AMD: RDNA 3 Speculation, Rumours and Discussion

[neckthrough]

So basically bad price/perf for RDNA3 too?

It's only "bad" if your benchmark is 2016 levels of generational improvement. Those baselines are irrelevant given today's silicon costs. Understand that Moore's Law is completely dead in the original $/transistor sense. You can get more performance, but it will cost more. Bleeding-edge silicon is *precious* and dGPUs are essentially large, dense concentrations of the material.

Companies can play temporary margin-sacrificing (or even "contra-revenue") games and hemorrhage themselves by selling at artificially low prices but at least for AMD it seems they just don't have to.

 

[Leoneazzurro5]

It's only "bad" if your benchmark is 2016 levels of generational improvement. Those baselines are irrelevant given today's silicon costs. Understand that Moore's Law is completely dead in the original $/transistor sense. You can get more performance, but it will cost more. Bleeding-edge silicon is *precious* and dGPUs are essentially large, dense concentrations of the material.

Companies can play temporary margin-sacrificing (or even "contra-revenue") games and hemorrhage themselves by selling at artificially low prices but at least for AMD it seems they just don't have to.

I meant in comparison to the competition. I know that silicon costs are increasing.
If NV has better raster and better RT performance ad a due price point, and AMD wants to sell cards at the same price point, good luck with it.
Also, the 4080 12GB seems to have a worse perf/price ratio than even the 4080 16GB and 4090. This is also bad pricing (or artificial driving the customers towards higher margin cards)

 

[DegustatoR]

If NV has better raster and better RT performance ad a due price point, and AMD wants to sell cards at the same price point, good luck with it.

I'd wager that it will be pretty much a repeat of RDNA2 vs Ampere: AMD will have a feature/performance deficit which they'll try to solve by setting slightly lower MSRPs.
I don't think that they will have any advantage in raster as I don't really this advantage between RDNA2 and Ampere already, and RDNA3 scaling seem more compute focused than anything.

 

[Leoneazzurro5]

I'd wager that it will be pretty much a repeat of RDNA2 vs Ampere: AMD will have a feature/performance deficit which they'll try to solve by setting slightly lower MSRPs.
I don't think that they will have any advantage in raster as I don't really this advantage between RDNA2 and Ampere already, and RDNA3 scaling seem more compute focused than anything.

Nah, these are not compute cards. The focus was very clear: maximum perf/area.

 

[DegustatoR]

Nah, these are not compute cards. The focus was very clear: maximum perf/area.

All GPUs are "compute cards", and by increasing compute is how you get maximum perf/area - in the absence of h/w like RT cores and XMX/TCs at least.
RDNA3 seemingly doubles compute throughput per WGP compared to RDNA2. The rest won't scale as well so you may predict where the biggest improvements will occur.
But as always -- we'll see.

 

[Bondrewd]

So basically bad price/perf for RDNA3 too?

Of course.
It's gonna be as expensive as the market tolerates.

 

[iroboto]

Of course.
It's gonna be as expensive as the market tolerates.

Damn. But what if the market can only tolerate sub 700 and team green is just way too out there for the mainstream crowd. I can’t see AMD fighting over a small piece of the market when the major bulk of the market is underserved.

 

[yuri]

APUs cover the IP development costs.

This is true. However, it makes one wonder why AMD keeps spending $$$ on developing and producing largish dGPUs on very expensive litho. Their market penetration has been tiny even with the relatively-well-competing RDNA2. Despite it's apparent flaws the Steam Survey shows a depressing picture for AMD dGPUs.

Is it even viable to design and produce such things as 6800 XT when a vanilla 3080 outsells it more than 10.5x? GFX IP for APUs, consoles, and compute accelerators could be developed w/o spending on the big die dGPUs.

 

[Bondrewd]

I can’t see AMD fighting over a small piece of the market when the major bulk of the market is underserved.

It's all N6/N5 wafers that could be spent railing Intel DCAI group into nonexistium.
Or making Intel client guys cry everyone a river.

Remember, AMD is first and foremost a server CPU vendor with everything else being hardly interesting to them at this point.

However, it makes one wonder why AMD keeps spending $$$ on developing and producing largish dGPUs on very expensive litho

The upper management at AMD likes to keep market presence even if they don't sell much stuff.
Those WS Threadrippers are meme-y but they still make them with a whole different platform to boot.

Is it even viable to design and produce such things as 6800 XT when a vanilla 3080 outsells it more than 10.5x?

Yea why not.
GPUs aren't particularly egregious to tape and validate.

 

[yuri]

The upper management at AMD likes to keep market presence even if they don't sell much stuff.
Those WS Threadrippers are meme-y but they still make them with a whole different platform to boot.

Yea, the WS TRs are quite close but... Catering a separate niche platform hosting binned server CPUs must not that expensive/hard compared to developing a top-to-bottom new GFX card.

These processors have been the perfect PR weapon - producing the fastest processor ever sounds great.

 

[Jawed]

When does the first APU with RDNA 3 launch?

Will it arrive before the final desktop discrete RDNA 3 card is launched? (Excluding refreshed desktop discrete.)

 

[Bondrewd]

Catering a separate niche platform hosting binned server CPUs must not that expensive/hard compared to developing a top-to-bottom new GFX card.

It's still a lot of platform work and validation and then a ton of OEM bootlicking to get those used anywhere.
WS markets do be dogshit slow.

These processors have been the perfect PR weapon

Which is why bigger AMD dGPUs still exist.
It keeps them afloat in those upper segments.

 

[Bondrewd]

When does the first APU with RDNA 3 launch?

Phoenix, CES.

Will it arrive before the final desktop discrete RDNA 3 card is launched?

Yea.

 

[Jawed]

So, when's the earliest that a Steam Deck 2 could launch, with RDNA 3, and not piss off everyone who's bought Steam Deck?

 

[Bondrewd]

So, when's the earliest that a Steam Deck 2 could launch, with RDNA 3, and not piss off everyone who's bought Steam Deck?

Late'23 or somewhere there.

 

[eastmen]

So, when's the earliest that a Steam Deck 2 could launch, with RDNA 3, and not piss off everyone who's bought Steam Deck?

My guess is going to be late 2024 or early 2025 and I bet they go rdna 4

 

[Frenetic Pony]

When does the first APU with RDNA 3 launch?

Will it arrive before the final desktop discrete RDNA 3 card is launched? (Excluding refreshed desktop discrete.)

Probably early next year for laptops, then I'd expect desktop to arrive after it, same execution pattern as usual.

I'd expect the desktop ones to by hyper popular. $300(ish) for a 6 core CPU and a 6(ish) teraflop GPU in one package? That's probably Spiderman Remastered @1080p with high settings (including some RT setting) and 30fps. Sign a hell of a lot of people up.

I do wonder if they'll do 2 configs for APU GPUs. If not immediately maybe later, a kind of M(x) Pro/Ultra competitor. Base APU should be like, 16CUs (half a 6600xt, or maybe 24?) but could they do one that's double? I don't see why not.

I wonder if Steam Deck 2 could launch next year towards the end? I don't see it as impossible, there's no reason to wait for RDNA4. 3 is a huge upgrade, as is Zen 4 over 2, and a better form factor/upgraded screen are needed due to competition reasons. A "Steamdeck light" that's the same perf as 1, but with an OLED, small form factor, and bigger battery life. Then a new mainline one that's 50% faster and a better screen, let people choose which one they want.

 

[Bondrewd]

I'd expect the desktop ones to by hyper popular

No DT PHX planned so far.

 

[Henry swagger]

I have a feeling the 7900xt will take the raster crown.. infinity cache is amd's trick. they must improve ray tracing to put more pressure on nvidia

 

[Flappy Pannus]

AMD has officially lowered prices on RDNA2.* Not 'huge' drops (well aside from the Covid Cards that made no sense otherwise - 6900XT from $999 to $699!), but drops nonetheless, so not bad - albeit you've been paying less on Newegg/Amazon already.

I don't expect RDNA3 to immediately supplant the entire lineup, maybe not even in the first year - but a drop is a drop and bodes a bit of hope that we will see cards that actually replace some in their lineup rather than being just enthusiast-enthusiast tier.

EDIT thanks to @Kaotik: Uh, no, doesn't look like MSRP after all. AMD's chart was just quoting average Newegg price as of Sept.