AMD's upcoming RDNA 3 GPU architecture will use a chiplet design and target a 50%+ performance-per-watt uplift versus RDNA 2; the first chips are coming in 2022
Ryan Smith / AnandTech :
Context & Ripple Effects
The 50% figure is becoming an AMD tradition: back at the Radeon RX 5700 Navi launch, the company pitched 50% better power efficiency over Vega, and RDNA 3 is now promising the same magnitude of step over RDNA 2. What is new is the mechanism — instead of relying on node shrinks alone, AMD is bringing its chiplet playbook from CPUs to graphics.
That playbook was proven on the CPU side earlier this year, when the Ryzen 7 5800X3D used 3D V-Cache chiplet stacking for a 15% gaming boost. AMD SVP Sam Naffziger's recent interviews elaborate on why the same disaggregation logic applies to GPUs, making this announcement less a surprise than the public confirmation of a direction AMD has been telegraphing all year.
First-order effects
- GPU buyers get a generation whose headline metric is performance-per-watt rather than raw peak throughput — directly relevant as power and cooling costs weigh on high-end card purchases.
- NVIDIA now faces a rival whose flagship architecture claims a 50%+ efficiency uplift, forcing the efficiency comparison onto NVIDIA's own roadmap rather than letting it compete purely on absolute frame rates.
Second-order effects
- A chiplet GPU changes AMD's supply economics: smaller dies mean better yields and the option to mix process nodes across a single card, which pressures competitors' monolithic-die cost structures.
- Packaging and interconnect suppliers gain a second major GPU customer for advanced stacking tech after AMD validated it with V-Cache, expanding demand beyond the CPU market.
Third-order effects
- If RDNA 3 ships as promised, chiplets stop being an AMD CPU experiment and become the company-wide design template — and likely an industry template — for scaling graphics beyond what monolithic dies can economically deliver.
- Efficiency-per-watt hardens into the primary competitive axis for consumer GPUs, mirroring how data center buyers already evaluate silicon.
The trend: GPU design is following CPUs toward chiplet-based architectures, with performance-per-watt displacing raw die size as the metric that defines a generation.