GDDR is specialized RAM optimized for the needs of video cards. Unlike system DDR, GDDR is capable of transferring gigantic volumes of data in parallel, sacrificing latency for the enormous bandwidth required for rendering textures and complex 3D scenes.
Description
| Simple description | Imagine a huge high-speed highway with dozens of lanes instead of a narrow city street. GDDR allows you to haul data in a wide stream, which is ideal for video cards where you need to simultaneously and quickly load heavy textures and geometry at high speed, sacrificing the first car's reaction time. |
| Mid-level description | A high-bandwidth synchronous graphics dynamic random-access memory (SGRAM) optimized for GPUs. It differs from system DDR by a wider bus and higher frequencies, providing a multiple increase in bitrate, but at the cost of higher latency that is offset by the massively parallel architecture of the GPU. |
| Technical description | A specialized SGRAM standard based on DDR technology with a multiplexed bus, using differential signaling (POD signaling) and QDR modes. The microarchitecture includes 16n/32n data prefetch. GDDR6 employs pseudo-channels that split a 32-bit array into two 16-bit ones, boosting efficiency. The interface is x32/x16 with double or quadruple data rate to achieve bandwidth exceeding 1 TB/s on top-tier chips. |
Technical Features
| Top-level category | Memory | Subcategory | Graphics Dynamic RAM |
| Developer/Author | ATI Technologies (originator), JEDEC (standardization) | Standardizing body | JEDEC |
| Entity type | Hardware component / DRAM subtype | Predecessor | SDR SGRAM, DDR SGRAM |
| Clock frequency | From ~1.2 GHz (GDDR5) to ~2.4 GHz and higher (GDDR6X) | Bandwidth | From ~64 GB/s to >1 TB/s per device |
| Data width / Block size | 32-bit wide interface per chip | Supply voltage | Decreased from 1.5V to 1.1–1.35V (PAM signaling) |
| Topology | Point-to-point (P2P) controller-to-chip connection | Error correction mechanism | ECC support at controller level, channel failure detection |
| Physical connector/interface | BGA soldering on the PCB, GPU memory bus | Cooling method | Passive heatsink or active airflow within the graphics card |
Advantages and Limitations
| Advantages | Extremely high bandwidth when working with large arrays of pixels and textures. Excellent energy efficiency per transferred watt due to low-voltage signaling. Specific features like address deduplication and write masking align perfectly with graphics pipeline workloads. |
| Limitations | High clock latencies (CAS Latency) compared to system DDR of the same generation. Poorly suited for tasks dominated by random small-block access (e.g., classic CPU computations). Structurally soldered and not user-expandable. |
Application Areas
| Product/Application area | Discrete gaming and professional graphics cards (NVIDIA GeForce RTX, AMD Radeon RX) |
| Product/Application area | Game consoles (SoC chips of PlayStation 5, Xbox Series X|S use GDDR6 as unified memory) |