Reported by TrendForce on . Summarized and analyzed by XAIPART Editorial.
What happened
According to Seoul Economic Daily, NVIDIA has asked Samsung, SK hynix and Micron to adopt TSMC's 3nm process for the base dies of its NVHBM memory, the custom HBM variant NVIDIA unveiled in August.S1 The push is intended to standardize production across suppliers so that end customers see consistent HBM performance and shorter product validation cycles.S1 Samsung's Device Solutions division reportedly signed off on the requirements after an internal review.S1
Industry estimates cited by Seoul Economic Daily put TSMC's 3nm yields above 90%, against roughly 60% for Samsung Foundry's 2nm (SF2) process, a gap that helps explain why NVIDIA is anchoring NVHBM base dies on TSMC's node.S1 NVIDIA has framed NVHBM's custom base die - which folds the memory controller usually located on the GPU into the HBM stack - as delivering up to 30% more memory bandwidth, 15% lower HBM power and 25% more XPU compute die area than standard HBM4E.S1
Samsung is hedging the dependency. The same Seoul Economic Daily report says Samsung is validating NVHBM base dies on its own 2nm process in case TSMC's 3nm lines become capacity-constrained.S1 For its in-house standard HBM4 (in mass production) and HBM4E (in validation), Samsung continues to use its own 4nm foundry node for base dies and 1c DRAM lines at Hwaseong and Pyeongtaek for core dies, and plans to keep a turnkey base-die path for custom HBM customers including AMD.S1
On the supply side, TSMC's October 15 earnings call flagged tight 3nm capacity; the node accounted for 33% of Q3 wafer revenue, its largest share of any process node.S1 Chairman and CEO C.C. Wei announced an additional 3nm fab in Tainan, on top of expansions already underway in Arizona, Kumamoto and Southern Taiwan Science Park, with mass production targeted for the first half of 2028.S1
| Metric | Reported value | Source |
|---|---|---|
| TSMC 3nm yield estimate | Above 90% | [S1] |
| Samsung SF2 (2nm) yield estimate | Around 60% | [S1] |
| NVHBM bandwidth uplift vs HBM4E | Up to 30% | [S1] |
| NVHBM power reduction vs HBM4E | 15% lower | [S1] |
| NVHBM XPU compute die area gain vs HBM4E | 25% more | [S1] |
| TSMC 3nm share of Q3 wafer sales | 33% | [S1] |
| Additional Tainan 3nm fab mass production | First half of 2028 | [S1] |
| Samsung foundry nodes for standard HBM4/HBM4E base dies | 4nm | [S1] |
Background timeline
- - TSMC's earnings call flagged tight 3nm capacity, with the node contributing 33% of Q3 wafer revenue, and announced plans for an additional 3nm fab in Tainan targeting first-half 2028 mass production.S1
- - TrendForce reports NVIDIA has asked Samsung, SK hynix and Micron to use TSMC 3nm for NVHBM base dies; Samsung's DS division has agreed to the requirements.S1
- not disclosed - NVHBM was unveiled by NVIDIA in August (date and venue not stated by source).S1
- not disclosed - Samsung and TSMC first collaborated roughly four years ago on TSMC GPU packaging compatibility for HBM; the current tie-up is the first extending into HBM manufacturing proper.S1
- not disclosed - Samsung's HBM4 entered mass production earlier in 2026, with HBM4E now undergoing production validation, both using Samsung Foundry 4nm base dies and 1c DRAM core dies at Hwaseong and Pyeongtaek.S1
XAIPART take
Technical impact
Moving the memory controller into the NVHBM base die shifts a critical analog/mixed-signal block from the GPU die to the HBM stack, which is why NVIDIA needs an advanced logic node at the base. The reported TSMC 3nm yield gap (above 90% vs around 60% on Samsung SF2) is a real engineering reason to standardize on TSMC, not just a sourcing preference. XAIPART analysis: buyers should expect base-die and GPU-die thermal/mechanical co-design to become a first-order constraint on NVHBM-based accelerator layouts, since the controller now lives next to DRAM rather than under the heatsink.
Market impact
TSMC's 3nm node is already its largest revenue contributor at 33% of Q3 wafer sales, and NVHBM base dies add a third major demand pool alongside AI accelerators and smartphone SoCs. Samsung keeping SF2 in validation is a credible hedge if TSMC allocates 3nm preferentially to higher-margin GPU/SoC wafers. XAIPART analysis: HBM buyers should model a two-supplier scenario (TSMC-base + Samsung-2nm-base) once Samsung completes validation, because dual sourcing is likely to become a contractual requirement for hyperscaler RFQs.
Action
Procurement teams pulling in HBM4, HBM4E or NVHBM should ask suppliers which foundry is producing the base die and at what node, since lead time and pricing will increasingly track TSMC 3nm wafer allocation rather than DRAM output alone. Design-in teams evaluating NVHBM should plan thermal and signal-integration margin around the integrated memory controller and confirm validation status for Samsung's 2nm fallback before locking the BOM. XAIPART analysis.
XAIPART Forecast
XAIPART expects at least one of the three NVHBM suppliers - Samsung, SK hynix or Micron - to publicly disclose NVHBM samples built on Samsung's SF2 base die by 2027-09-30 if Samsung's 2nm validation progresses on a typical 12-month cycle for a new base-die tape-out. XAIPART analysis
Facts are as reported by the cited source. The design-in, sourcing and forward-looking statements above are XAIPART's analysis, not the source's claims. Uncertainty: NVHBM performance numbers (30% bandwidth, 15% power, 25% compute die area) are NVIDIA-stated and not independently benchmarked in this report; yield estimates (TSMC 3nm >90%, Samsung SF2 ~60%) are industry estimates attributed by the source, not audited figures; SK hynix and Micron responses to NVIDIA's standardization request are not disclosed.
Corrections: none at publication. This brief will be updated with a timestamp if the source corrects its report.
Alternatives et cycle de vie
| Référence | Relation | Cycle de vie | Fondement |
|---|---|---|---|
| — | Aucune alternative référencée DONNÉES REQUISES | — | Nécessite match_type avant que toute mention de prédécesseur ne puisse apparaître |
Les mentions « prédécesseur » et « alternative » proviennent uniquement d'un type de correspondance enregistré, d'un statut de cycle de vie ou du texte source. Lorsqu'aucune correspondance n'est enregistrée, aucune mention n'est affichée. Nos propres annonces ne constituent pas une référence de marché.
Références concernées
Aucune correspondance dans le catalogue. Cette information ne correspond à aucune référence actuellement listée sur XAIPART. Nous ne déduisons pas de correspondances et n'affichons pas de liste fictive.
Pression sur les coûts par dimension
| Dimension | FRED series ID | Unit | Lag | As of |
|---|---|---|---|---|
| Input cost | PCOPPUSDM, PALUMUSDM, PNICKUSDM | USD/mt | 52 d | |
| Supply/demand | PCU334413334413, IPG3344S | index(1982=100) / index(2017=100) | 46 d | |
| FX & energy | DEXCHUS, DCOILBRENTEU, DHHNGSP | rate / USD/bbl / USD/MMBtu | 5 d |
No free authoritative interface for: gold · silver · rare earth · polysilicon · lithium carbonate · market share — listed rather than approximated.
Comment ce contenu a été produit
Traçabilité du processus — récupération, contrôle de la source, modèle et contrôle
| Source fetched | TrendForce · www.trendforce.com · 3,502 chars |
|---|---|
| Source sanity (S1) | ok — passed before any model call (no s1_ block recorded) |
| Model / prompt | MiniMax-M3 · prompt Blog Agent - News Summary (v2.8-final) · 7,330 tokens · 18.7s · run completed |
| Gate | verdict hold · 5/6 applicable checks passed · spec v2.9 · C10 ratio 0.029751 · C10 rounds 0 |
| Time contract | 3 <time datetime> values in this brief — 0 violations (+08:00 or date-only) |
Chaque ligne ci-dessus est produite par le pipeline, et non rédigée à la main.
Commentaires
Soyez le premier à commenter.
🔐 Veuillez vous connecter pour publier un commentaire.
Vous n'avez pas de compte ? vous inscrire