Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

Man I have been dying to make this one for a while now.

First rumours for RDNA3 are here so new thread time!

Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[Aapje]

I'll repeat what I said earlier... I hope you guys aren't expecting sub-300 Zen 4 either.

Of course I expect that to eventually happen. But the 5600X was $300 at launch, so the successor is unlikely to start below that.

 

[beginner99]

The old AMD is over buddy. Plus TSMC price hikes, etc.

And Inflation. $1500 for the top halo part wouldn't surpirse anyway and then maybe $1100 for the "afforable" top part. Mabye a cut-down N31 for $850 I could see. But not any lower.

 

[Saylick]

I expect the following price ranges and performance:
$400 - $599: N33-based products with raster performance roughly on par with N21, RT performance to be noticeably better
$600 - $999: N32-based products with raster performance roughly 50% better than N21, RT performance to be >2x N21
>$1000: N31-based products with raster performance roughly 100% better than N21, RT performance to be ~3x N21
>$1500: N31-based halo product with raster performance roughly 125% better than N21, RT performance to be >3x N21

 

[gdansk]

Anyone who has been paying attention care to guess who will have the fastest 300±25W GPU this generation? I think that's where I'll end up

Cutdown N31, N32, AD104, AD103. Not sure what the options are.

 

[Aapje]

AMD is surely going to have the best performance/watt.

 

[Timorous]

And Inflation. $1500 for the top halo part wouldn't surpirse anyway and then maybe $1100 for the "afforable" top part. Mabye a cut-down N31 for $850 I could see. But not any lower.

I think top halo will be $2,000 if the v-cache SKU rumour is accurate Either as a 7900XT3D or maybe some prosumer style naming like the Titan series. The non v-cache SKU with the fully enabled die may be $1500 and I can see a cut version with 20GB ram coming in at $1100. I think $750-800 will be where N32 starts.

I expect the x600 and x700 tier to come in at 6x50XT pricing. There was no 6850XT but again 7800XT will likely see a bump to $750ish which would be about the same kind of increase the 6750XT was over the 6700XT.

7900XT will be a lot more expensive but performance will also see a greater uplift I expect of > 2x. If 2.5x is anywhere near true then even a 50% price hike to $1,500 or so represents a good improvement in perf/$

 

[biostud]

N33 is supposedly as fast as N21. At least in 1080p, not sure about 1440p, highly doubt in 2160p.
I compared RX 6600XT vs RX 6900XT
1080p: +60%
1440p: +81%
2160p: +117%
FullHD looks realistic.

Agree, the memory bandwidth will make sure it doesn't completely obsolete the 6800XT/6900XT when playing higher resolutions.

 

[Kepler_L2]

I expect the following price ranges and performance:
$400 - $599: N33-based products with raster performance roughly on par with N21, RT performance to be noticeably better
$600 - $999: N32-based products with raster performance roughly 50% better than N21, RT performance to be >2x N21
>$1000: N31-based products with raster performance roughly 100% better than N21, RT performance to be ~3x N21
>$1500: N31-based halo product with raster performance roughly 125% better than N21, RT performance to be >3x N21

AMD is not selling a 128-bit GPU for $600, doesn't matter how fast it is.

 

[Saylick]

AMD is not selling a 128-bit GPU for $600, doesn't matter how fast it is.

lol, but what if it were 10x as fast as a 6900XT?? /s

How do the other performance / price ranges look to you?

 

[Timorous]

lol, but what if it were 10x as fast as a 6900XT?? /s

How do the other performance / price ranges look to you?

It would run out of VRAM at 4K and be slower in any modern game...

 

[Stuka87]

It would run out of VRAM at 4K and be slower in any modern game...

Huh? Bus width and VRAM capacity are two distinctly different things.

 

[Kepler_L2]

lol, but what if it were 10x as fast as a 6900XT?? /s

How do the other performance / price ranges look to you?

N33 for $400 and lower.
N32 for $500-800
N31 for $1000 and up.

 

[GodisanAtheist]

AMD is not selling a 128-bit GPU for $600, doesn't matter how fast it is.

AMD is not selling a 128-bit 256-bit GPU for $600 $1000, doesn't matter how fast it is.

 

[Timorous]

Huh? Bus width and VRAM capacity are two distinctly different things.

Shall I be explicit.

A GPU with a 128 bit bus at current GDDR6 ram speeds and density is going to lack bandwidth and capacity to handle modern 4K games.

Of course with GDDR7 and higher density ram you might be able to make a 128bit GPU with enough bandwidth and VRAM to handle then modern 4K games but then PS6 and Xbox X will release with beefier GPUs and more ram so textures will grow again since they will probably do 4k native and 8k upscaled and VRAM requirements will grow but that does seem to be slowing down.

 

[tomatosummit]

I expect the following price ranges and performance:
$400 - $599: N33-based products with raster performance roughly on par with N21, RT performance to be noticeably better
$600 - $999: N32-based products with raster performance roughly 50% better than N21, RT performance to be >2x N21
>$1000: N31-based products with raster performance roughly 100% better than N21, RT performance to be ~3x N21
>$1500: N31-based halo product with raster performance roughly 125% better than N21, RT performance to be >3x N21

I'd say those a good chunk above what they'll sell at, the market is going to be more resistant to higher msrps this year although aibs might have more premium versions to push.
This generation of gpus are launching with actual competition on a bitcoin downturn so neither side can go crazy with gpus they expect to sell, halo parts are of course another matter. n31xt I would put at $800 and the halo n31xtx at $1300+ and if there's a need for the n31pro-5mc version then around $650 due to how late n32 is going to be, sort of like the 6800 today, only exists to fill a temporary gap. $1000 are a big ask without the mind share to go along with it.
The previous gen had reasonable rrps before stock dried up, mining and scalping kicked in so I think there's less flex in price increases than people think.

Same goes for zen4 people have mentioned in the thread. Zen3 launched WITHOUT COMPETITION. The days after alder lake launched zen3 saw unofficial price drops and then official drops to under $200 shortly afterwards.
AM5 also has the problem of being a fresh platform and against alder/raptor so a cheaper mainstream cpu is needed to get motherboards off shelves and with no whispers about desktop rembrandt yet there might/should be a ~$200 raph a few months after launch.

 

[Stuka87]

Shall I be explicit.

A GPU with a 128 bit bus at current GDDR6 ram speeds and density is going to lack bandwidth and capacity to handle modern 4K games.

Of course with GDDR7 and higher density ram you might be able to make a 128bit GPU with enough bandwidth and VRAM to handle then modern 4K games but then PS6 and Xbox X will release with beefier GPUs and more ram so textures will grow again since they will probably do 4k native and 8k upscaled and VRAM requirements will grow but that does seem to be slowing down.

16GB of RAM is plenty for 4K during this point in time. With a 128bit bus, and 2GB GDDR6 chips, thats 16bits per chip.

An RTX 3090 has a 384bit bus, but it has 24 ram chips, which means each chip has... 16bits per chip.

Now yes, the 3090 RAM runs at a higher clock, as it uses a GDDR6X. But a card having a 128bit bus, does not mean it will lack capacity. And the whole point of AMD having the giant cache is to make memory bandwidth less important.

 

[Timorous]

16GB of RAM is plenty for 4K during this point in time. With a 128bit bus, and 2GB GDDR6 chips, thats 16bits per chip.

An RTX 3090 has a 384bit bus, but it has 24 ram chips, which means each chip has... 16bits per chip.

Now yes, the 3090 RAM runs at a higher clock, as it uses a GDDR6X. But a card having a 128bit bus, does not mean it will lack capacity. And the whole point of AMD having the giant cache is to make memory bandwidth less important.

Taking Samsung they have upto 24gbps parts at 16Gb density sampling and in mass production they have 16Gb density at 14gbps. On a 128 bit bus it gets you the capacity and your bandwidth would match that of the 6700XT so it might work for 1440p / entry level 4k. Of course 24gbps chips are only sampling so they might be rather expensive and low volume making them unsuitable for a high volume product like N33. The 14Gbps parts lack the bandwidth required on a 128bit bus.

So my point stands. With current and sampling GDDR6 there is not the required combination of speed and density to make a 128bit card suitable for 4K. Once GDDR7 is available that probably changes and you could see a 128 bit card with 16GB of ram coming in at $600 to fill the 1440p tier but we are not there yet.

 

[Mopetar]

Again, with the yields as good as they are there shouldn't be any need to have more than one cut for each level of the stack.

Defects aren't the only reason to cut chips. Even a fully enabled chip may not be able to hit the same clock speeds with the same voltage as other parts.

Companies create separate bins for different performance categories and to target different market segments.

Expect to see both companies doing what they always have. 7nm yields were already great well before RDNA 2 parts taped out, yet AMD did as the GPU makers always have. If there are some parts of the hardware that hold back the whole chip, they get disabled. Same with CPUs.

 

[jpiniero]

Defects aren't the only reason to cut chips. Even a fully enabled chip may not be able to hit the same clock speeds with the same voltage as other parts.

Companies create separate bins for different performance categories and to target different market segments.

Have to think both companies definitely bin for quality. But they don't make official different SKUs to differentiate. The amount of chips that are dropped to a cut model because of quality is probably not much if any.

nVidia did this specifically with Turing, at least with TU102. The crap bin version were not allowed to be overclocked.

 

[Frenetic Pony]

Have to think both companies definitely bin for quality. But they don't make official different SKUs to differentiate. The amount of chips that are dropped to a cut model because of quality is probably not much if any.

nVidia did this specifically with Turing, at least with TU102. The crap bin version were not allowed to be overclocked.

I'd forgotten this type of bin exists.

It'll be interesting to see how AMD handles binning gpu chiplets (assuming that's correct). They've got the opportunity to go a lot of routes, though I'm guessing they're going for a very narrow range (maybe as low as 1 initially) for binned products to ensure good supply. Even with more bins than just defects there's still a high amount of good chiplets, witness the high prices of the cut bin 6800xt, with a lot of models still sold out, versus the ever dropping prices of the full 6900xt.

Still... maybe there'll be some "premium" overclocked/overclockable version of an SKU and a non-overclockable one as well. Or maybe they'll save the best dies up to make another AMD only liquid cooled version next year. 600 watts, overclocked memory and core clocks, $2k?

 

[Aapje]

They've been cutting down perfectly good chips for decades now, for commercial reasons. In the past the only difference was the bios and you could often reflash the bios to upgrade the card. Nowadays they make that impossible.

The reason for the high prices and low supply of the 6800 was that AMD was unable to meet demand anyway, so then there was little reason to sell 6800's when you can sell the same chip as a 6900. Now it would be a good idea to change that, but it may be too late, with the cards that come on the market already being in the pipeline and new production being ended or at the very least greatly reduced.

However, the market is normalizing, so they will surely need the 7800 to bridge the gap between the 7700 and 7900, to get that sweet, sweet upselling action going.

 

[Mopetar]

Have to think both companies definitely bin for quality. But they don't make official different SKUs to differentiate. The amount of chips that are dropped to a cut model because of quality is probably not much if any.

Look at sales for any generation. The model numbers change , but the distribution of sales is similar. If the performance capabilities are normally distributed they need to cut a little under the sigmas and it's a nice bottom 25, mid 50, top 25 which is a consistent spread as far as the market is concerned. The only difference are the final specs.

They can adjust and cut however they need to best address the market.

 

[Frenetic Pony]

They've been cutting down perfectly good chips for decades now, for commercial reasons. In the past the only difference was the bios and you could often reflash the bios to upgrade the card. Nowadays they make that impossible.

I don't see that happening if its multiple chiplets, simply because they don't need to. They could easily have 2 or 3 SKUs with all good, even very good, top tier chiplets that are fully enabled and then just shuffle them around based on demand.

This is what I mean by saying chiplets will fundamentally change how binning works. All the old strategies don't apply because you can do a lot of new ones that are more efficient.

 

[maddie]

I don't see that happening if its multiple chiplets, simply because they don't need to. They could easily have 2 or 3 SKUs with all good, even very good, top tier chiplets that are fully enabled and then just shuffle them around based on demand.

This is what I mean by saying chiplets will fundamentally change how binning works. All the old strategies don't apply because you can do a lot of new ones that are more efficient.

Unfortunately, It doesn't seem as if there are multiple compute chiplets on a GPU. Maybe next gen.

 

[Leeea]

Unfortunately, It doesn't seem as if there are multiple compute chiplets on a GPU. Maybe next gen.

But it looks like their are multiple chiplets.

It appears they are breaking the IO off the main compute chiplet. Memory and PCI lanes use a lot of space for all the traces, power, and noise control. If AMD also breaks the cache off into its own chiplet, that would allow even more space for compute.

This does not seem far fetched, AMD already breaks the IO off their CPUs. The 5900x has one IO chiplet and 2xcompute chiplets.


if you look at a 5900x:

The IO chiplet is already bigger then the compute chiplets!