GDDR (Graphics Double Data Rate)

Overview

GDDR is a specialized type of DRAM designed for GPUs and high-bandwidth applications. While DDR is optimized for low latency (CPU workloads), GDDR is optimized for high bandwidth (GPU workloads). Modern GDDR6 and GDDR6X achieve significantly higher bandwidth than DDR5 through wider interfaces and higher clock speeds.

GDDR vs DDR

Key insight: GDDR trades latency for bandwidth. GPUs need massive bandwidth (hundreds of GB/s), not low latency.

GDDR Generations

GDDR3 (2004)

• Data rate: up to 3.2 Gbps/pin
• Used in: Xbox 360, PS3, older GPUs

GDDR5 (2008)

• Data rate: up to 8 Gbps/pin
• Voltage: 1.35-1.5V
• Used in: NVIDIA GTX 10 series, AMD RX 500 series

GDDR5X (2016)

• Data rate: up to 14 Gbps/pin
• PAM4 signaling (first in consumer memory)
• Used in: NVIDIA GTX 1080 Ti

GDDR6 (2018)

• Data rate: up to 24 Gbps/pin
• Voltage: 1.35V
• Two 16-bit channels per chip
• Used in: NVIDIA RTX 30 series, AMD RX 6000 series

GDDR6X (2020)

• Data rate: up to 24 Gbps/pin (effective 48 Gbps with PAM4)
• PAM4 signaling: 4 voltage levels per symbol = 2 bits per symbol
• Used in: NVIDIA RTX 3090, RTX 4090

PAM4 Signaling (GDDR6X)

Traditional NRZ (Non-Return-to-Zero) uses 2 voltage levels:

NRZ: High = 1, Low = 0 → 1 bit per symbol

PAM4 (Pulse Amplitude Modulation, 4 levels) uses 4 voltage levels:

PAM4: Level 3 = 11, Level 2 = 10, Level 1 = 01, Level 0 = 00 → 2 bits per symbol

graph LR
    subgraph "NRZ (2 levels)"
        N0["Low = 0"]
        N1["High = 1"]
    end
    subgraph "PAM4 (4 levels)"
        P0["Level 0 = 00"]
        P1["Level 1 = 01"]
        P2["Level 2 = 10"]
        P3["Level 3 = 11"]
    end

Benefit: Double the data rate at the same symbol rate. Cost: Tighter voltage margins → more susceptible to noise → requires better signal integrity.

GDDR6 Memory Subsystem

graph TD
    subgraph "GPU"
        MC["Memory Controller"]
    end
    subgraph "Memory Channels"
        Ch0["Channel 0<br/>32-bit"]
        Ch1["Channel 1<br/>32-bit"]
        Ch2["Channel 2<br/>32-bit"]
        Ch3["Channel 3<br/>32-bit"]
    end
    subgraph "GDDR6 Chips"
        C0["Chip 0<br/>16-bit × 2 channels"]
        C1["Chip 1"]
        C2["..."]
        C11["Chip 11"]
    end
    MC --> Ch0
    MC --> Ch1
    MC --> Ch2
    MC --> Ch3
    Ch0 --> C0
    Ch0 --> C1
    Ch1 --> C2
    Ch3 --> C11

RTX 3090 Example

• Memory: 24 GB GDDR6X
• Bus width: 384-bit
• Data rate: 19.5 Gbps/pin
• Bandwidth: 384 × 19.5 / 8 = 936 GB/s

RTX 4090 Example

• Memory: 24 GB GDDR6X
• Bus width: 384-bit
• Data rate: 21 Gbps/pin
• Bandwidth: 384 × 21 / 8 = 1008 GB/s

GDDR6 vs HBM

Power Consumption

GDDR is power-hungry due to high clock speeds and wide interfaces:

GDDR6 power per chip: ~4-7W
GDDR6X power per chip: ~7-12W

Total memory power:
RTX 3090: 12 chips × ~8W = ~96W (memory alone!)
RTX 4090: 12 chips × ~10W = ~120W

This is why GPUs have massive power budgets (300-450W).

Interview Questions

1. Q: Why do GPUs use GDDR instead of DDR? A: GPUs need massive bandwidth to feed thousands of cores, not low latency. GDDR provides much higher bandwidth per chip (16-48 GB/s vs 8 GB/s for DDR5) through wider interfaces and higher clock speeds. The latency penalty doesn’t matter much for GPU workloads (thousands of threads hide latency).

2. Q: What is PAM4 signaling in GDDR6X? A: PAM4 uses 4 voltage levels instead of 2 (NRZ), encoding 2 bits per symbol instead of 1. This doubles the data rate at the same symbol frequency, but requires tighter voltage margins and better signal integrity.

3. Q: Calculate the bandwidth of a 384-bit GDDR6X bus at 21 Gbps/pin. A: Bandwidth = 384 pins × 21 Gbps / 8 bits = 1008 GB/s ≈ 1 TB/s.

4. Q: Why is GDDR power consumption so high? A: High clock speeds (up to 21 Gbps/pin), wide interfaces (384-bit), and PAM4 signaling all consume significant power. Each GDDR6X chip can consume 7-12W, and a GPU may have 12 chips.

Common Mistakes

• ❌ Confusing GDDR with DDR (different optimization targets)
• ❌ Not knowing PAM4 signaling in GDDR6X
• ❌ Assuming GDDR has lower latency than DDR (it’s actually higher)
• ❌ Forgetting that GPU bandwidth is much higher than CPU bandwidth

Summary

GDDR is optimized for bandwidth over latency, making it ideal for GPUs. GDDR6 uses NRZ signaling up to 24 Gbps/pin; GDDR6X uses PAM4 for up to 48 Gbps effective. A 384-bit GDDR6X bus can achieve ~1 TB/s bandwidth. The trade-off is higher power consumption and latency compared to DDR.

Cross-References

• DRAM — Base DRAM technology
• HBM — Stacked alternative
• GPU — Why GPUs need high bandwidth
• Memory Hierarchy — Where GDDR fits

Cross References

• DDR
• HBM
• GPU