얀도링
11.2K posts

얀도링
@dara8887
- 헛소리 많으므로 팔로비추 (님들의 탐라가 쓰잘데없는 소리로 가득찰겁니다)


공대 지망이면 물화는 필수로 끼는 거고 자연대 지망이어도 물화는 필수로 끼는 거고 메디컬 지망이어도 물화는 필수로 끼는 거임 애초에 이과 지망이면 물화는 필수로 끼는 게 맞음


내 말은 이거였지 라고 할 거면 트윗을 내리는 게 좋지 않을까요? 늙크크 + 맥락맹이신 걸 남의 트윗 위에 글까지 써 가면서 티내실 필요는 없어 보이는데


[소신] 고등학교 물화 별거아니니까 그냥 대학 들어가서 배워도됨 일물 일화 전공기초로 가르쳐주잖아 뭔 특별한거 배우는줄 별거아님 쫄지마 뭔 중학교 내신을 잘따려면 미리미리 구몬 학습지를 해둬야해요 ㄷㄷ 보는것같네


신박한 소신 발언 접어두시죠. 물리 화학 배우고 못 배우고 얘기가 아닙니다. 대학에서 선택과목으로 물리 화학을 해야 좋은 평가를 받는다는 겁니다. 중학교 내신 따겠다고 구몬하냐는 신박한 소리는, 본인이 지금 하고 있는 말이에요.





キオクシア(旧東芝メモリ)の2026年売上は2.3兆円。 この会社の屋台骨を作った発明者が、東芝から受け取った対価は、8,700万円やった。 異常やった。 舛岡富士雄さん。 スマホ・SSD・USBメモリ、全部に入ってる「NAND型フラッシュメモリ」の発明者。 1987年に東芝で作り上げた技術が、40年後に世界を変えた。 でも発明した後、東芝社内で何が起きたか。 擁護してくれた上司が1991年に急逝。 翌年、提示されたポストは「研究所付技監」。 肩書きはあるけど、部下なし・研究費なし。 舛岡さんはこう言ってる。 「研究の場を失うのは、エンジニアにとって死を意味します」 3年間抵抗したけど、1994年に東芝を去った。 その後、2004年に「せめて40億円は認めてほしい」と提訴。 2006年に和解が成立して、受け取った額が8,700万円。 要求額のたった2%。 その間、東芝は舛岡さんの発明で生まれた特許を、インテルに二束三文で売り払った。 2.3兆円の売上を生む会社を作った発明が、8,700万円で終わった。 日本の企業が「天才を閑職に追いやる」と何を失うか、この数字が全部言ってる。 発明の価値は、会社が決めるもんやない。




Cerebras's insane yield strategy: One perfect giant chip vs. a flexible swarm of many small cores Imagine a flagship smartphone AP chip. It packs dozens of CPU cores, a powerful GPU, NPU, 5G modem, image processor, audio DSP, Bluetooth, security engine — everything in one big die. If even a single block has a defect during manufacturing — calls keep dropping, audio crackles, or the camera glitches — the entire chip is thrown away. That's the harsh reality of traditional large-chip yields. Every single function must be 100% perfect to pass. As dies get bigger, defect probability skyrockets. In cutting-edge 3nm or 2nm processes, hitting even 50-60% yield is considered a big win for mass production. Cerebras $CBRS completely flipped this rule. They take an entire 300mm wafer and turn it into one massive chip (the WSE-3). Inside are 970,000 tiny AI cores, all identical, each about 0.05mm². They operate independently and connect through a smart switch network. The key difference is how they define a good chip: "90,000+ working cores = Pass" - Defective cores are automatically bypassed. - Surrounding wiring simply acts as pass-through fabric. - Only the SRAM attached to the surviving cores is used. The result? 93% silicon utilization. They receive the whole wafer, carve 44GB of on-chip SRAM into 970,000 pieces, and use only what's attached to the 900,000+ good cores. The bad parts? Just discard them. The cost is already baked into the full-wafer price anyway. While traditional chip companies struggle to push yields past 60% on advanced nodes, Cerebras already achieves over 93% on a full wafer. This isn't a simple trick. It's a fundamental paradigm shift. - Traditional approach: Chase one perfect large die → extremely vulnerable to defects - Cerebras approach: Connect many slightly imperfect small cores → tolerate defects while delivering extreme performance Thanks to this, Cerebras can build AI inference chips that compete directly with Nvidia, without suffering the usual yield headaches. Of course it's incredibly hard to copy. It required a decade of work on dynamic routing for hundreds of thousands of cores, full-wafer testing, custom software stacks, and deep foundry partnerships. But the outcome is clear: Cerebras has become one of the few companies to overcome yield limits through architecture. And that crazy originality is now heading to the public market through their IPO. x.com/SVTrivo/status…




