TL;DR
- Micron-developed graphics memory standard introduced 2020 using PAM4 four-level signalling.
- Per-pin speeds up to 24 Gb/s; common bus widths of 256-384 bits.
- Used in NVIDIA GeForce RTX 3080/3090/4090 and certain prosumer/data centre cards.
- Higher bandwidth-per-pin than GDDR6 but more thermal and power demanding.
Overview#
GDDR6X is Micron's proprietary extension of the JEDEC GDDR6 standard. It introduces PAM4 (Pulse Amplitude Modulation with four levels) signalling — each pin transmits two bits per cycle rather than one — doubling effective bandwidth versus equivalent NRZ-signalled GDDR6.
Commercially it is best known from NVIDIA's GeForce RTX 30 and 40 series (Ampere and Ada consumer flagships). In the data centre it appears in select cards; the dominant choices for data centre AI cards are HBM3/HBM3e for high-end parts and GDDR6 (non-X) for mid-range.
Specifications#
| Metric | GDDR6X |
|---|---|
| Signalling | PAM4 (4-level) |
| Pin speed | Up to 24 Gb/s (effective) |
| Typical bus width | 256-384 bits |
| Typical card bandwidth | ~700 GB/s to ~1 TB/s |
| Vendor | Micron (proprietary) |
| Voltage | 1.35 V |
| Typical use | GeForce RTX 30/40 series |
Architecture Notes#
PAM4 signalling allows higher symbol rates but lower per-symbol amplitude margin, complicating equalisation and error correction. Micron uses a dedicated controller and signalling stack; GDDR6X is not interoperable with standard GDDR6 controllers without explicit support.
Thermal envelopes are elevated relative to GDDR6 — early RTX 3000 launches had documented memory thermal concerns; later boards added explicit memory cooling.
When to Pick GDDR6X (or Cards Using It)#
- Consumer / prosumer workstations and AI dev boxes where RTX-class cards are the right form factor.
- Mid-tier inference workloads where 24 GB GDDR6X (RTX 4090, etc.) is sufficient.
- Single-card deployments — multi-card GDDR6X setups don't have NVLink.
- For data centre AI — HBM-based cards (L40S, A100, H100) are the recommended path.
Pitfalls#
- Memory thermals require attention in dense deployments.
- Proprietary to Micron; supply is single-sourced.
- ECC support depends on card-level implementation; consumer cards typically omit it.
References
- Micron GDDR6X Product Information · Micron