AMD’s patent has previously revealed that the company is working on one chiplet-based design for graphics processors, where several smaller circuits work together on a substrate. When the patent was published, speculation was raging about whether the technology might take place already in the next game-oriented architecture RDNA 3. Now several details are being sought out on the web, which indicates that this will be the case.
3DCenter sums up all the rumors, which are based on statements from the four famous leaks greymon55, KittyYYuko, kopite7kimi and Bondrewd. In terms of hardware resources, traditional raster performance seems to be able to increase by up to three times what a Radeon RX 6900 XT offers. The memory bus stops at relatively tame 256-bits, but is compensated by Infinity Cache, which is instead expected to increase up to two or even fourfold. AMD is also said to be shifting its focus from naming Computing Units (CU) in favor of Work Group Processors (WGP).
Circuit | Navi 21 | Navi 31 |
---|---|---|
Manufacturing node | TSMC 7 nm | TSMC 5/6 nm |
Type | Monolithic | Multi-Chip Module (MCM) |
Shader Engines | 4 st. | 6 st. |
Working groups (WGP) | 40 st. | 2 x 30 st. |
Raster units per workgroup | 128 st. | 256 st. |
Calculation units | 80 st. | 2 x 120 st. |
Kernels | 5 120 st. | 2 × 7 680 st. |
Memory bus | 256-bit | 256-bit |
Minnestyp | GDDR6 | GDDR6 |
Infinity Cache | 128 MB | 256–512 MB |
On paper, the Navi 31 looks to be a performance monster. One of the rumors, however, completely omitted information about is whether AMD has implemented any changes regarding Ray Accelerator, which are dedicated units in each CU unit to calculate ray tracing.
It also remains to be seen how well an extended Infinity Cache will scale, something that will be extra important as the design is rumored to continue to use regular GDDR6 memory. Regarding launch windows, the third or fourth quarter of 2022 is mentioned.