Initial RDNA 4 details reveal Navi 41 to be 2x faster than Navi 31 as RDNA 4 could have 50-60% better performance/watt vs RDNA 3

Back in June, AMD shared that it would be using an “Advanced Node” for its RDNA 4 architecture. The company also revealed that the Navi 4x GPUs would come in 2024. This was all the information that we received from AMD at the time. RedGamingTech has now detailed some early rumors regarding RDNA 4.

According to Paul, RDNA 4 is a brand-new architecture with far-reaching changes. One of the biggest is the presence of a single Graphics Compute Die (GCD) that is divided into one or more GCXs. Similar to RDNA 3, RDNA 4 could have 128/64 Arithmetic Logic Units (ALU) per Compute Unit (CU). There will allegedly be 48 CUs and 3 Shader Engines (SE) per GCX. However, the number can change depending on a host of factors.

Additionally, Memory Cache Die (MCD) chips in RDNA 4 will reportedly retain their memory and cache controller role with GDDR7 being used in place of GDDR6. When it comes to cache, Paul mentions a big increase in cache sizes, as well as third-gen Infinity Cache (IC) but the leaker doesn’t provide any figures.

There could also be significant under-the-hood improvements in the way of new data prefetching, better software dependency manager, Wave Matrix Multiply Accumulate (WMMA) version two that can speed up performance per CU for certain instructions by 2x, etc. Finally, because of the architectural improvements, AMD may also increase the number of professional and server SKUs.

Finally, Paul reports that RDNA 4 will use PCIe 5 and the GCDs will be manufactured using TSMC’s 3 and 4 nm processes.

RDNA 4 performance estimate

The leaker alleges that RDNA 4 could have a 50-60% performance/watt advantage over RDNA 3. But he is “less confident” in this number, so take it with a grain of salt. The clock frequencies will reportedly see an enormous uplift over the current RDNA 3 boards as RDNA 4 could effortlessly hit 3.5+ GHz provided AMD meets its target which is something the company has supposedly failed to do with RX 7900 XT and RX 7900 XTX.

Moreover, RDNA 4’s performance targets also appear to be quite ambitious with an expected increase of more than 2x vs RDNA 3. AI/ML performance could also see a big bump. Paul claims that Navi 41, a GPU that could appear in a possible RX 8900 XT/XTX if AMD follows the current naming scheme, will have double the performance of Navi 31 featured in the RX 7900 XT/XTX.

Navi 41, Navi 42, and Navi 43 specs

Navi 41, the flagship RDNA 4 GPU, will reportedly have 48 CUs per GCX and a total of three GCXs per GCD. This results in 144 CUs for a fully-enabled Navi 41 GPU. As a comparison, the Navi 31 chip featured in the RX 7900 XTX has 96 CUs. The total ALU count of Navi 41 could land at 18,432.

Moving on to VRAM, the leaker mentions two possible configurations, a 4 MCD configuration with 32 GB of GDDR7 and a 6 MCD setup with 24/48 GB of GDDR7. Paul thinks that the second configuration is more likely.

Navi 42, which will be a tier below Navi 41, apparently has two GCXs resulting in 96 CUs and 12,288 ALUs in total and 4 MCDs. The CU and ALU figures are similar to Navi 31. Even then, Navi 42 could enjoy a 2x performance increase over the unreleased Navi 32.

Navi 43, the weakest of the RDNA 4 lineup, will allegedly have a single GCX for a total of 48 CUs. There will be 2 MCDs with GDDR7 memory. The performance target of Navi 43 is similar to its higher-end brethren with a 2 times performance increase over Navi 33.

Only time will tell how much of the information provided by RedGamingTech ends up being true. We are at the beginning of the RDNA 3’s lifecycle with a slew of AMD Radeon GPUs to look forward to. As such, it is too early to speculate on the final specs and performance targets of RDNA 4. Therefore, take the information presented here as early rumors and nothing concrete.

Buy the ASUS Dual AMD Radeon RX 6700 XT on Amazon