Monstrous 129 FP32 TFLOPS GPU, GDDR7, 3.5GHz clock rates, and more

RedGamingTech has posted a lot of information about the AMD RDNA4 architecture and upcoming Navi 41, Navi 42, and Navi 43 GPUs. This is one of the first leaks from a well-known leaker, and while RGT has a reliable track record, this post is still considered a rumor at this point because This early information is always in a state of flux and is subject to change. So with salt shakers in hand, let’s dive into this ultimate RGT leak remedy.

AMD Navi 41, Navi 42, and Navi 43 GPUs Get Popular Preliminary Specs for Upcoming Radeon RX 8000 Series

On the first work order, RGT confirms that the number of ALUs for RDNA4 CUs will remain exactly the same as RDNA 3 – that is, it will remain at 128 ALUs per CU. GDDR7 memory is definitely in the works but not confirmed at the moment although it is very likely that PCIE 5.0 will be the preferred standard against PCIE 6.0 as the latter will likely not be ready in time. Interestingly, the RDNA4 architecture will target a boost clock frequency of 3.5GHz, and that combined with other architecture moves means you’re looking for a net performance-per-watt gain of about 50-60%.

One thing that AMD’s RDNA3 architecture was lacking was ray tracing and AI/ML performance and AMD is looking to fix that with RDNA4. The caches will get big upgrades and 3Research and development Generation Infinity Cache will be launched alongside the new data prefetching system. WMMA (Wave Matrix Multiply-Accumulate) V2 will also enable multiply performance per controller (most likely) for matrix instructions allowing use of the entire SIMD lane. Although the overall increase in CUs is about 50%, the generation-to-generation performance increase is 100% – which is nice to hear. Here are the separate templates:

Also Read:  Samsung's changing development plans to focus on OLED panels for the 2024 iPad Pro models

AMD Navi 41 – 144 CUs

Here’s where things get even crazier, Navi 41 is touted to have up to 32GB of vRAM – possibly of the GDDR7 variety. There is a second configuration as well with 24 or 48 GB of GDDR7 and 6 MCDs (the 32 GB variant has 4 MCDs). Navi 41 will have 144 controllers, which are clocked at 3.5GHz and when you factor in 128 ALUs per controller, should result in 129 TFLOPs of FP32 performance. Needless to say, that’s the level of Godzilla performance there is.

AMD Navi 42-96 CUs

The Navi 42 will be slightly muted and has a net total of 12,288 CU clocked at 3.5GHz, which should result in about 86 TFLOPs of FP32 performance. 96 CU is also exactly the number of RDNA3’s flagship, the Radeon 7900 XTX, so this SKU would represent apples to apples generation compared to the generation between RDNA3 and RDNA4 (or more accurately, apples to apples it would ever get). As RGT notes in its slides, specs like this *could* be changed prior to final revision.

AMD Navi 43-48 CUs

Finally, we have Navi 43, which will have 48 currency units. At about 3.5GHz, this will yield about 43 TFLOPs of FP32 performance. This is probably the lower end of the AMD RDNA4 lineup and just continues to show how much performance is being packed into the upcoming RDNA4 architecture as the “low end” easily exceeds 40 TFLOPs. The memory standard will still be GDDR7 and will come with 2x MCDs. This is also the only SKU where RGT notes that the GPU may not be of MCM design – despite being told that this is currently the case.

Which AMD Navi GPU are you most interested in?

Also Read:  Wrapped Versions of Bitcoin and Ethereum (wBTC and wETH) Safe, Binance Still Least CEX Risk for Now

Share this story



Leave a Comment