iFred retweeted
[단독] 삼성전자, ‘차세대 HBM’ TSMC와 함께 만든다
Samsung, $TSM, $NVDA
엔비디아 ‘NVHBM’ 공동 개발·생산
HBM의 핵심 구성품 ‘베이스 다이’
삼성 파운드리 대신 TSMC 3나노 적용
삼성, HBM 제조 ‘턴키’ 경쟁력에도
빅테크의 까다로운 요구에 한발 물러서
삼성전자(005930)가 엔비디아의 차세대 맞춤형 고대역폭메모리(HBM)를 개발하면서 핵심 구성품인 ‘베이스다이’를 세계 최대 파운드리(반도체 위탁 생산) 업체인 TSMC 공정을 활용하기로 했다. 그간 HBM 공급 시 메모리칩 역할인 코어다이와 시스템반도체 역할인 베이스다이를 모두 자체 생산해온 삼성전자가 엔비디아의 기술적 요구 사항을 수용해 ‘턴키(일괄공급)’ 기반의 HBM 제조 방식을 바꾸기로 한 것이다.
18일 업계에 따르면 엔비디아는 올 8월 공개한 커스텀 HBM인 ‘NVHBM’ 생산을 맡긴 삼성전자와 SK하이닉스(000660)·마이크론에 베이스다이를 TSMC의 3㎚(나노미터·10억분의 1m) 공정으로 만들어줄 것을 요구했다. 코어다이 역시 6세대 D램 공정(1c)으로 통일을 요청했다. NVHBM은 차세대 HBM인 HBM4E(7세대) 구조를 기반으로 한다.
엔비디아의 이 같은 요구는 HBM 공급 업체마다 제조 방식이 다르면 설계한 제품 성능을 일관되게 구현하기 어렵기 때문이다. 업체마다 제품 검증을 별도로 해야 하는 부담도 줄일 수 있다. 파운드리 사업을 거느린 삼성전자 DS 부문은 내부 검토 끝에 엔비디아 요구를 수용하기로 했다.
삼성전자는 HBM을 그래픽처리장치(GPU)와 연결하는 과정에서 4년 전 TSMC의 패키징 공정과 호환을 위해 협력한 적은 있지만 HBM 제조 단계부터 손을 잡는 것은 이번이 처음이다.
삼성은 다만 HBM 턴키 전략은 표준 HBM 제품에서 그대로 유지한다. 삼성전자는 올 초 세계 최초로 양산한 6세대 HBM4와 양산 검증 중인 7세대 HBM4E의 베이스다이에 삼성 파운드리의 4나노 공정을 적용하고 있다. 두 제품의 코어다이 역시 삼성전자 화성·평택캠퍼스의 1c 라인에서 만들고 있다.
삼성전자는 AMD 등 다른 고객사의 커스텀 HBM에서도 턴키 방식을 고수할 방침이다. 아울러 NVHBM 개발·생산에도 자체 파운드리 활용 가능성을 남겨뒀다. 업계 관계자는 “삼성은 NVHBM용 베이스다이를 자사 파운드리에서도 만들 수 있는지 검증하고 있다”며 “NVHBM 수요가 TSMC 생산능력을 넘어설 경우를 대비한 것”이라고 전했다.
iFred retweeted
The usefulness of Amazon’s 6-page memo has been nuked by AI-generated text (employees sit in meetings suspecting text is AI and tune out).
Suspicious memos tagged as AI;DR (AI, didn’t read).
iFred retweeted
Some of you may remember the old "Meteostar" forecast tables. I think its no way to do real analysis, but it can be a nice summary graphic. So I made something similar for SatelliteWx. These are in the meteogram tab->Hourly Table->Rows (classic).
This is was only two cards.
$500,000 worth of ZOTAC RTX 5090 shipments reportedly "go missing" at Brazil's São Paulo airport. 🔗 wccf.tech/1lebo
Someone responding to the lack of F500 uptake with “enterprises are lethargic and behind the times” is pure cope.
“This is why we don’t hire people with children or anyone over 30.”
- a recruiter bragging to me about how their team stays locked in
i have no horse in the race here - but to everyone trashing the 8am-8pm requirement, availability weekends and nights, and 3 year commitment...
this is the tacit/unsaid requirement of every job I have ever had that had real growth, learning, and ownership baked into it. the only difference is VK is being explicit about it in the JD
All points I agree with, but to add:
- the added capacity from the big three isn't NAND or consumer DDR, its HBM and high density DDR packages
- timelines for new fabs dont have production scale impacts until well beyond 2030. Probably looking at '32-35 if the demand curve keeps
- any consumer savior will either be through new entrants (Intel+SK is one example) or through new tech that pushes density (zram, 1t2c, exoctic materials)
- china isnt going to save you, look at where advanced ymtc nand and advanced cxmt dram is ending up
- Starting with nvidia, and soon others, it will make more economical sense to outright abandon a customer line to accommodate AI infra; see the aib distress, g202 allocation, and emergence of the RTX 5500
- any fall in training demand is being eaten up by inference, and inference demand is proving to be durable; so any AI infra companies succumbing to a bubble will not cause prices to crash
Thoughts on Memory
1. I do not agree with the idea of a memory supply glut in 2029. I expect demand growth to accelerate further.
2. Around 2029 to 2030, the memory business will undergo a fundamental transformation. Think of how we no longer view NVIDIA as simply a gaming GPU company.
3. China will remain a competitive threat, but equipment constraints will make capacity expansion harder than many expect. Its impact on the supply and demand balance in 2029 will also be smaller than the market fears.
4. I expect memory stocks to begin rerating within the next six months. This also ties into a broader rerating of the KOSPI.
5. I believe memory needs to rerate first for other AI themes to follow. That view has not changed.
Source: me
iFred retweeted
what this actually means:
> ex-openai employee tried reading ai neurons by hand years ago, only ~2% made sense
> now ai agents do the reading, and most of them make sense, even in trillion-param models
> if it holds, alignment stops being guesswork
bitter lesson strikes again. the dumb thing at absurd scale just works
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Telling spam callers that you're going to prison and that you are legally required to collect their information is a great way to never get a spam call again.
iFred retweeted
en.gamegpu.com/news/zhelezo/…
According to unconfirmed reports, NVIDIA has stopped supplying GB202 graphics processors to gaming partners. We believe the company may be shifting manufacturing capacity toward AI server accelerators while preparing the next generation of GeForce graphics cards.
Our forecast suggests the GeForce RTX 6090 could debut at CES 2027 in January, with sales beginning in late January or early February. Rumored specifications include a 3nm GR202 chip, over 150 billion transistors, 768 Tensor cores, 48 GB of GDDR7 memory, 2 TB/s bandwidth, and a 600 W TGP.
We estimate a potential launch price of $2,999, compared with $1,999 for the RTX 5090. If shortages become severe, retail prices could reach $12,000–$15,000. Full details will only become known after NVIDIA's official announcement.
#NVIDIA #RTX6090 #RTX5090 #GeForce #GDDR7 #CES2027 #GraphicsCards
Variety would have enjoyed it if there was some Marvel characters in it.
“Digger” cratered in its second weekend with $3.2 million, a 60% decline from its dismal $8 million debut.
variety.com/2026/film/box-of…
Inference will prove to be a useless boondoggle and every GPU will turn to dust as soon as they hit their expiration date.
I think @TankerTrackers confirmed this as fake.
iFred retweeted
This is actually completely backwards, the largest customers (OAI/Ant/Meta) only take/highly prefer baremetal.
Managed k8s/slurm is great for small/medium labs. But even amongst these, you'd find a preference for baremetal.
The framing makes it sound like baremetal is easy; it's not.
to be a real neocloud you need to build software, not just supply power and GPUs
lots of people are buying sites and filling them with B300s
that’s cool, they’ll probably sell a few racks, but they’re gonna get blown out of the water by companies that build software
spot is fundamentally a software product. the platform is constantly shifting workloads around and cleaning up between jobs, and you’re solving many orchestration problems at scale
i’d be bullish on anyone doing spot or serverless like beam.cloud
and bearish on anyone buying power and GPUs and calling themselves a neocloud
I still this this is shifting for Rubin. We’re one quarter out.
🚨 I have TWO theories on why NVIDIA may be killing the RTX 5090 and moving GB202 entirely to RTX PRO. 👀
• Theory #1: NVIDIA may be trying to tighten the GB202 supply chain and reduce the chances of RTX 5090s being diverted into China
• One of the reports behind the RTX 5090 EOL rumor originated in Simplified Chinese, which makes the China angle worth watching
• Micro Center has already reportedly required some RTX 5090 buyers to show ID and sign a “NO EXPORT” declaration
• That declaration asks who the actual end user is, where the hardware will be installed and what it will be used for
• Theory #2: NVIDIA may have accidentally discovered just how valuable GB202 really is outside gaming
• The RTX 5090 launched at $1,999 but listings have climbed as high as roughly $9,600
• We’ve already seen reports of AI companies buying GeForce RTX 5090s by the pallet for AI servers
• At that point, why keep allocating GB202 dies to a $1,999 gaming product when professional AI customers could justify dramatically higher margins?
• The new rumor claims future GB202 production could be reserved entirely for RTX PRO
• None of this has been officially confirmed by NVIDIA yet
My theory is NVIDIA watched the RTX 5090 turn from a halo gaming GPU into an AI commodity and realized GB202 may simply be worth too much for GeForce.
If controlling where those chips ultimately end up is becoming of higher importance, killing the consumer GB202 product solves two problems at once. 🤔
With the GB300 cutoff, I’m assuming this is driven more by Rubin allocation than “fuck them kids” behaviors. Looking to be disappointed by the 10,000 dollar 6090 in January!
