Advertisement
AMD has confirmed its RDNA 4 “Radeon RX 8000” next-gen gaming GPUs and CDNA 3 “Instinct MI300” data center APUs in its latest roadmap.
AMD GPU Roadmap 2022-2024 confirms RDNA 3 for gaming chips “Radeon RX 8000” and CDNA 4 for “Instinct MI300” chips
While AMD unveiled the first details of its next-gen RDNA 3 GPUs, they also unveiled the generation of gaming GPUs that is coming after it, and surprisingly it’s called RDNA 4.
AMD RDNA 4 gaming GPUs are coming by 2024
Expected to launch in 2024, the new Navi 4x range will be based on an advanced process node. For its CPU lineup, AMD has announced both 4nm and 3nm nodes, so it could be any of those nodes, but I’d bet on the 4- set nm nodes which will make for some good marketing. AMD didn’t share numbers, but at least we know that RDNA 4 is a real thing and is coming after RDNA 3.
Rumor has it that RDNA 4’s MCM GPUs will include two different processing nodes similar to RDNA 3. The RDNA 4 architecture will power the Navi 4X line of GPUs, and the flagship chip, the Navi 41, was leaked a year ago. The information states that the RDNA 4 series, which will power the Radeon RX 8000 series graphics cards, will share a common architecture and will not split them into two different architectures as expected with RDNA 3.
There are reports that AMD’s RDNA 3 graphics architecture will only be featured on the Navi 31, 32 and possibly the Navi 33 GPU, while the rest of the lineup will be based on an RDNA 2 refresh. The top GPUs Navi 31 and Navi 32 RDNA 3 feature an MCM architecture and use a 5nm node for the GCD (Graphics Compute Die) and a 6nm node for the MCD (Multi-Cache/IO Die ). The RDNA 2 refresh GPUs are based on a 6 nm process node and are presented as a refresh within the new graphics family.
For RDNA 4, it looks like AMD plans to use the same graphics architecture across its Navi 4X lineup and no updates from older GPU architectures like RDNA 3 or RDNA 2. AMD will use two different nodes for its MCM chips, such as the Navi 41, namely 5nm and 3nm. The GCD is based on the 3nm process node while the MCD is based on the 5nm node. With RDNA 3 scheduled to launch in late 2022, the RDNA 4 lineup will launch sometime in 2024, competing against the successor to NVIDIA’s Ada Lovelace and Intel’s ARC Battle Mage or Celestial.
AMD RDNA Generational GPU Lineup
| Radeon Lineup | Radeon RX5000 | Radeon RX6000 | Radeon RX7000 | Radeon RX8000 |
|---|---|---|---|---|
| GPU architecture | RNA 1 | RDNA 2 | RDNA 3 / RDNA 2 | RNA 4 |
| process node | 7nm | 7nm | 5nm/6nm? | 5nm/3nm? |
| GPU family | Navi 1X | Navi 2X | GPS 3X | GPS 4X |
| flagship GPU | N / A | Navi 21 (5120 SPs) | Navi 31 (15360 SPs) | GPS 41 |
| High-end GPU | Navi 10 (2560 SPs) | Navi 22 (2560 SPs) | Navi 32 (10240 SPs) | GPS 42 |
| Mid-tier GPU | Navi 12 (2560 SPs) | Navi 23 (2048 SPs) | Navi 33 (5120 SPs) | GPS 43 |
| entry-level GPU | Navi 14 (1536 SPs) | Navi 24 (1024 SPs) | Navi 34 (2560 SPs) | GPS 44 |
AMD’s next generation cDNA 4 “Instinct MI300 APU” by 2023
David Wang also announced the Compute GPU roadmap, which includes Instinct-class chips for the AI and data center segments. It can now be confirmed that AMD is indeed working on a multi-chip and multi-IP Instinct accelerator that will not only feature the next generation of CDNA 3 cores, but also come with Zen 4 CPU cores of the equipped with the next generation. The Instinct MI300 GPU (technically an APU) is slated to launch by 2023.
As for the details, AMD will use the 5nm process node for its Instinct MI300 “CDNA 3” GPUs. The chip will be equipped with the next generation of Infinity Cache and features the 4th generation Infinity architecture, which enables the support of the CXL 3.0 ecosystem.
The Instinct MI300 accelerator will rock a unified memory APU architecture and new math formats that allow for a 5x performance per watt increase over CDNA 2, which is huge. AMD also predicts over 8x AI performance compared to the CDNA 2-based Instinct MI250X accelerators.
The CDNA 3 GPU’s UMAA connects the CPU and GPU to a unified HBM memory package, eliminating redundant memory copies while delivering a low total cost of ownership.
