AMD: RDNA 3 Speculation, Rumours and Discussion

Usual leakers at it again. Kepler is hinting at >50% perf/W increase for RDNA3 respect to RDNA2. Also there are some numbers floating for the TDP, Kepler seems confident in a 350W SKU while Greymon is declaring that Lovelace may reach up to 800W (!) on some high-end SKUs, with baseline at 450W. To me it seems too much, more so for the TDP gap between the two architectures.

 

Kepler?

 

In a reality where it really was 350 watt vs 800, I wonder how many gamers would pick the Nvidia option. What performance advantage would be needed for people to choose it? For me it would have to be well over double, maybe 3-4x.

 

I don't think there's that many people looking at power consumption. For most its all about price/performance. If you can get more or less the same performance for more or less the same price I can see people wondering if getting a card that uses twice as much power is worth that bit of extra performance but otherwise I'd say most don't really give it any thought.

 

Maybe they don't look too much at power consumption numbers, but , imho, they care about noise and heat. And not blowing up their psu

 

This applies at power levels that are low enough where there are no real consequences. Few should care when its 180 vs 250 watts or something along those lines. Although when Nvidia was more efficient yet slower and more expensive it certainly seemed to matter a ton.

 

800W is waaaaaay too much. Wtf.

If AMD manages to get similar performance using only 350 watts, there is something seriously wrong with Lovelace.

Nvidia should overhaul the complete architecture already. If the rumors is too, it's basically just Turing plus plus (given Ampere is just Turing with minor improvements) which would be really sad. Turing was revolutionary in 2018 and holds up very well today, but the architecture is nearly 4 years old. Time for something new.

I believe AMD will have a lead here. RDNA3 will overhaul the architecture completely and it is going to be the first consumer GPU using MCM.

 

"Holds up" as in beats everything else?

 

AMD has to come with a redesign since their behind in basically everything with rdna1/2.

In a world where consoles have 350w psu’s, (the highest ever for a console), we are indeed seeing increases, but 800w to 1.2kw for gpus doesnt seem realistic to happen, not even for rdna3 which likely will need higher clocks than rdna2 does.

 

RDNA is behind in RT. They need a separate unit to accelerate more of the ray process without bogging down the rest of the chip. DLSS is much more of a software issue to solve than a hardware one. I don’t expect a worthy DLSS competitor from AMD for years. In rasterization RDNA is superior.

 

Nope. RDNA2 is about on par with Ampere in rasterization but even that is dependent on resolution - in higher resolutions RDNA2 is significantly slower due to memory bandwidth limitations.

 

Rdna in its current form is behind in ’dlss’ because it lacks hardware acceleration for it, as explained by another user recently in a different topic. Not just software.

https://forum.beyond3d.com/posts/2243177/

In rasterization, rdna2 on pc (the full rdna2 with IC) is competitive, not ahead. Ampere might be ahead even there going forward seeing its compute/wide oriented design. Ampere arguably has the more advanced feature set aswell, and isnt reliant on an infinity cache-like solution as rdna2 is, which can become a bottleneck at high bw/resolution situations.

 

Realistic to happen? My guess is it feels like that just now. Look into the past.

The ATi 9700 Pro review labels its power consumption as high and highlights the need of an external floppy power cable. TDP of the chip was 37W.

The legendary 'turbo sucker' nV FX 5800 features a whooping 44W TDP. Similarly, the legendary 'Thermi' GTX 480 is rated at mild 250W.

In the CPU world, the pinnacle of Intel's NetBurst era - Prescott - running at 3.6GHz got 115W TDP.

A 2002 highend Radeon R300 user would be utterly shocked by today's 350+W TDP. 20 yeras of progress with almost 10 times the TDP.

 

I don't know RDNA3 performance except the usual leaks but I know Lovelace target performance and some of the uarch improvements so I can say this:
It looks like RDNA3 will have the lead in rasterization but I have high doubts that it will win in RT. Lovelace bring several RT improvements and a big leap in compute vs Ampere. In fact, Lovelace is ~3 times faster than Ampere in RT thus a quick napkin math tells me that RDNA3 must be 5~6 times faster than RDNA2 in RT to beat Lovelace...
Next gen will be so fast in rasterization that it won't be the main metric to take the crown. Reaching 200fps or 250fps at 4k in all AAA games on a $2k USD GPU makes no difference but if you tank below 80fps with RT ON when your competitor is above 150fps is a big deal (random number to prove my point)
So wait and see, but I don't believe RDNA3 will be faster than Lovelace in triple A games with RT ON...

 

Nothing would prevent AMD using wider and faster memory bus over IC. IC is there as alternative solution to wide and fast bus.
AMD hasn't been tied to specific memory option for any architecture since forever.

 

To be fair the 800W figure was given by the leaker for a really, really high-end model, while the baseline for this "AD102" board was given as 450W (which is already reached by custom 3090 cards). Nevertheless, the leaker seems confident on next gen Nvidia high-end cards to consume a lot of power, and more than equivalent AMD parts.

Being weaker in RT is triue for RDNA2 but we don't know if RDNA3 will have improvements in that regard.

 

Wow, nice.

Any source for that claim?

 

I think it's a very safe bet that RDNA 3 will still be substantially slower in RT than Lovelace.

AMD could add tensor cores to their GPU without much issue. The real work is on the software side. Even if RDNA 2 had the cores there would be no DLSS equivalent.

 

A non-trivial portion of the potential high power consumption likely comes from the memory system. Ampere also supports acceleration for ray traced motion blur but no game developer ever asked for this feature so who else knows what other dumb features they included in their hardware design ? Most of the products featuring their latest architecture don't support double rate calculation for FP16 precision so there's virtually no hope for future architectures to make large strides in power efficiency if hardware designers don't want to take advantage of lower precision ...

Also, this might be contested but they could save a noticeable amount of die space if they didn't include tensor cores in their hardware design since there's only one end user oriented application (DLSS) for this specialized hardware. The worst case scenario you have with feature creep is that they could decide to add more features to only find out that you spent more hardware running the same number of available applications as before ...

 

Only if you cherry-pick View instancing level. In other DirectX 12 Ultrimate features RDNA 2 is more advanced than Ampere (RDNA 2 supports 44bit virtual addressing, Ampere 40bit. RDNA 2 supports Sampler feedback 1.0, Ampere just 0.9. Ampere doesn't support Stencil reference value from Pixel Shader at all, RDNA 2 does).

Are you sure that you understand the way in which the cache works? It doesn't make the product more bandwidth bottlenecked, but less bandwidth bottlenecked. I could say that 6MB LLC of GA102 makes Ampere more bandwidth bottlenecked than 128MB LLC of Navi 31.

Anyway, when Nvidia increased LLC size in Maxwell by up to eightfold (from 256 kB of GK107 to 2048 kB of GM107), it was praised by all the reviews as a great bandwidth- and power-saving feature. Now I see all of the reviews were wrong and the big cache in fact limited Maxwell greatly!