AMD’s Next-Gen RDNA 3 Flagship GPU to Have Four Times the FP32 Compute Power of Radeon RX 6900XT
The GPU field is heating up as both the big players, NVIDIA and AMD are set to unleash their next-gen graphics cards towards the end of this year. Intel, a company largely limited to developing only processors, is also stepping into the world of discrete graphics with Arc A-Series this time. With three companies ready to battle it out for the people’s crown, competition is at an all-time high, and with such high stakes, leaks are a given.
Rumors and reports from sources around the world have kept us fed good when news from official channels is scarce. And today is no different. Both NVIDIA and AMD’s top-end flagship SKUs had massive leaks today going over their theoretical compute performance. Unsurprisingly, the numbers were nearly identical, once again giving us a trailer of just how heated the market is about to become.
In this article, we’ll talk about RDNA 3. AMD’s next-generation GPU architecture is said to bring significant performance improvements over the current-gen RDNA 2. With RDNA 3, AMD is shifting from a monolithic to a MCM (Multi-Chip-Module) design in the flagship SKUs, a first for the graphics card industry. That means AMD will be using 3D stacking technology to stack chips on top of each other to maximize performance.
Navi 31 TFLOPs
Case in point, popular leaker Greymon55 took to Twitter today to point out the kind of performance we can expect from Navi 31, the top-end RDNA 3 GPU and the direct successor to Navi 21. We’ll most likely see this GPU used in the Radeon RX 7900XT at launch and it will be part of the very first few SKUs that AMD releases. The RX 7900XT will replace the current-gen RX 6900XT as the new flagship series.
As you can see in the tweet attached above, Greymon mentions “92T FP32” referring to the FP32 performance of the Navi 31 GPU and 92T, of course, means 92 Teraflops (TFLOPs). According to the leaker, Navi 31 will hit 92 TFLOPs of FP32 performance, a number that is higher than the already-previously-leaked value of 75 TFLOPs. For context, the RX 6900XT flagship GPU only has 23 TFLOPs of FP32 performance, so we’re looking at a 4x improvement here.
One important thing to keep in mind here is that 92 TFLOPs is the theoretical maximum, meaning that number is in accordance to the factory clock speeds. If you were to overclock the GPU, you could potentially get even closer to that 100 TFLOPs barrier. Speaking of which, Greymon55 says that RX 7900XT (Navi 31) will feature a 3Ghz boost clock which is already insane in of itself, so just imagine the overclocking possibilities.
While all this sounds nice, AMD’s evil twin is treading on an oddly similar line. NVIDIA’s next-gen Ada Lovelace is rumored to actually break the 100 TFLOPs barrier and even offer clock speeds similar to AMD’s RDNA 3 offerings. Most of that can attributed to NVIDIA’s switch from Samsung‘s inferior chips to TSMC‘s bleeding-edge 4nm process node, the same node AMD RDNA 3 will be manufactured on.
As for the specs, the current rumors are largely unchanged from what we’ve been hearing for the past few months. Navi 31 is said to pack 15,360 Stream Processors and 60 WorkGroup Processors. There are conflicting reports pertaining to how how AMD is dropping its Compute Units (CUs) in favor of WorkGroup Processors on RDNA 3.
Kepler has revealed that AMD might end up keeping Compute Units around after all but rework them to now be a part of the WorkGroup Processor. In the tweet below, the leaker explains how AMD has doubled the amount of SIMD32 clusters on RDNA 3 as compared to RDNA 2. With AMD’s next-gen RDNA 3 GPUs, we’ll see four SIMD32 clusters per each Compute Unit, instead of two per CU. All in all, we’re looking at the same amount of CUs per WGC in RDNA 3, but with twice the SIMD clusters.
According to Linux drivers, it appears that Compute Units still exist, and WGPs are still DualCUs. The big change is going back to 4x SIMD per CU like GCN, but now with RDNA's 32 ALU design. pic.twitter.com/hAzGs9M9E6
— Kepler (@Kepler_L2) April 30, 2022
AMD is hard at work at making sure RDNA 3 is their most competitive product ever. Where RDNA 2 surprised everyone by offering genuine alternatives to NVIDIA’s RTX 30-series, RDNA 3 will be AMD’s attempt at finally stealing the crown the Red Team lost to NVIDIA decades ago. Whether their efforts materialize into success or not remains to be seen.
And we won’t have to wait for long as the company is rumored to announced Radeon RX 7000 series at Computex 2022 in May. Paper availability is expected in the third quarter of this year. Let’s just hope that this gargantuan uplift in performance doesn’t come at the cost of, well, cost. Higher prices at a time like this could only worsen the severe GPU drought that we’ve just started to recover from. Though, luxury pricing could help alleviate high demand which can keep prices in check as a whole, so I guess there’s a silver lining after all.