボ @Ruslan_BJ]

@Isichan_Hitori Interesting split between refusal and evasion. Removing the refusal direction fixes the obvious part, but evasion is spread across the whole model, so retraining the answering style makes sense.

[FINIK | GPU・グラボ @Ruslan_BJ]

@sinovoip Meanwhile in Japan, 1GB Raspberry Pis cost almost ¥9,000. A 32GB RK3588 board suddenly looks like a bargain.

[FINIK | GPU・グラボ @Ruslan_BJ]

@Yacamochi_db This is exactly what multi-GPU LLM builds need. A slot can look like x16 and still run at x4 through the chipset, and you only find out after the second card is in.

[FINIK | GPU・グラボ @Ruslan_BJ]

@PCONES Small note: SUPRIM is MSI's lineup, not ASUS. Looks like the shop tag got it mixed up too.

[FINIK | GPU・グラボ @Ruslan_BJ]

@sozoraemon The Pi 3 B+ costing more than a Pi 5 is wild. At that point the newer board is the only sane pick.

[FINIK | GPU・グラボ @Ruslan_BJ]

Running your own AI in the office sounds simple.
This is what it actually takes.
Dual Xeon, 64GB ECC sticks in every slot, and a wall of NVIDIA GPUs.
No cloud, no data leaving the building.
社内でAIを動かすなら、これくらい必要。
#ローカルLLM #自作PC

[FINIK | GPU・グラボ @Ruslan_BJ]

@hokazuya Honestly the most productive impulse buying I've ever seen. And B300 would be the logical next step: 288GB of HBM3e at around 8TB/s.

[FINIK | GPU・グラボ @Ruslan_BJ]

@Tono_Ken3 14kWh at ~480W is about 29 hours of full inference. Two days works if it spends part of the time idling.

[FINIK | GPU・グラボ @Ruslan_BJ]

@gollirapanty Most of that price jump is RAM. AI data centers are eating the DRAM supply, so DDR5 got hit the hardest.