Wednesday, November 30th 2022
AMD RDNA3 Second-largest Navi 32 and Third-largest Navi 33 Shader Counts Leaked
The unified shader (stream processor) counts of AMD's upcoming second- and third-largest GPUs based on the RDNA3 graphics architecture, have been leaked in some ROCm code, discovered by Kepler_L2 on Twitter. The "performance.hpp" file references "Navi 32" with a compute unit count of 60, and the "Navi 33" with 32 compute units. We know from the "Navi 31" specifications that an RDNA3 compute unit still amounts to 64 stream processors (although with significant IPC uplifts over the RDNA2 stream processor due to dual-instruction issue-rate).
60 compute units would give the "Navi 32" silicon a stream processor count of 3,840, a 50% numerical increase over the 2,560 of its predecessor, the "Navi 22," powering graphics cards such as the Radeon RX 6750 XT. Meanwhile, the 32 CU count of the "Navi 33" amounts to 2,048 stream processors, which is numerically unchanged from that of the "Navi 23" powering the RX 6650 XT. The new RDNA3 compute unit has significant changes over RDNA2, besides the dual-issue stream processors—it gets second-generation Ray Accelerators, and two AI accelerators for matrix-multiplication.
No other specs of the "Navi 32" and "Navi 33" are known at this point. If AMD is sticking with the chiplet design for the "Navi 32," it could feature a similar design to the "Navi 31," with a 5 nm GCD that has these 60 RDNA3 compute units; and either three or four 6 nm MCDs, depending on whether AMD decides to give it a 192-bit or 256-bit wide GDDR6 memory interface. It remains to be seen if the chiplet design carries on to even the smaller GPUs such as the "Navi 33." There's also no word on when AMD launches the rest of its Radeon RX 7000-series graphics card lineup. The RX 7900 series sees a December 2022 debut.
Sources:
Kepler_L2 (Twitter), HotHardware
60 compute units would give the "Navi 32" silicon a stream processor count of 3,840, a 50% numerical increase over the 2,560 of its predecessor, the "Navi 22," powering graphics cards such as the Radeon RX 6750 XT. Meanwhile, the 32 CU count of the "Navi 33" amounts to 2,048 stream processors, which is numerically unchanged from that of the "Navi 23" powering the RX 6650 XT. The new RDNA3 compute unit has significant changes over RDNA2, besides the dual-issue stream processors—it gets second-generation Ray Accelerators, and two AI accelerators for matrix-multiplication.
26 Comments on AMD RDNA3 Second-largest Navi 32 and Third-largest Navi 33 Shader Counts Leaked
The 7900 XTX is supposedly 60% faster than the 6950 XT while having only 20% more stream processors (6144 vs 5120). This makes it 33% faster per stream processor. If we carry over this same logic when comparing Navi 33 to 23 (which is not a given, considering 33 might be 6nm monolithic), then it would land somewhere between the 6750 XT and 6800 in performance, or roughly on par with the 3070.
Cut N32 will go in the 7700XT with 3 MCDs.
Not sure if the cut will be 2WGPS per SE for 48CUs, 3SEs and 96 ROPs or if AMD will cut an entire SE out for 40CUs and 64 ROPs.
It will depend on 3 things IMO. 1) How much do AMD need to push full N32 to hit the desired gen on gen gain over the 6800XT. 2) do they plan on making a 7800 available and 3) what is the lowest TBP required to hit the 7700XT performance target.
Given we are talking about a 200mm^2 die I expect yields to be excellent which means a 7800 part may not be needed unless AMD are really pushing the 7800XT bin and they might.
To make a 7800XT really worthwhile (except the Extra RT perf), i think they should clock it to at least 2.75 GHz to get about 20-25% more performance. If it's more well, the better it.
But also, the important factor will be the cost. If they offer a 2.50 GHz N32 at 600$, it could maybe make sense.
Q3 gaming revenue of NV and AMD was on par, although AMD's includes console business.
Around 5 million consoles are sold each quarter, how much that would be off 1.6 billion revenue? Even if AMD gets 200 for each chip, it leaves us with 600 million vs 1.6 billion.
Even purely from revenue perspective that is 28% vs 72%.
On top of it, NV's cards are way pricier.Surely, hordes of Steam users who bought 3050Ti over cheaper and faster 6600 are playing with "RT on"... :D
AMD took a reasonable approach, as few games support RT, and no GPU is capable of it properly even nVidia themselves say so, hence DLSS.
Why nVidia is leader is more due to their huge budget, they can spend huge amounts in marketing (you are a proof of their marketing effectiveness, so are lots of so called reviewers), they have reasonably good drivers for some time now, while AMD was nearly bankrupt (thanks to Intel's bribes).
People are swallowing the RayTracing marketing, with the same ease that they are swallowing the total numbers of cores marketing with Intel's hybrid CPUs. The end result is people buying RTX 3050 to see RayTracing effects and 13th gen CPUs, because they offer more cores than 7000 CPU models at a lower price. P cores, E cores, who cares. For the average Joe they are just cores. AMD is losing for the first time in the last 5-6 years in both CPU and GPU retail market, because the competition is offering competitive products backed up by superior marketing. If they where increasing their RT performance significantly with RX 7000 series, I mean like 2.5-3 times, they would be negating Nvidia's strongest marketing card. They didn't and until they come up with FSR 3.0, Nvidia will be happily selling framerates to the masses with DLSS 3.0, even in raster. Even if FSR 3.0 does come soon, AMD's options while having one strong advantage, they also have a big disadvantage. The advantage of course is that they are open, more easily to be adopted by developers. The disadvantage is that they are somewhat inferior solutions, because they do not use custom hardware and they are reactions to Nvidia's announcements, meaning Nvidia programmers have much more time in advance to optimize their features. This is enough for the tech press to keep repeating in every review that "DLSS is superior", making AMD's options look in every way inferior. This is something that Nvidia tried to achieve first with PhysX, but failed. Today they are succeeding.
Also, Nvidia probably wanted to cash in on crypto miners with 4000-series products, or at least expected them (miners) to clear store shelves of 3000-series ones before 4000 launched. Unfortunately for them, crypto crashed way within the development cycle of the 4000-series, stores got stuck with remaining 3000-series, which left them sitting between a rock and a hard place.
RDNA 2 wasn't very popular with miners, so AMD didn't have this problem. Their problems consist of what you said, and the RTX bandwagon.