HBM3E and HBM4 news over the past six months has been dominated by SK Hynix and Micron. Samsung's straightforward push for 12- and 16-layer DRAM die stacks looks like a losing bet in the short term. So they're taking a different approach: zHBM (zoomed-in HBM). The prototype just surfaced, though community discussion around it has been conspicuously quiet.
Here's what we can confirm and what we can't.
How zHBM Differs from Conventional HBM
Conventional HBM works like this: DRAM dies are stacked layer by layer, connected through TSVs (through-silicon vias), and the entire stack sits next to a GPU or ASIC on the same package substrate, linked via a silicon interposer or more advanced RDL fan-out. The GPU core and HBM are neighbors, not layers.
The core change in zHBM in one sentence: stack the memory directly on top of the AI accelerator die.
This sounds like flipping HBM upside down, but the packaging logic is fundamentally different:
- Physical distance: Conventional HBM routes signals through an interposer, creating relatively long paths. zHBM uses vertical stacking, cutting memory-to-compute distance by an order of magnitude.
- Bandwidth potential: Shorter distances mean lower parasitic capacitance, enabling higher transfer rates at lower power or wider I/O configurations.
- Thermal design: A double-edged sword. Memory and logic chips pressed together create extreme heat density, forcing thermal solutions to be redesigned from scratch.
- Packaging process: No longer standard CoWoS or SoIC. zHBM requires more aggressive hybrid bonding or new TSV-less approaches.
The bet here is simple: packaging innovation is cheaper than process node upgrades. Each generation of 7nm, 5nm process technology raises costs by 30-50%, but a different packaging architecture might extract performance gains from more mature nodes.
Why Samsung Chose This Path Over Stacking Layers
Current state of the HBM race:
- SK Hynix: HBM3 ramped early; HBM3E is already shipping in volume to NVIDIA H200, leading in 12-layer stacking.
- Micron: HBM3E 24GB 12-Hi started sampling last year, mass production in mid-2024. Not getting the lion's share of B100/B200 but grabbing a piece.
- Samsung: HBM3 ramped late, HBM3E still catching up, 12-Hi trailing competitors.
Samsung is 1-2 quarters behind in the main HBM arena, and that gap can't be closed by stacking more layers alone. Yields, capacity, and customer validation all take time.
But Samsung holds one card: advanced packaging.
Samsung has built expertise in FOPLP (fan-out panel-level packaging) and I-Cube (2.5D packaging), which are relative weak spots for SK Hynix and Micron. By going the packaging route with zHBM, Samsung effectively shifts the battlefield from "who stacks more DRAM dies" to "who can stitch memory and compute together more tightly."
Whether this approach can sidestep JEDEC HBM interface standards (the defined physical layer and protocol) is a separate question. Conventional HBM must comply with HBM3/HBM4 specs to be adopted by NVIDIA or AMD GPUs directly. If zHBM is custom-ASIC-only, it won't need JEDEC compliance, but the customer pool shrinks dramatically.
It's a Prototype — Don't Get Too Excited
What we can confirm:
- Samsung did unveil a zHBM prototype (source: thelec.net reporting)
- The packaging architecture and stacking method differ from conventional HBM
- It's positioned as AI accelerator-specific
What we can't confirm:
- Production timeline — not disclosed in the source; industry analysts estimate 2026-2027 at the earliest
- Customer list — zero public design wins
- Specific bandwidth numbers — none released; likely "works in the lab" territory, not production specs
- Yield and cost — meaningless to discuss at prototype stage
- Whether it can actually bypass HBM interface standards — technical details remain undisclosed
In other words, any conclusion along the lines of "Samsung is about to overtake everyone" or "HBM is being disrupted" is just PR talking-point regurgitation. What we're seeing is an engineering signal, not a product signal.
The Community Is Cold, But It Shouldn't Be Ignored
The original Hacker News thread sits at 16 points with 3 comments. That's unusually low engagement. Possible reasons:
- The source is a paywalled Korean/English mix with low information density
- "Samsung has another new prototype" is a high-frequency event for longtime followers, offering little marginal information
- Memory packaging stories naturally fail to resonate with software engineers
But the underlying issue, the memory wall, is a shared bottleneck for both training and inference of large models.
Look at the numbers: H100's HBM3 bandwidth is 3.35 TB/s, but during LLM inference, KV cache access patterns routinely push bandwidth utilization above 80%. For a GPT-4-class model, a single inference call can require moving hundreds of GB of KV cache data. Without enough bandwidth, even powerful GPUs end up idle.
Any attempt to shorten memory-to-compute distance at the packaging level can theoretically offset part of the memory wall. Even a 30% improvement in bandwidth density would have structural implications for inference cost.
So this story deserves tracking, even if the current hype is low.
If zHBM Works, Who's the First Customer
Speculating on potential customer profiles, ranked by likelihood:
Most likely: LLM inference chip vendors
- Groq's LPU architecture already goes the SRAM-on-chip route, starving for low-latency memory
- Cerebras's wafer-scale engine already takes an extreme "compute on memory" philosophy, aligned with zHBM's thinking
- Dataflow architecture vendors like SambaNova and Tenstorrent are also potential customers
Next likely: HPC and custom ASIC cloud vendors
- AWS Trainium2 / 3 has already validated custom silicon viability; if zHBM can reduce bandwidth costs, AWS has motivation to experiment
- Google TPU already has established HBM supply relationships; switching packaging architecture carries high migration costs
- Microsoft Maia and Meta MTIA are both potential customers, especially vendors sensitive to inference TCO
Least likely: traditional GPU vendors
- NVIDIA won't change socket and packaging design in the short term; HBM supply still comes from SK Hynix
- AMD MI300/MI325 sticks with standard HBM; packaging architecture is already locked in
This suggests zHBM's market positioning is closer to a "memory solution for custom accelerators" than a "universal HBM replacement." The market ceiling is lower, but per-customer revenue could be higher because custom solutions command premium margins.
Editor's Take
This story is being underestimated because packaging innovation has stayed in the background of the AI compute narrative. For the past decade, process nodes (7nm, 5nm, 3nm) got the most attention, architectures (GPU vs TPU) got second billing, and packaging barely registered.
But the weight of the packaging layer is rising. Three reasons:
- Process node upgrades deliver diminishing returns — 3nm to 2nm gains only 10-15% performance while costs jump 30%+
- Chiplets turn "assembly" into the mainstream — single dies no longer chase size, making the packaging layer the critical bottleneck
- The memory wall problem amplifies in the LLM era — inference and training are both starving for bandwidth
zHBM isn't replacing HBM; it's giving custom chips another option. Same logic as NVIDIA acquiring Mellanox or AMD acquiring Pensando: the second half of the compute competition plays out at the system level, not the chip level.
My take: if Samsung can deliver a production-ready zHBM by end of 2026 and secure at least one design win with a major AI chip vendor, packaging will become the next hyped battleground. Otherwise, this story will likely fade into the RSS feeds of industry watchers, joining the long list of "Samsung HBM revolution" announcements from years past.
Unresolved concern: zHBM's thermal design is a major open question. With memory and logic chips pressed together, power density could exceed 200W/cm². Whether existing cooling solutions can handle it remains completely undocumented. That's a potential physical constraint that could collapse the entire plan, worth continued attention.