GDDR7 is the seventh generation of Graphics Double Data Rate memory, a type of DRAM designed for the very high bandwidth that graphics processors need. Standardized by JEDEC in 2024, it serves as VRAM on graphics cards and AI accelerators and introduces PAM3 signaling, which carries more data per clock than the binary signaling of earlier generations.
In graphics hardware, memory bandwidth often limits performance as much as shader throughput does. GDDR7 raises the per-pin data rate from the roughly 20 to 24 Gbps of mainstream GDDR6 to 28 to 32 Gbps in first products, with the standard designed to scale further, which lets a GPU feed its cores without widening the memory bus.
How GDDR7 transfers data with PAM3
Earlier GDDR generations use NRZ signaling, where each symbol on a wire is either high or low and carries one bit. GDDR7 uses PAM3, three-level pulse amplitude modulation, in which each symbol takes one of three voltage levels. Two consecutive symbols give nine combinations, enough to encode three bits per two symbols, or 1.5 bits per symbol.
This approach is a compromise. PAM4, used by GDDR6X, packs two bits per symbol but places the voltage levels closer together, which makes the signal more sensitive to noise. PAM3 keeps a wider gap between levels, giving better signal margins at high speed while still delivering roughly 50 percent more data per symbol than NRZ.
GDDR7 also reorganizes each memory device into four independent 8-bit channels instead of the two 16-bit channels of GDDR6. Smaller channels allow more parallel accesses, which improves efficiency when a GPU issues many small, scattered requests.
Faster signaling also demands better link training and equalization. The GPU memory controller and each GDDR7 chip calibrate timing and voltage levels when the card starts, and GDDR7 adds features that help the controller detect and correct transmission errors on the link. These mechanisms are part of the reason the memory can run reliably at speeds that would overwhelm a GDDR6 interface.
Key specifications of GDDR7
The following table lists the main GDDR7 parameters with typical values. The standard allows higher speeds and densities than those shipping in the first generation of products, so later cards are expected to use faster and larger chips on the same bus widths.
| Parameter | Typical value | Significance |
|---|---|---|
| Signaling | PAM3 | More data per symbol than NRZ |
| Per-pin data rate | 28 to 32 Gbps initially, roadmap toward 48 Gbps | Main driver of bandwidth gains |
| Channels per device | Four 8-bit channels | Better efficiency for small accesses |
| Device density | 16 Gb and 24 Gb, up to 32 Gb in the standard | 2 GB or 3 GB per chip |
| Core and I/O voltage | About 1.2 V | Lower than the 1.35 V of GDDR6 |
| Reliability features | On-die ECC, data poisoning, error check and scrub | Fewer silent memory errors |
The 24 Gb density is notable because it gives 3 GB per chip. On a 256-bit bus with eight chips, this allows 24 GB of VRAM instead of 16 GB, without adding chips or widening the bus.
Calculating GDDR7 memory bandwidth
Total memory bandwidth equals the per-pin data rate multiplied by the bus width, divided by eight to convert bits to bytes. Because each GDDR7 chip has a 32-bit interface, the bus width is simply the number of chips multiplied by 32.
| Bus width | Data rate | Bandwidth | Typical card class |
|---|---|---|---|
| 128-bit | 28 Gbps | 448 GB/s | Entry-level gaming |
| 192-bit | 28 Gbps | 672 GB/s | Mainstream gaming |
| 256-bit | 30 Gbps | 960 GB/s | Upper mid-range |
| 512-bit | 28 Gbps | About 1.79 TB/s | Flagship consumer GPU |
These figures explain why a narrower bus with GDDR7 can match or exceed a wider bus with GDDR6. Large on-chip caches on modern GPUs further reduce how often data must travel to VRAM, so effective bandwidth can be higher than the raw number suggests.
GDDR7 compared with GDDR6, GDDR6X and HBM
GPU designers choose memory based on bandwidth, cost, capacity and power. GDDR7 targets the same consumer and workstation market as GDDR6, while HBM serves data-center accelerators where cost matters less.
| Memory type | Signaling | Per-pin rate | Typical use |
|---|---|---|---|
| GDDR6 | NRZ | 14 to 24 Gbps | Mainstream graphics cards and consoles |
| GDDR6X | PAM4 | 19 to 24 Gbps | Upper-tier cards of one vendor |
| GDDR7 | PAM3 | 28 Gbps and higher | Current consumer and workstation GPUs |
| HBM3E | NRZ, very wide stacked bus | About 8 to 10 Gbps per pin | Data-center AI accelerators |
HBM reaches higher total bandwidth by using a 1024-bit interface per stack on a silicon interposer, but it is far more expensive to package. GDDR7 remains the cost-effective choice for products that sit on a standard printed circuit board.
Practical considerations and limitations
GDDR7 brings clear gains, but its real effect depends on the whole graphics design. Buyers and engineers should keep the following points in mind:
- Capacity still matters: fast memory does not help when a game or model exceeds the available VRAM.
- Bus width varies widely: a low-end card with GDDR7 on a narrow bus can have less bandwidth than an older high-end card.
- Board design is demanding: higher signaling rates require careful routing and signal integrity work on the circuit board.
- Thermals: memory chips running at high data rates need proper cooling pads and airflow.
- Workloads differ: AI inference and high resolutions benefit most, while CPU-limited games may show little change.
In plain words
A graphics card has its own fast memory, and GDDR7 is the newest and quickest common type. Think of the memory as a highway between storage and the graphics chip. GDDR7 does not add more lanes; it lets each lane carry more cars per second by using a smarter way of sending signals.
The practical result is that the graphics chip waits less for data, which helps with high resolutions, detailed textures and AI tasks.
A real-world example
A video editor upgrades from a graphics card with 16 GB of GDDR6 at 18 Gbps on a 256-bit bus, giving 576 GB/s, to a newer card with 16 GB of GDDR7 at 30 Gbps on the same bus width, giving 960 GB/s.
In 8K timeline playback with several color-grading layers and AI-based noise reduction, the new card keeps playback smooth where the old one dropped frames. Export times for a long project fall noticeably, mainly because the effects pipeline is limited by memory bandwidth rather than by compute.
The editor notes that capacity did not change, so the largest projects still approach the 16 GB limit. The upgrade solved the bandwidth bottleneck, not the capacity one.