Radeon RX 7000 & MI300: RDNA 3 comes with chiplets, but only CDNA 3 stacks them

0
163

For Financial Analyst Day 2022, AMD not only announced the IPC increase from Zen 3 to Zen 4, but also the first details about RDNA 3 as the basis of Radeon RX 7000 and CDNA 3 as a basis the HPC accelerator called Instinct MI300. A chiplet design has been confirmed for RDNA 3, CDNA 3 even puts the GPUs on the Infinity Cache.

RDNA 3 uses chiplets, but stacks they don't

The fact that AMD relies on a multi-chip approach for RDNA 3 has been in the rumor mill since summer 2021. But it had to be June 2022 for AMD to confirm exactly that. AMD calls this technology “Advanced Chiplet Packaging”. AMD, on the other hand, left it open which chiplets are used in which RDNA 3 GPU – and so the coming weeks should again see a wide variety of reports on the number of compute and infinity cache chiplets making the rounds through the rumor mill.

What AMD has confirmed is that RDNA 3, like Zen 4, will be manufactured in a 5nm process, using new compute units and an overhauled rendering pipeline. The fast Infinity Cache is also said to have been launched in a new generation. It remains to be seen whether this means more than outsourcing it to separate chips that also contain the GDDR6 memory controller.

All of the manufacturing, packaging, and architecture changes are designed to give RDNA 3 over a 50 percent increase in performance over RDNA 2 for the same power consumption.

Image 1 of 2

RDNA roadmap as of June 2022
50 percent more performance with RDNA 3

AMD also mentioned AV1 encoding capabilities and DisplayPort 2.0 compatibility as additional features of the RDNA 3 architecture on the audio track. With Navi-4x GPUs based on RDNA 4, the successor to the RDNA 3 architecture should appear by 2024, but AMD didn't say another word about it.

CDNA 3 puts two GPUs on the infinity cache

AMD also spoke for the first time about the next generation of the HPC GPU architecture CDNA 3, whose current expansion stage CDNA 2 on the Instinct MI250(X) graphics cards will equip the world's fastest supercomputer at the end of the year. CDNA 2 already relies on a multi-chip package with two GPUs and HBM on one package, but still lacks Infinity Cache. This will change with CDNA 3.

AMD has confirmed that CDNA 3 based on the 4th generation Infinity Fabric will be the first to stack chips of a GPU. The technology is referred to as “3D Chiplet Packaging”. The Infinity Cache is placed on the “base layer” like the HBM memory, the GPUs – again two according to the figure – sit above it.

CDNA 3 with 3D chiplet packaging (Image: AMD)

With the announcement of stacking technology for CDNA 3, it is also a certainty that RDNA 3 will not rely on stacked chips just yet. The fact that RDNA relies on new packaging technologies a generation later than CDNA could become the rule in the future.

Like RDNA 3 and Zen 4, CDNA 3 should also have a 5- nm process and AMD promises serious performance advantages for Instinct MI300 products based on it for AI workloads by a factor of more than 5 compared to CDNA 2. However, this is also necessary if the next generation is to at least compete against Nvidia Hopper – the GH100 -GPU is much stronger than CDNA 2 in AI workloads.

The Zen 4 CDNA 3 APU is coming

AMD has also confirmed an APU based on a Zen 4 CPU and a CDNA 3 GPU a unified memory interface – i.e. the CPU and GPU share the HBM memory. Rumors about it had surfaced in May. CDNA 3 should be released by the end of 2023, it remains to be seen whether this also applies to the APU.

Image 1 of 2

The Zen 4 CDNA 3 APU uses unified memory
CDNA Roadmap