HBM vs GDDR6: what is the difference?
Graphics processors and AI accelerators are often limited by how fast data can be moved in and out of memory rather than by compute. HBM (High Bandwidth Memory) and GDDR6 (Graphics Double Data Rate 6) are the two main answers. Which one a chip uses decides its packaging, cost structure and supply chain as much as its performance.
The short answer
Both are DRAM built for high bandwidth, but they get it in opposite ways. HBM stacks several DRAM dies with through-silicon vias and places the stack beside the processor on an interposer, giving a 1024-bit interface per stack at a moderate pin speed; GDDR6 uses separate 32-bit chips soldered on the circuit board around the processor and runs each pin much faster. HBM delivers more bandwidth per watt and per board area, while GDDR6 is cheaper and simpler to build in.
Side-by-side comparison
| Aspect | HBM | GDDR6 |
|---|---|---|
| Physical form | Stack of thinned DRAM dies on a base die, connected by through-silicon vias | Individually packaged DRAM chips |
| Interface width | 1024 bits per stack (HBM3 and HBM3E); 2048 bits in HBM4 | 32 bits per chip, organised as two 16-bit channels |
| Per-pin data rate | Moderate: 6.4 Gb/s in the HBM3 standard, around 9 to 10 Gb/s in HBM3E products | High: typically 14 to 20 Gb/s per pin |
| Bandwidth per device | About 0.8 TB/s per HBM3 stack and above 1.2 TB/s per HBM3E stack | About 56 to 80 GB/s per chip |
| Connection to processor | Short traces through a silicon interposer or bridge in a 2.5D package | Traces on the printed circuit board |
| Energy per bit transferred | Lower, thanks to very short, slower wires | Higher, because fast signals travel over board traces |
| Board area | Small: memory sits inside the processor package | Large: chips surround the processor on the board |
| Cost and complexity | High: TSVs, stacking, known-good-die testing and advanced packaging | Low: standard packages and conventional board assembly |
| Typical uses | AI training and inference accelerators, HPC processors, high-end networking | Consumer graphics cards, game consoles, some inference cards and networking equipment |
How HBM reaches its bandwidth
HBM trades pin speed for width. Each stack has 1024 data lines in HBM3 and HBM3E, far more than a circuit board can route, so the stack is placed millimetres from the processor and wired through a silicon interposer or embedded bridge, as in TSMC's CoWoS family of packages. Because the wires are short, each pin can run at a moderate rate with low energy per bit, and several stacks around one processor add up to many terabytes per second.
The cost is manufacturing complexity. DRAM wafers must be thinned to a few tens of micrometres, filled with thousands of through-silicon vias, tested as known-good dies and bonded into stacks of 8 or 12 dies, then integrated in a 2.5D package. Yield losses at any of these steps are expensive, and packaging capacity has at times limited how many accelerators could be built.
How GDDR6 reaches its bandwidth
GDDR6, standardised by JEDEC as JESD250, goes the other way. Each chip has a 32-bit interface split into two independent 16-bit channels, and every pin runs fast. A GPU reaches high total bandwidth by placing many chips around the die: a 384-bit bus built from twelve chips at 20 Gb/s per pin gives 960 GB/s.
Because the chips are standard packages soldered to the board, GDDR6 needs no interposer, no stacking and no special assembly. That keeps cost down and makes supply flexible, which is why it dominates consumer graphics. The penalty is power: driving very fast signals across board traces costs more energy per bit, and the chips take up board area. Micron's GDDR6X and the newer JEDEC GDDR7 standard push per-pin rates further using multi-level signalling.
Which to use when
When memory bandwidth per package and per watt is the limiting factor, as in training large AI models or memory-bound HPC codes, HBM is the usual choice despite its cost. When cost per gigabyte and supply flexibility matter more, as in gaming graphics cards and many inference or networking products, GDDR remains the practical option. Capacity also differs: HBM stacks put a large amount of memory inside the package, while GDDR capacity grows with the number of chips on the board.
In Japan, suppliers serve both paths: DISCO's grinding and dicing tools are used to thin and singulate HBM wafers, and Japanese firms supply substrates, underfills and other packaging materials.
FAQ
Is HBM faster than GDDR6?
Per device and per package, yes. One HBM3E stack delivers more than 1.2 TB/s, while one GDDR6 chip delivers well under 100 GB/s. Each HBM pin is actually slower than a GDDR6 pin; HBM wins because it has many times more pins per device.
Why do gaming graphics cards not use HBM?
Cost and packaging. HBM needs an interposer-based package and expensive stacked dies, which would raise the price of a consumer card considerably. Some professional and earlier consumer GPUs did use HBM2, but GDDR has remained the standard for gaming.
Are HBM and GDDR6 both DRAM?
Yes. Both store data in standard 1T1C DRAM cells and are made by the same three companies: Samsung, SK hynix and Micron. They differ in interface design, die layout and packaging, not in the memory cell.
What does GDDR7 change?
GDDR7, published by JEDEC in 2024, uses three-level PAM3 signalling to carry more data per pin than GDDR6. It narrows the per-chip gap but keeps the same basic approach of separate chips on the board, so HBM still offers far more bandwidth per package.
Terms compared
- 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.
- DRAM (Dynamic Random-Access Memory)Volatile memory that stores each bit as charge on a tiny capacitor and must be refreshed; the main memory of computers, phones and servers.
- 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.
- CoWoS (Chip on Wafer on Substrate)TSMC's 2.5D packaging technology that places logic dies and HBM stacks side by side on an interposer; the standard package for AI accelerators.
Japanese companies involved
- DISCO · Dicing, grinding and polishing equipment
- Micron Memory Japan (Micron Hiroshima) · DRAM fab in Hiroshima (Micron Technology)
- Tokyo Electron (TEL) · Coat/develop, etch and deposition equipment
- Ajinomoto · ABF insulating film for package substrates
- IBIDEN · IC package substrates (FC-BGA)
More memory comparisons
Sources
Semiconductor comparisons
- All semiconductor comparisons
- HBM3E vs HBM4: what is the difference?
- 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?
Last updated: November 2, 2026