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

 

[TESKATLIPOKA]

CUs still exist and they are smaller than RDNA2 CUs @ 7nm, though N33 CUs @ 6nm and N31/32 CUs @ 5nm are slightly different.

Any other interesting Info?

 

[Kaluan]

PULSE AMD Radeon™ RX 6500 XT 8GB

<p>Up to 2855 MHz, 8GB/64 bit DDR6. 18 Gbps Effective</p>

www.sapphiretech.com

Sapphire begs to differ.

For anyone interested, I finally found someone testing (Sapphire's) 6500 XT 8GB performance, it's hidden in this Arc A380 review:

And a few other more uncommon/older/light games tested, with those the difference is predictably smaller.
Testbed is Ryzen 5 3600X, B550 and 2x8GB DDR4-3200.

Now, looping back to N33 and discussions of 8x lane being a weakness, and similarly that 4x was the 6500 XT's biggest weakness, where I said I don't think that's true, the 4GB VRAM, in tandem with 4x lanes and too little effective bandwidth (144GB + 16MB IC) made a deadly performance-killing/stutter fest combo. Glad to see a proper frame buffer size fixed most of that and then some.
With the supposed end-to-end compression RDNA3 has, penalties for going more often out of VRAM and back would be lessen on a 4x, 4GB 6500XT, if it was hypoethically RDNA3 uArch.

There's no 3.0 vs 4.0 tests (this isn't a 6500 XT 8GB review after all), but I think we'd see an even wider gap between 4GB and 8GB.

Back when Sapphire's 6500 XT 8GB was announced, there were many saying so much VRAM is a waste and N24 can't make use of it lol

That being said, I am curious what will the eventual RX 7500/7400 family be based off... Cut down N33? Later announced N34 (similar to how N24 was only discovered much later)? N23 N6 refresh? And will AMD not mess up it's entry lineup again like that?

 

[Timorous]

For anyone interested, I finally found someone testing (Sapphire's) 6500 XT 8GB performance, it's hidden in this Arc A380 review:

View attachment 69633

And a few other more uncommon/older/light games tested, with those the difference is predictably smaller.
Testbed is Ryzen 5 3600X, B550 and 2x8GB DDR4-3200.

Now, looping back to N33 and discussions of 8x lane being a weakness, and similarly that 4x was the 6500 XT's biggest weakness, where I said I don't think that's true, the 4GB VRAM, in tandem with 4x lanes and too little effective bandwidth (144GB + 16MB IC) made a deadly performance-killing/stutter fest combo. Glad to see a proper frame buffer size fixed most of that and then some.
With the supposed end-to-end compression RDNA3 has, penalties for going more often out of VRAM and back would be lessen on a 4x, 4GB 6500XT, if it was hypoethically RDNA3 uArch.

There's no 3.0 vs 4.0 tests (this isn't a 6500 XT 8GB review after all), but I think we'd see an even wider gap between 4GB and 8GB.

Back when Sapphire's 6500 XT 8GB was announced, there were many saying so much VRAM is a waste and N24 can't make use of it lol

That being said, I am curious what will the eventual RX 7500/7400 family be based off... Cut down N33? Later announced N34 (similar to how N24 was only discovered much later)? N23 N6 refresh? And will AMD not mess up it's entry lineup again like that?

I think 7500 will be cut N33 with 6GB.

Below that not sure, APU perhaps or maybe there will be an N34 at some point.

 

[GodisanAtheist]

For anyone interested, I finally found someone testing (Sapphire's) 6500 XT 8GB performance, it's hidden in this Arc A380 review:

View attachment 69633

And a few other more uncommon/older/light games tested, with those the difference is predictably smaller.
Testbed is Ryzen 5 3600X, B550 and 2x8GB DDR4-3200.

Now, looping back to N33 and discussions of 8x lane being a weakness, and similarly that 4x was the 6500 XT's biggest weakness, where I said I don't think that's true, the 4GB VRAM, in tandem with 4x lanes and too little effective bandwidth (144GB + 16MB IC) made a deadly performance-killing/stutter fest combo. Glad to see a proper frame buffer size fixed most of that and then some.
With the supposed end-to-end compression RDNA3 has, penalties for going more often out of VRAM and back would be lessen on a 4x, 4GB 6500XT, if it was hypoethically RDNA3 uArch.

There's no 3.0 vs 4.0 tests (this isn't a 6500 XT 8GB review after all), but I think we'd see an even wider gap between 4GB and 8GB.

Back when Sapphire's 6500 XT 8GB was announced, there were many saying so much VRAM is a waste and N24 can't make use of it lol

That being said, I am curious what will the eventual RX 7500/7400 family be based off... Cut down N33? Later announced N34 (similar to how N24 was only discovered much later)? N23 N6 refresh? And will AMD not mess up it's entry lineup again like that?

-Wow, that's like up to 30% additional performance with the addition of some RAM, that's basically an additional tier of card between the 6600 and 6500xt.

If that performance kept up across a broad range of titles it's frankly a little dumb that AMD didn't just throw 2gb chips on the 6500xt and just release it as an official card.

 

[jpiniero]

If that performance kept up across a broad range of titles it's frankly a little dumb that AMD didn't just throw 2gb chips on the 6500xt and just release it as an official card.

The original version does use 2 GB chips. The 8 GB version has 4 chips on a 64-bit bus.

 

[Leeea]

For anyone interested, I finally found someone testing (Sapphire's) 6500 XT 8GB performance, it's hidden in this Arc A380 review:

Wow, nice find!

 

[TESKATLIPOKA]

RX 6500XT 8GB looks way better with more Vram.

I would love to see these specs for N34.
Full N34: 10 WGP; 20 CU; 2560 SP; 80-160? TMU; 32 ROP; 96-bit GDDR6 6GB Vram
cutdown N34: 8 WGP; 16CU; 2048 SP; 64-128? TMU; 32 ROP; 96-bit GDDR6 6GB Vram
Die size could be 120-130mm2.
I gave N34 more WGPs compared to N24, so the cut down version wouldn't end up just a bit faster than Phoenix IGP rumored to have 6 WGPs.
I wouldn't be surprised If there won't be any N34 chip.

 

[Joe NYC]

FYI, AMD officially announced the event for RDNA 3, "together we advance_gaming":

AMD to Host Livestream Event to Unveil AMD RDNA 3 Graphics Generation

Browse AMD’s company-wide and financial press releases.

ir.amd.com

The show premieres at 1:00 p.m. PDT on Thursday, November 3, on the AMD YouTube channel.

RDNA 3, Nov 3, 3pm Austin, TX time.

 

[Joe NYC]

N32 is later I reckon. If they stuck to the same timing as RDNA 2, that would be Marchish for N32. I could see earlier.

Delaying N32 months after N31 would be a mistake. But then, AMD marketing which come up with these plats is the weakest link at AMD, they seem to be screwing up everything.

Such as releasing B670E before there is anything to plug into the PCIe5 GPU and M.2 slots, with some 25 USB slots that no one needs, inflating the cost, while delaying the more sane and economical 650 series.

With RDNA3, delaying high volume N32, releasing low volume N31 and paying penalties to TSMC for not taking enough wafers would be another own goal of AMD marketing...

5800x3d launch fiasco, delaying it to the point of irrelevance, was another masterpiece of AMD marketing.

One has to really wonder Robert Hallock, AMD marketing director, left on his own or was pushed...

 

[Joe NYC]

Difference in gaming frequency is only 5% so 6800M is also pushed very high.
This just shows how much better a binned chip can be.

Now let's go back to greymon55's leak.
I have to wonder how high is this N32 actually clocked, because 80% higher performance at ISO power is not so much If I think about N32's specs.
N32 has 50% more WGPs than N22(30 vs 20) and each WGP should be significantly better.

Expected performance per ISO power(~150W)
RX 6800M : 100%
N32 mobile : +80%

N32 has +50% WGPs and If we say RDNA 3 WGP is 50% better than RDNA2 WGP then ideally that's 100*1.5*1.5=225% or +125% higher performance, but It's actually only +80%. I have to lower clockspeed by 20% from 2300MHz to 1840MHz to have the expected performance.

This leak from Greymom55 could actually be a cutdown version of N32, because I don't expect a mobile N32 to be downclocked from >3GHz down to <2GHz just to be within 145-165W, when N22 was downclocked by only a few %.

For top end mobile part, isn't it better to have full die down clocked than cutdown die clocked high? (From performance / power perspective)

 

[jpiniero]

Delaying N32 months after N31 would be a mistake.

It wouldn't be a delay. AMD did the same thing with the RX 5000 series - Navi 14 was 3 months later.

 

[TESKATLIPOKA]

For top end mobile part, isn't it better to have full die down clocked than cutdown die clocked high? (From performance / power perspective)

Yes, It would be better, but performance is king, so they clock everything high.

Delaying N32 months after N31 would be a mistake. But then, AMD marketing which come up with these plats is the weakest link at AMD, they seem to be screwing up everything.

Such as releasing B670E before there is anything to plug into the PCIe5 GPU and M.2 slots, with some 25 USB slots that no one needs, inflating the cost, while delaying the more sane and economical 650 series.

With RDNA3, delaying high volume N32, releasing low volume N31 and paying penalties to TSMC for not taking enough wafers would be another own goal of AMD marketing...

5800x3d launch fiasco, delaying it to the point of irrelevance, was another masterpiece of AMD marketing.

One has to really wonder Robert Hallock, AMD marketing director, left on his own or was pushed...

What?
Do you seriously think the marketing department is the one who decides when to release or delay a product?

 

[linkgoron]

Delaying N32 months after N31 would be a mistake. But then, AMD marketing which come up with these plats is the weakest link at AMD, they seem to be screwing up everything.

Such as releasing B670E before there is anything to plug into the PCIe5 GPU and M.2 slots, with some 25 USB slots that no one needs, inflating the cost, while delaying the more sane and economical 650 series.

With RDNA3, delaying high volume N32, releasing low volume N31 and paying penalties to TSMC for not taking enough wafers would be another own goal of AMD marketing...

5800x3d launch fiasco, delaying it to the point of irrelevance, was another masterpiece of AMD marketing.

One has to really wonder Robert Hallock, AMD marketing director, left on his own or was pushed...

Maybe stuff was just not ready? Do you have any proof that these things are delayed just because of marketing?

 

[Glo.]

For anyone interested, I finally found someone testing (Sapphire's) 6500 XT 8GB performance, it's hidden in this Arc A380 review:

View attachment 69633

And a few other more uncommon/older/light games tested, with those the difference is predictably smaller.
Testbed is Ryzen 5 3600X, B550 and 2x8GB DDR4-3200.

Now, looping back to N33 and discussions of 8x lane being a weakness, and similarly that 4x was the 6500 XT's biggest weakness, where I said I don't think that's true, the 4GB VRAM, in tandem with 4x lanes and too little effective bandwidth (144GB + 16MB IC) made a deadly performance-killing/stutter fest combo. Glad to see a proper frame buffer size fixed most of that and then some.
With the supposed end-to-end compression RDNA3 has, penalties for going more often out of VRAM and back would be lessen on a 4x, 4GB 6500XT, if it was hypoethically RDNA3 uArch.

There's no 3.0 vs 4.0 tests (this isn't a 6500 XT 8GB review after all), but I think we'd see an even wider gap between 4GB and 8GB.

Back when Sapphire's 6500 XT 8GB was announced, there were many saying so much VRAM is a waste and N24 can't make use of it lol

That being said, I am curious what will the eventual RX 7500/7400 family be based off... Cut down N33? Later announced N34 (similar to how N24 was only discovered much later)? N23 N6 refresh? And will AMD not mess up it's entry lineup again like that?

Its as fast as GTX 1660 Super.

P.S. There will not be Navi 34. There was development of this die, but no point for it when APUs will replace the low end.

Yes, N34 would be faster(around RX 6600 XT), but for segmentation reasons its better to hold down the low end for now.

P.S.2 - would anybody care that they have no dGPU if their APU performs like RTX 3050?

 

[Glo.]

RX 6500XT 8GB looks way better with more Vram.

I would love to see these specs for N34.
Full N34: 10 WGP; 20 CU; 2560 SP; 80-160? TMU; 32 ROP; 96-bit GDDR6 6GB Vram
cutdown N34: 8 WGP; 16CU; 2048 SP; 64-128? TMU; 32 ROP; 96-bit GDDR6 6GB Vram
Die size could be 120-130mm2.
I gave N34 more WGPs compared to N24, so the cut down version wouldn't end up just a bit faster than Phoenix IGP rumored to have 6 WGPs.
I wouldn't be surprised If there won't be any N34 chip.

Last time I heard, N34 was 2x of N24. 2048 ALUs, 128 bit bus, 8x PCIe 4.0. The project is scrapped. Phoenix is going to be used for this price segment.

 

[Joe NYC]

What?
Do you seriously think the marketing department is the one who decides when to release or delay a product?

For certain things - yes, I think so.

670 before 650
N31 before N32

Early low volume launch vs. later high volume launch

 

[Joe NYC]

Maybe stuff was just not ready? Do you have any proof that these things are delayed just because of marketing?

I don't think it is any more challenging to release lower end 650 vs. higher end 670, or lower end N32 vs. higher end N31. That was a marketing decision.

In the time of shortage, maybe it makes sense to release low volume product before high volume product.

But this time around, AMD is drowning in pre-purchased wafers from TSMC. Paying pre-payment penalties, while holding back easier the release, high demand, high volume product is not the best strategy...

This is not 2020-2021 anymore, when AMD could easily sell everything it could manufacture.

The depressed economic times in 2022 call for smart strategy, and AMD is not pursuing smart strategy - if, as some say, N32 will come in ~March.

 

[jpiniero]

But this time around, AMD is drowning in pre-purchased wafers from TSMC. Paying pre-payment penalties,

There hasn't been any speculation that they are doing any of the sort for N5.

Also I assume laptop OEMs will get first dibs on the N32 wafers.

 

[Aapje]

Companies will always release the best product first, unless that is a total dud (Hey Intel).

 

[Joe NYC]

There hasn't been any speculation that they are doing any of the sort for N5.

Also I assume laptop OEMs will get first dibs on the N32 wafers.

Not about AMD yet, but lots of smoke coming from others - NVidia, Intel, that they may end up paying penalties without taking the wafers.

It may happen to AMD as well.

Just months before "it" (with a sh) hit the fan, Graymon posted this:

https://twitter.com/x/status/1557025592075517952

That's a lot of wafers. No way AMD can come close to selling products at this pace.

- Genoa is delayed, and ramp of DDR5 based servers will be slow
- Zen 4 was released with the wrong chipset, suppressing demand
- N31 at $1,200 will only sell trivial number of cards

Without Zen 4 V-Cache and N32 available from January 1, I think AMD can sell ~15k wafers worth of products...

 

[RnR_au]

That's a lot of wafers. No way AMD can come close to selling products at this pace.

idk... I think AMD is very bullish on the server space. Only a few weeks ago in this forum...

Question - AMD to double EPYC production rate in 2023

https://www.nextplatform.com/2022/10/03/the-steady-hand-guiding-amds-prudently-expanding-datacenter-business/ I’m not sure if I’m reading it correctly. Forrest Norrod says “And just to be clear, we are planning for doubling year-on-year over time.” in response to the question of why AMD is...

forums.anandtech.com

 

[HurleyBird]

That's a lot of wafers. No way AMD can come close to selling products at this pace.

Of course they can. There's plenty of market share to take from Intel and Nvidia even in a retracting market, and AMD has cost advantages. They just need to stop clinging onto the fantasy of pandemic margins and get into the mindset of growing marketshare, even if it isn't immediately that profitable to do so. The hesitation to go down this path is understandable given that it has done little for AMD in the past, but the context is very different this time. With the recession the market will reward companies that provide better value, as opposed to punishing them with a value brand stigma, and given the probable relative strength of Zen 5 and RDNA4 against the competition, AMD should be able to hold onto any gains for a reasonable amount of time. The goal for AMD should be to greatly increase margins when the economy recovers, while keeping share it captured. To do that, AMD needs competitive products when that happens.

 

[Joe NYC]

idk... I think AMD is very bullish on the server space. Only a few weeks ago in this forum...

Question - AMD to double EPYC production rate in 2023

https://www.nextplatform.com/2022/10/03/the-steady-hand-guiding-amds-prudently-expanding-datacenter-business/ I’m not sure if I’m reading it correctly. Forrest Norrod says “And just to be clear, we are planning for doubling year-on-year over time.” in response to the question of why AMD is...

forums.anandtech.com

Dylan Patel posted some server forecast for 2023, and according to his forecast, only ~18% of server CPUs sold will be DDR5. That means (for AMD) that 82% of server CPUs AMD will be Milan and only 18% will be Genoa.

So even if AMD sales double, most of the sales will be N7 based Milan.

The 60k wafers for Q4 that Graymon mentioned are N5 wafers.

 

[Joe NYC]

Of course they can. There's plenty of market share to take from Intel and Nvidia even in a retracting market, and AMD has cost advantages. They just need to stop clinging onto the fantasy of pandemic margins and get into the mindset of growing marketshare, even if it isn't immediately super profitable to do so. The hesitation to go down this path is understandable given that it has done little for AMD in the past, but the facts are way different this time because with the recession the market will reward companies that provide better value, and given the probable relative strength of Zen 5 and RDNA4 versus the competition, AMD should be able to hold onto any gains for a good amount of time. The goal should be that when the economy does recover AMD is able to greatly increase margins while keeping share it captured during the recession.

I agree, but the key product is N32, which can sell 10x or even higher volume than N31. If N32 is not for sale (pronto), no amount of price cutting of N31 is going to bring sufficient volume.

On Zen 4 side, while AMD is still missing B650E motherboards and Zen 4 V-Cache, the sales will continue to be anemic.

The only one niche, that is being served reasonably well is high end, productivity niche, with 7950x. And again, like N31, this segment is tiny.

More mainstream, budget and gaming-oriented segments are served poorly.

So, imagine that Zen 4 V-Cache and N32 are released in March. That would mean AMD squandered 2 quarters of potentially good sales, mostly due to poor marketing decisions.

Nvidia has a huge gap in its product line up, where N32 would fit perfectly - if AMD released it timely.

 

[TESKATLIPOKA]

Last time I heard, N34 was 2x of N24. 2048 ALUs, 128 bit bus, 8x PCIe 4.0. The project is scrapped. Phoenix is going to be used for this price segment.

Phoenix has 6 WGPs so an 8WGP N34 would be kinda pointless.
Why would 8WGP, 16CU, 2048SP need 128 bit bus + 8x PCIe 4.0?
N33 also uses only 128 bit bus + 8x PCIe 4.0, yet It has 16WGP, 32CU, 4096SP. That doesn't make sense.
I also wonder If 8WGP N34 would really perform close to 16WGP N23(RX 6600XT).