Beta
MemoryReading · ~3 min · 56 words deep

GDDR7

GDDR7 is the newest graphics DRAM · 32-40 Gbps per pin · used in consumer GPUs (RTX 50-series) and pro cards for inference.

TL;DR

GDDR7 is the newest graphics DRAM · 32-40 Gbps per pin · used in consumer GPUs (RTX 50-series) and pro cards for inference.

Level 1

GDDR7 is the successor to GDDR6X, used primarily in consumer + prosumer GPUs. Speeds: 32 Gbps at launch (RTX 5090, RX 9000 series), 36-40 Gbps in 2026 refreshes. Adopts PAM3 signaling (3-level) for higher bandwidth at lower power vs GDDR6X's PAM4. Deployed on consumer GPUs for gaming and ML/AI enthusiasts.

Level 2

GDDR7 bandwidth per card: RTX 5090 hits 1,792 GB/s (512-bit bus × 28 Gbps). Compared to HBM3e at ~4 TB/s (server-class), GDDR7 is cheaper per GB but lower bandwidth. Sweet spot: inference on consumer + prosumer GPUs running 7-32B models. Samsung, SK Hynix, and Micron produce GDDR7. Capacity per die: 3GB, enabling 24-48GB on consumer cards.

Level 3

GDDR7 matters for AI inference at the edge and prosumer level. RTX 5090 with 32GB GDDR7 can serve a 70B 4-bit model interactively. Workstation cards (RTX Pro 6000 Blackwell) with 96GB GDDR7 hit a middle ground between consumer + data center. Micron 3GB GDDR7 die enables 48GB on consumer 384-bit bus cards · a niche but growing segment.

The takeaway for you
If you are a
Researcher
  • ·PAM3 signaling · 32-40 Gbps per pin
  • ·Consumer + prosumer GPU memory
  • ·Samsung, SK Hynix, Micron supply
If you are a
Builder
  • ·Consumer AI inference on 5090-class cards
  • ·32-48GB capacities unlock 70B+ local inference
  • ·Cheaper per GB than HBM but lower bandwidth
If you are a
Investor
  • ·Volume consumer GPU memory · large but margin-thin
  • ·Prosumer AI segment driving 32-48GB demand
  • ·Complements HBM · different market
If you are a
Curious · Normie
  • ·Fast memory in gaming + consumer AI graphics cards
  • ·Makes GPUs run AI and games fast
  • ·Not the same as AI data-center memory
Gecko's take

GDDR7 is the consumer-AI memory · RTX 5090 at 32GB is the enthusiast inference floor for 2026.

HBM3e has 3-5× bandwidth but is on-package, expensive. GDDR7 is DIMM-like, cheaper, lower bandwidth. Consumer vs server split.