HBM3E vs HBM4: what is the difference?
HBM3E is the extended-speed version of HBM3 that went into volume production in 2024 and powers most current AI accelerators. HBM4, standardised by JEDEC as JESD270-4 in April 2025, is the next generation. The jump is larger than earlier ones because it widens the interface, not just the pin speed, and changes how the bottom die of each stack is made.
The short answer
HBM4 doubles the interface of each memory stack from 1024 to 2048 bits and the channel count from 16 to 32. At the JEDEC standard's 8 Gb/s per pin that gives about 2 TB/s per stack, against roughly 1.2 TB/s for HBM3E products running at about 9 to 10 Gb/s per pin. HBM4 also allows taller stacks of up to 16 dies and 64 GB, and moves the base die towards logic processes.
Side-by-side comparison
| Aspect | HBM3E | HBM4 |
|---|---|---|
| Standard | Speed extension of HBM3 (JEDEC JESD238 family), defined largely by vendor products | JEDEC JESD270-4, published April 2025 |
| Interface per stack | 1024 bits | 2048 bits |
| Channels per stack | 16 channels (32 pseudo-channels) | 32 channels (64 pseudo-channels) |
| Per-pin data rate | About 9.2 to 9.8 Gb/s in products | Up to 8 Gb/s in the standard; makers have announced faster parts |
| Bandwidth per stack | More than 1.2 TB/s | Up to about 2 TB/s at 8 Gb/s |
| Stack height | 8-high and 12-high | 4, 8, 12 and 16-high |
| Capacity per stack | 24 GB (8-high) and 36 GB (12-high) with 24 Gb dies | Up to 64 GB with sixteen 32 Gb dies |
| Base die | Mainly an I/O buffer made on a DRAM process | Increasingly made on a logic process, allowing customer-specific functions |
| Production timing | Volume production since 2024 | SK hynix announced completed development and mass-production readiness in September 2025; Samsung and Micron are following |
Wider, not just faster
Each HBM generation up to HBM3E gained bandwidth mainly by raising the per-pin rate on the same 1024-bit interface. HBM4 instead doubles the width to 2048 bits and the number of independent channels from 16 to 32. That lets it reach about 2 TB/s per stack at 8 Gb/s, a lower pin speed than HBM3E products already use, which eases signal integrity and power per bit.
The doubled width has a cost: twice as many connections between the stack and the processor. The interposer or bridge in the 2.5D package must route far denser wiring, and the processor's memory interface (PHY) must be redesigned. HBM4 stacks are therefore not drop-in replacements for HBM3E in an existing accelerator design.
The base die becomes a logic chip
At the bottom of every HBM stack is a base die that handles input and output and testing. Through HBM3E it was made on a DRAM-type process. For HBM4, memory makers are moving the base die to logic processes, either at a foundry or in their own logic fabs, and approaches differ between companies. A logic base die can use faster, lower-power transistors and can carry functions that a customer specifies, which is why custom HBM has become a topic for accelerator designers.
This changes the supply chain. An HBM4 stack may combine DRAM dies from a memory maker with a base die from a logic foundry, then be assembled in a package built by yet another company, so co-ordination between memory makers, foundries and accelerator designers matters more than before.
Taller stacks and the packaging challenge
HBM3E stacks of 12 dies are in volume production, and HBM4 allows up to 16. The overall stack height is constrained by the package, so more dies means thinner dies and thinner bonding layers. Current stacks are joined with micro-bumps using either mass-reflow molded underfill or thermo-compression with non-conductive film; hybrid bonding, which joins copper pads directly without bumps, is being evaluated for the tallest stacks.
Wafer thinning, through-silicon via formation, die-to-die bonding and known-good-die testing are the steps that decide HBM yield. Japanese suppliers are active here: DISCO's grinders and dicing saws thin and singulate the wafers, and Japanese companies supply underfills, mould compounds and test equipment used in HBM production.
FAQ
What is the difference between HBM3 and HBM3E?
HBM3E keeps the HBM3 architecture, including the 1024-bit interface and 16 channels, but raises the per-pin rate from the 6.4 Gb/s of the original HBM3 standard to about 9 to 10 Gb/s, and moves to 24 Gb dies and 12-high stacks. Per-stack bandwidth rises from about 0.8 TB/s to more than 1.2 TB/s.
Is HBM4 backward compatible with HBM3E?
Not at the package level. The 2048-bit interface needs a new processor PHY and denser interposer wiring. The JEDEC standard is designed so a controller can support both HBM3 and HBM4, but a given accelerator is designed around one generation.
Which accelerators use HBM3E and HBM4?
HBM3E is used in current AI accelerators such as NVIDIA's H200 and Blackwell-generation GPUs and AMD Instinct accelerators from the MI325X onward. HBM4 is designed for the following generation, including NVIDIA's Rubin platform.
Who makes HBM?
Three DRAM makers: SK hynix, Samsung and Micron. All three produce HBM3E and are developing or ramping HBM4.
Terms compared
- HBM3E (High Bandwidth Memory 3E)The extended version of HBM3 that entered volume production in 2024 and became the main memory of AI accelerators.
- HBM4 (High Bandwidth Memory 4)The fourth generation of stacked high-bandwidth DRAM, standardised by JEDEC in April 2025 with a 2048-bit interface per stack.
- HBM (High Bandwidth Memory)A memory made by stacking DRAM dies with through-silicon vias, giving a very wide interface placed next to a GPU or other processor.
- Through-Silicon Via (TSV)A copper-filled hole that passes vertically through a silicon die or wafer, connecting stacked dies over the shortest possible distance; the key interconnect of HBM and 3D ICs.
Japanese companies involved
- Advantest · Automated test equipment (ATE) for SoC and memory
- DISCO · Dicing, grinding and polishing equipment
- Tokyo Electron (TEL) · Coat/develop, etch and deposition equipment
More memory comparisons
Sources
Semiconductor comparisons
- All semiconductor comparisons
- SiC vs GaN: what is the difference?
- SiC vs silicon: what is the difference?
- Gallium oxide vs SiC: what is the difference?
- SiC MOSFET vs IGBT: what is the difference?
- CoWoS vs SoIC: what is the difference?
- CoWoS vs InFO: what is the difference?
- Chiplet vs monolithic: what is the difference?
Last updated: November 2, 2026