HBM4 Memory Architecture
HBM4 Memory Architecture
HBM4 is the memory breakthrough that unlocks the next generation of AI inference and training — but the bigger story is the architectural shakeup that accompanies it. Three generations matter now: HBM4 (2026), HBM4E (2026–2027), and C-HBM4E (2026–2027). Two structural shifts make this the largest HBM change since the standard was introduced:
- Base dies move from DRAM to logic processes. HBM4 base dies are manufactured on TSMC 12FFC, N5, or N3P — not traditional DRAM nodes. This makes HBM base dies ~2× more power-efficient than HBM3E's and opens the door to embedding custom logic (C-HBM4E).
- Interface doubles to 2,048 bits with 32 channels per stack, directly targeting the AI memory wall.
Debuting in NVIDIA's Rubin GPU, HBM4 delivers 22 TB/s per GPU (2.8× over Blackwell's 8 TB/s HBM3e) with 288 GB capacity. HBM4E pushes this to 3 TB/s bandwidth per stack — 2.5× over HBM3E — at up to 12.8 GT/s demonstrated speeds. The Rubin Ultra GPU is spec'd for 1 TB of HBM4E across 8 stacks, yielding potential 16 TB/s per accelerator.
C-HBM4E is the real structural change. It retains standard HBM4E devices but allows custom base dies in three escalating levels: (1) logic integration on the base die, (2) custom die-to-die interface with memory controller on the base die, and (3) near-memory compute — basic processing capabilities inside the memory device itself. NMC requires topology-aware software, runtime, and OS evolution for heterogeneous memory domains — but reframes memory from a passive bottleneck to an active compute substrate. One source calls this "potentially the biggest shift in how computers work in decades."
Key Claims
- 22 TB/s bandwidth per GPU (HBM4, Rubin) — 2.8× over Blackwell. Evidence: strong (NVIDIA Rubin)
- 288 GB capacity per GPU (HBM4, Rubin) — 1.5× over Blackwell's 192 GB. Evidence: strong (NVIDIA Rubin)
- HBM4E: 3 TB/s per stack — 2.5× bandwidth over HBM3E. Evidence: strong (HBM4 Shakeup)
- 2,048-bit interface — doubled vs HBM3E's 1,024. Evidence: strong (HBM4 Shakeup)
- 32 channels per stack — doubled concurrency. Evidence: strong (HBM4 Shakeup)
- 12.8 GT/s demonstrated (Cadence PHY) — above spec. Evidence: strong (HBM4 Shakeup)
- Base dies on TSMC logic processes — 12FFC / N5 / N3P; 2× power efficient vs DRAM-based HBM3E. Evidence: strong (HBM4 Shakeup)
- C-HBM4E with near-memory compute — basic processing inside memory devices; requires topology-aware software. Evidence: moderate (HBM4 Shakeup)
- Rubin Ultra: 1 TB HBM4E, 16 TB/s potential — 8 HBM4 stacks per accelerator. Evidence: strong (HBM4 Shakeup)
- Memory subsystems with 48 TB/s bandwidth possible — with custom D2D interfaces. Evidence: moderate (HBM4 Shakeup)
- HBM3E 12-Hi 36GB carries ~2/3 of 2026 HBM shipments; HBM4 16-Hi 48GB targets Q4 2026 production. Evidence: moderate — January-2026 vendor outlook, now six months stale, and the raw ingest is a search-snippet summary (WebFetch blocked) (SK hynix outlook)
- SK Hynix ~60% HBM share; its 2026 capacity pre-booked by NVIDIA + OpenAI. Evidence: moderate — same January-2026 outlook; supersede with the 2026-07-29 print (SK hynix outlook)
- Memory is 40–50% of accelerator BOM (45–50% for a custom accelerator); HBM took 23% of DRAM wafer output in Q1–Q2'26; DRAM contract prices rose ~90% QoQ in Q1'26. Evidence: weak — TrendForce figures cited inside a single-author preprint; summary-derived twice over, unverified against memory-maker filings (Matsuoka)
- The HBM4 logic base die can serve as a stack-local control plane (cross-layer DMA, address translation, gather/broadcast, inline quantization) — the first C-HBM4E-class design with a named workload. Evidence: moderate — academic architecture proposal, abstract-level only (TokenStack)
Benchmarks & Data
| Metric | HBM3e (Blackwell) | HBM4 (Rubin) | HBM4E | Improvement HBM3E→HBM4E |
|---|---|---|---|---|
| Bandwidth per GPU | 8 TB/s | 22 TB/s | up to 48 TB/s (C-HBM4E) | 2.5× (stack) / 6× (system) |
| Per-pin speed | 9.4 Gbps | 8–12.8 GT/s | 12 Gbps (12.8 demo) | 1.3× |
| I/O width | 1,024-bit | 2,048-bit | 2,048-bit | 2× |
| Channels | 16 | 32 | 32 | 2× |
| Capacity per GPU | 192 GB | 288 GB | 1 TB (Rubin Ultra) | 5× |
| Base die process | DRAM | TSMC 12FFC/N5 | TSMC N5/N3P | Logic |
| Operating voltage | 1.1V | 0.75–0.8V | 0.679–0.963V | ~35% lower |
C-HBM4E Customization Levels
| Level | What's Custom | Example Benefit |
|---|---|---|
| 1 | Logic + caches on base die | Enhanced performance, standard interface |
| 2 | Custom D2D interface | More stacks per SoC, no package expansion |
| 3 | Near-memory compute (NMC) | Processing inside memory — "biggest shift in decades" |
Roadmap
| Variant | Availability | Status |
|---|---|---|
| HBM4 | 2026 | GUC PHY tape-out N3P (Mar 2025); silicon validation Q1 2026 |
| HBM4E | 2026–2027 | In development |
| C-HBM4E | 2026–2027 | In development |
Manufacturers
- DRAM: Micron (high-volume HBM4 for NVIDIA Vera Rubin), SK Hynix, Samsung
- Base dies: TSMC (12FFC / N5 / N3P)
- IP: GUC (PHY), Rambus (controller + C-HBM4E guidance), Cadence (12.8 GT/s PHY), Siemens EDA, Synopsys
2026 Supply Mix — SK Hynix Outlook (HISTORICAL, January 2026)
Staleness flag. The source below is a January 2026 vendor-outlook analysis and is now six months old. Treat it as the entry state of 2026, not current posture. SK Hynix reports on 2026-07-29 and Samsung on 2026-07-30 — both dates verified against primary filings this week — and those prints supersede everything in this section.
| Product | Stack | Per-stack capacity | Share of 2026 HBM shipments | Timing |
|---|---|---|---|---|
| HBM3E | 12-Hi | 36 GB | ~2/3 of total HBM shipments | Mainstream through 2026 |
| HBM4 | 16-Hi | 48 GB | Ramp begins H2 2026 | 16-Hi production targeted Q4 2026 |
- SK Hynix at ~60% of the HBM market by share; its 2026 HBM capacity reported pre-booked by NVIDIA + OpenAI — which means new entrants are gated on supply until 2027 regardless of design merit. See SK Hynix.
- Samsung + SK Hynix reportedly raised HBM3E prices ~20% heading into 2026 (Digitimes cross-reference).
- The headline "three market headwinds" in the source title were not captured in the ingest — the raw file records that WebFetch was blocked and its content is a search-snippet summary. That gap is unresolved.
The commercially important read is the mix, not the roadmap: the generation that actually ships the volume in 2026 is HBM3E 12-Hi, not HBM4. HBM4's bandwidth numbers above describe the leading edge; two-thirds of 2026 units are the prior generation.
Memory as a Priced Input, Not Just a Spec
HBM stopped being a bill-of-materials line and became the term that sets AI-buildout economics. The figures below travel with this page but are weakly sourced — they are TrendForce numbers cited inside a single-author preprint, i.e. summary-derived twice over, and none has been verified against a memory-maker filing by this KB:
| Data point | Value |
|---|---|
| DRAM contract price rise, Q1'26 vs Q4'25 | ~90% |
| Memory share of accelerator bill-of-materials | 40–50% |
| HBM share of DRAM wafer output, Q1–Q2'26 | 23% |
| HBM share of a custom accelerator's BOM | 45–50% |
| Meaningful new fab capacity arrival | 2027–2028 |
The consequence developed on Memory Scarcity & Inference Economics: because HBM is "priced by the same three suppliers under the same shortage" regardless of who designs the logic die, a custom accelerator removes merchant GPU margin but not the memory premium. DRAM→HBM wafer reallocation is treated there as the root cause of the broader commodity-DRAM surge — HBM allocation competes with commodity DRAM for the same wafer starts.
The Base Die Becomes a Controller
The HBM4 logic base die was introduced above as a power/efficiency story. The first concrete architecture to treat it as a control plane is TokenStack (arXiv:2605.05639), which uses the HBM4 base die as a stack-local controller handling cross-layer DMA, layered address translation, attention-side gather/broadcast coordination, and inline quantization during migration — with the stack split into dense capacity layers and PIM-enabled compute layers. That is C-HBM4E level-3 near-memory compute with a named workload attached (hot KV-cache attention) rather than a roadmap slot. See KV Cache Management and Processing-In-Memory. It is an academic proposal, not a product.
Analyst View — Memory Shortage (Attributed, Unverified)
Dylan Patel (SemiAnalysis), on Podcast Alpha (2026-07-10), characterizes the current memory tightness as structural rather than cyclical: "Memory is not a cycle top. It is a multi-year structural shortage." He frames the driver as a growth-rate mismatch — memory capacity grows 20-30%/year while AI demand doubles (~100%/year). Evidence: weak — Patel claims/estimates, unverified, no primary data cited, disclosed conflict (SemiAnalysis is a reported Anthropic enterprise customer) (Dylan Patel / Podcast Alpha).
This is directionally consistent with this page's own capacity/demand claims (HBM4/HBM4E roadmap tightness, TSMC logic-process base-die bottleneck) but Patel's specific growth-rate figures (20-30%/yr vs ~100%/yr) are his own estimate, not sourced to a named dataset — treat as a claim to test against primary capacity data (Micron/SK Hynix/Samsung disclosures), not as a settled number.
Open Questions
- What do the SK Hynix (2026-07-29) and Samsung (2026-07-30) prints say about HBM4 mix, ASP and capacity — and how far has the January outlook drifted?
- Do the ~90% DRAM price rise / 23% HBM wafer share figures hold up against primary memory-maker disclosures, or are they TrendForce estimates that have been laundered into fact by repetition?
- What are HBM4 yield rates and cost premiums versus HBM3e?
- Does Patel's 20-30%/yr capacity-growth-vs-~100%/yr-demand-growth framing hold up against primary DRAM-maker capex/output disclosures?
- Does C-HBM4E near-memory compute find a real workload beyond simulation benchmarks?
- Can ASIC vendors (TPU, Trainium) get HBM4/HBM4E at competitive timelines vs NVIDIA?
- What programming models emerge for NMC — CUDA extensions, or a new stack entirely?
- Does 1 TB HBM4E per accelerator change model-parallelism tradeoffs enough to shrink racks?
Related Concepts
- Rack-Scale AI Compute — HBM4 bandwidth enables rack-scale serving
- Custom Silicon vs GPU — memory access is a key differentiator in the ASIC vs GPU debate
- Nanosheet GAA Transistor — logic process nodes (N3P) are the new home for HBM base dies
- Processing-In-Memory — C-HBM4E level-3 NMC is the commercial wedge for memory-centric computing
- KV Cache Management — hot KV attention is the first named workload for base-die compute
- Memory Scarcity & Inference Economics — HBM price is the exogenous input driving those cost models
- Advanced Packaging & CoWoS Capacity — the second capacity gate on the same accelerator
Backlinks
Pages that reference this concept:
- NVIDIA Vera Rubin Platform
- Custom Silicon Inflection 2026
- HBM4 Architectural Shakeup
- SK hynix 2026 HBM Outlook
- TSMC
- SK Hynix
Changelog
- 2026-07-22 — Added the 2026 supply mix from the SK hynix January outlook (HBM3E 12-Hi ~2/3 of shipments, HBM4 16-Hi Q4'26, ~60% share, capacity pre-booked) with an explicit six-months-stale flag pointing at the 2026-07-29/30 prints; added the "memory as a priced input" block (BOM share, DRAM price rise, HBM wafer share) labelled weak/twice-summary-derived; added the HBM4 base-die-as-controller reading from TokenStack. +3 sources.
- 2026-07-14 — Added Patel/Podcast Alpha attributed "multi-year structural shortage" claim (capacity +20-30%/yr vs AI demand doubling) as an Analyst View — unverified, flagged for cross-check against primary DRAM-maker data.
- 2026-04-09 — Initial compilation from NVIDIA Rubin + SemiAnalysis.
- 2026-04-17 — Major update: added HBM4E (2.5× bandwidth, 2,048-bit interface, 32 channels), C-HBM4E (3 customization levels including near-memory compute), and TSMC logic-process base dies. Cross-linked to Nanosheet GAA Transistor.
Related Concepts
Advanced Packaging & CoWoS Capacity
Active FrontierCustom Silicon vs GPU
Active FrontierKV Cache Management
Active FrontierMemory Scarcity & Inference Economics ($/PB)
Active FrontierNanosheet GAA Transistor
Rack-Scale AI Compute
Active FrontierTheses that depend on this concept
These research positions cite this concept in their evidence. If the concept changes materially, these theses may need re-scoring.
Test Your Understanding
Hardware Concepts: The 2026 Substrate
Rack-scale compute, HBM4, custom silicon, quantum error correction, photonics, and memory-centric computing
Hardware Concepts Sprint
Fast recall on CoWoS, HBM4E, Willow's Λ, Nanosheet GAA, iceberg codes, neutral atoms, and PIM
Sourcing Frontier AI Compute in 2026
Make the chip, packaging, memory, and power calls the evidence supports