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

 

[Mopetar]

It also ended the bear / short attacks on AMD stock.

AMD is on a pretty steady upwards trajectory as far as growth is concerned so anyone trying to do that strikes me as pretty foolish.

 

[Joe NYC]

AMD is on a pretty steady upwards trajectory as far as growth is concerned so anyone trying to do that strikes me as pretty foolish.

Some things that seem obvious in retrospect were not so clear prospectively. Just a few hurdles that AMD overcame in last 6 months, which could have tripped up AMD
- Rocket Lake
- Ice Lake
- capacity

Just take the capacity. If AMD showed no growth for 2 quarters, because it could not secure any additional capacity from TSMC, the bears / shorts would have made a killing, and AMD stock would have been half of what it is.

 

[Mopetar]

Some things that seem obvious in retrospect were not so clear prospectively. Just a few hurdles that AMD overcame in last 6 months, which could have tripped up AMD
- Rocket Lake
- Ice Lake
- capacity

Just take the capacity. If AMD showed no growth for 2 quarters, because it could not secure any additional capacity from TSMC, the bears / shorts would have made a killing, and AMD stock would have been half of what it is.

What Intel did or might do would have little bearing on AMD. They hadn't beat Intel before Zen 3 and their revenue and profits were only going up. The notion that if Intel were once again overtake AMD it would lead to some kind of downfall for the company is rather shortsighted and anyone wanting to bail out for that deserves to lose money for their lack of knowledge.

Really the only concern is their ability to grow which is the third point. Now that AMD is making more than can afford to invest more. Considering they're the partner that unveiled TSMC's 3D stacking technology does seem to suggest they're a fairly important partner to TSMC.

The notion that shot sellers would have stopped AMD from rising is simply ludicrous. Maybe they can wreck a failing company to profit a bit more off the corpse, but you can't reliably ruin a profitable and growing company that way. You'll just get screwed as the company delivers and people want to buy the stock.

 

[Joe NYC]

What Intel did or might do would have little bearing on AMD. They hadn't beat Intel before Zen 3 and their revenue and profits were only going up. The notion that if Intel were once again overtake AMD it would lead to some kind of downfall for the company is rather shortsighted and anyone wanting to bail out for that deserves to lose money for their lack of knowledge.

These investors make subjective decisions, and, would you be surprised that some came to believe that with Pat Gelsinger, IDM 2.0, Intel is back and will crush AMD day after tomorrow - rather than a realistic time frame of 2+ years ago when things CAN turn around for Intel? And as a consequence, AMD loses steam.

Really the only concern is their ability to grow which is the third point. Now that AMD is making more than can afford to invest more. Considering they're the partner that unveiled TSMC's 3D stacking technology does seem to suggest they're a fairly important partner to TSMC.

I believe it, but there are others who think AMD is just another customer.

Other than AMD unveiling TSMC's 3D stacking, AMD was able to double it's capacity from TSMC, in time of shortage, in last 12 months.

But suppose that AMD did not get this capacity, and the revenue stagnated for the last 6-12 months.

With lack of capacity, revenue and profits would stall, and AMD stock, priced on high growth would have gone in reverse.

The stock price is a combination of current earnings and prospect of revenue and earnings growth. It is the growth potential why AMD is priced at much greater multiple than Intel, and why Nvidia is priced at even higher multiple than AMD.

 

[Mopetar]

If you believed anything of what you just typed why didn't you use that to invest accordingly?

Maybe it sounds good when you type it, but if you believe it why not invest in that belief accordingly. Talk is cheap.

 

[Joe NYC]

If you believed anything of what you just typed why didn't you use that to invest accordingly?

Maybe it sounds good when you type it, but if you believe it why not invest in that belief accordingly. Talk is cheap.

No, none of those were my opinions, I thought the opposite. I am just describing opinions of the others, that were not 100% nuts at the time, as it would seem in retrospect.

 

[Joe NYC]

Not that this is a huge surprise, but software support for the the upcoming RDNA 3 chips, Navi 31 and Navi 33 is making its way to some of AMD software tools:

AMD Next-Gen RDNA 3 Navi 31 & Navi 33 GPUs Listed In ROCm OpenGPU Platform (wccftech.com)

 

[Gideon]

Any guesses what the MSRP of Navi 31 would be?

Of which I'm 100% certain is that even in a normal market (e.g. no shortage) it would not be a Navi 21 replacement but a level up in pricing. Cosidering the amount of silicon and packaging effort involved (essentially a GPU threadripper)

My wild guess, without the GPU shortage around, it will be:

• $1799 for the top "gaming card"
• $1299 for the 6800 (non-xt) replacement
• N32 filling the N21 price bracket

I'm sure people will yell and stomp their feet at these prices, but that's a totally realistic ballpark. In the current GPU situation with the rumoured performance i could easily see $2500 - 3000 for the top of the line card

 

[Timorous]

Any guesses what the MSRP of Navi 31 would be?

Of which I'm 100% certain is that even in a normal market (e.g. no shortage) it would not be a Navi 21 replacement but a level up in pricing. Cosidering the amount of silicon and packaging effort involved (essentially a GPU threadripper)

My wild guess, without the GPU shortage around, it will be:

• $1799 for the top "gaming card"
• $1299 for the 6800 (non-xt) replacement
• N32 filling the N21 price bracket

I'm sure people will yell and stomp their feet at these prices, but that's a totally realistic ballpark. In the current GPU situation with the rumoured performance i could easily see $2500 - 3000 for the top of the line card

If top N31 does get close to 2.7x then calling it an N21 replacement seems a bit off since that performance uplift would be unprecedented. The closest you could get is 9700 Pro or 8800GTX and even they were only upto ~2.5x in outliers and were closer to ~2x most of the time.

Scaling from N21 MSRP I would think that a 50% gen on gen uplift is considered pretty good in a given price bracket. That puts 1.5x 6900XT at around $1000. If AMD get close to 2.7x then I agree that $2,000 MSRPs may be on the table for the top part sans current market conditions, maybe more as that would be close to 1:1 price/perf scaling which never happens for halo products.

Not sure how AMD are going to build their stack but I could see 79xx and 78xx being built from various N31 MCM configs and then 77xx being built from N33.

 

[Stuka87]

Any guesses what the MSRP of Navi 31 would be?

Of which I'm 100% certain is that even in a normal market (e.g. no shortage) it would not be a Navi 21 replacement but a level up in pricing. Cosidering the amount of silicon and packaging effort involved (essentially a GPU threadripper)

My wild guess, without the GPU shortage around, it will be:

• $1799 for the top "gaming card"
• $1299 for the 6800 (non-xt) replacement
• N32 filling the N21 price bracket

I'm sure people will yell and stomp their feet at these prices, but that's a totally realistic ballpark. In the current GPU situation with the rumoured performance i could easily see $2500 - 3000 for the top of the line card

How is an 225% price increase "totally realistic"?? Are you also suggesting a 6600 replacement is going to be $700? If mining is not around, why on earth would anybody pay such a high price for a card? Yes, people have payed 1000-1500 for the absolute best card ever. But to pay that for a mid range card is crazy.

Where is the idea behind these prices coming from?

 

[Timorous]

How is an 225% price increase "totally realistic"?? Are you also suggesting a 6600 replacement is going to be $700? If mining is not around, why on earth would anybody pay such a high price for a card? Yes, people have payed 1000-1500 for the absolute best card ever. But to pay that for a mid range card is crazy.

Where is the idea behind these prices coming from?

If N31 is the unprecedented performance uplift some rumours say it is, 2.7x has never been seen before as an average in raster in the history of GPUs, then in effect top end N31 is a brand new tier of part that never existed before and could occupy a new pricing tier that obviously exists because people have paid $2k for GPUs in the current climate.

So given that it means AMD either re-align their stack so 79 = new tier, 78 replaces 69, 77 replaces 68, 76 replaces 67 etc or they keep the alignment and top tier N31 gets a fancy name like Fury or Vega or something to distinguish it from the rest of the stack.

OTOH the rumours could be bogus and we are looking at a more standard 1.7x performance uplift which might not see much MSRP inflation.

 

[Stuka87]

If N31 is the unprecedented performance uplift some rumours say it is, 2.7x has never been seen before as an average in raster in the history of GPUs, then in effect top end N31 is a brand new tier of part that never existed before and could occupy a new pricing tier that obviously exists because people have paid $2k for GPUs in the current climate.

So given that it means AMD either re-align their stack so 79 = new tier, 78 replaces 69, 77 replaces 68, 76 replaces 67 etc or they keep the alignment and top tier N31 gets a fancy name like Fury or Vega or something to distinguish it from the rest of the stack.

OTOH the rumours could be bogus and we are looking at a more standard 1.7x performance uplift which might not see much MSRP inflation.

1: Gamers are no spending $2K on GPUs, miners are. Sure there are 1%ers spending whatever, but the vast majority are not.

2: Every time we get rumors of a new chip, its said to be amazing and super ultra fast.

3: Just because the N31 is that fast, does not mean the pricing of their entire stack shifts upwards by 225%. Maybe that one card because AMD's new Fury card (akin to Titan). And everything else stays right where it is.

 

[eek2121]

1: Gamers are no spending $2K on GPUs, miners are. Sure there are 1%ers spending whatever, but the vast majority are not.

2: Every time we get rumors of a new chip, its said to be amazing and super ultra fast.

3: Just because the N31 is that fast, does not mean the pricing of their entire stack ships upwards by 225%. Maybe that one card because AMD's new Fury card (akin to Titan). And everything else stays right where it is.

You are foolish if you think this. Scalpers are making good money from gamers.

The top end RDNA3 chip, if it performs as rumored (I have doubts), will not be cheap.

Gamers buy RTX 3090s all the time. I use my 3090 for other things, but even if I did not, I would STILL have purchased the 3090.

If AMD rolls out a $2,000 part that is nearly 3X faster in gaming, I will probably buy it.

 

[Mopetar]

The problem will be that even if it's a worthy $2,000 gaming card, that probably makes it at least a $4,000 mining card.

 

[Stuka87]

You are foolish if you think this. Scalpers are making good money from gamers.

The top end RDNA3 chip, if it performs as rumored (I have doubts), will not be cheap.

Gamers buy RTX 3090s all the time. I use my 3090 for other things, but even if I did not, I would STILL have purchased the 3090.

If AMD rolls out a $2,000 part that is nearly 3X faster in gaming, I will probably buy it.

I am sure lots of gamers are spending 500-600 on RTX 3060's and such. Very very few gamers buy cards that are over $750 in a regular market, much less an over inflated market.

 

[Mopetar]

There were plenty of people balking at the $700 price tag of the 3080 when it launched. No one close to that mindset is paying even more for a 3060.

It's quite clear that miners are driving the market price because if you look at the relationship between cards and their prices you'll see that MH/s is what largely determines the price because that's what primarily dictates how long it will take a miner to see a return on their investment at a given cryptocurrency price.

 

[blckgrffn]

Miners might be driving it but I have a good friend that bought a $4k prebuilt for 3090 and helped another friend and a cousin secure 3080's, one for $1200 the other for $1k.

Zero mining, all gaming.

15% of American households have a net worth of over $1M. People buy Jet Skis and other toys for many, many thousands (tens!) of dollars to use twice a year with no economic (mining type) application.

PC parts, even now, are a "cheap" hobby compared to many high end hobbies and with the prices of old parts increasing, it becomes even more "obvious" the best time to get a new gaming card is at release while the terrible supply issues keep the value of the old hardware inflated.

Point being - plenty of people will buy these uber-cards at uber-prices. I didn't chime in on @moonbogg thread on prices "should be higher" but I believe they will be. I think the PS5 Pro and xbox series XXX could easily be $1k consoles and sell out.

The 90's with "good computers" starting at like $1500+ and stretching to $3K+ seem to be staring us in the face again - IMO.

 

[Stuka87]

Miners might be driving it but I have a good friend that bought a $4k prebuilt for 3090 and helped another friend and a cousin secure 3080's, one for $1200 the other for $1k.

But that is anecdotal information. Just because a subset of gamers are willing to spend that much, does not mean the masses will. And it certainly does not mean OEM's are going to increases MSRPs by 225% as suggested above.

 

[blckgrffn]

But that is anecdotal information. Just because a subset of gamers are willing to spend that much, does not mean the masses will. And it certainly does not mean OEM's are going to increases MSRPs by 225% as suggested above.

I gave non-anecdotal information about the financial situation in the US. Read the rest of my post.

OEMs and boutique shops sell PCs for $2000-$4000 constantly, right now. You are right I don’t know the sales mix but they are frequently unavailable for purchase and they certainly have had more cards made available to them than the typical enthusiast.

My point is that we cannot know if it (much higher MSRP graphics cards) will fail until it’s been tried, and we can be sure the marketing and sales departments are well aware of the potential price points currently untapped and that this information is informing decisions. I cannot fathom that both AMD wouldn’t have products aimed squarely at $2k MSRP next go round with retail 6900xts and 3090s selling above that with little issue.

And by masses... what does a card need to sell to be a Halo like a Titan Black or something? 10K units? 100K units? 1MM units? Frankly I don't know but somewhere there are marketing and engineering teams doing the math.

 

[jpiniero]

There's also ProViz to sell to.

 

[Kepler_L2]

Any guesses what the MSRP of Navi 31 would be?

Of which I'm 100% certain is that even in a normal market (e.g. no shortage) it would not be a Navi 21 replacement but a level up in pricing. Cosidering the amount of silicon and packaging effort involved (essentially a GPU threadripper)

My wild guess, without the GPU shortage around, it will be:

• $1799 for the top "gaming card"
• $1299 for the 6800 (non-xt) replacement
• N32 filling the N21 price bracket

I'm sure people will yell and stomp their feet at these prices, but that's a totally realistic ballpark. In the current GPU situation with the rumoured performance i could easily see $2500 - 3000 for the top of the line card

$1999 minimum.

The problem will be that even if it's a worthy $2,000 gaming card, that probably makes it at least a $4,000 mining card.

Navi31 will have the mining perf of the RTX 3060Ti, maybe *slightly* higher. It will probably be the worst MH/$ GPU.

 

[Bigos]

Can I ask for any "prediction" to be mentioned as that? The certainty of some posters here is frightening.

Talking about things that do not exist and speaking as if your live hangs on that is not smart, it is just sad.

 

[ModEl4]

If the rumors are true for a 440mm2 128bit memory bus etc based Navi 33, then I think is very difficult for Navi 32 (2x) to be below 690mm2 and Navi 31 below 940mm2 (including everything, infinity cache etc). In David Wang presentation there was a slide shown indicating +50% performance/watt initial target for RDNA3 vs RDNA2 (not present in Tech ARP video with the static split screen slides...) and also Rich Bergman indicated similar performance/watt improvements as with RDNA2 vs 1

AMD is expecting the next-gen RDNA 3 GPUs scheduled for 2021 to deliver 50% improved performance-per-watt over the new RX 6000 series

We have not even got a chance to explore the limits of the new RX 6000 GPUs based on the RDNA 2 architecture, and AMD is already planning some impressive performance gains for the upcoming RDNA 3 cards that should release in late 2021. The performance uplift should be facilitated through MCM...

www.notebookcheck.net

I think AMD will be able to double (+100%) the performance/watt vs RDNA2 for some of the MCM designs so at 420W TBP (which I think is the limit for AMD reference designs with 3x 8pins) we will have 2.8X max increase in performance vs 6900XT.
So according to TPU Nvidia will need 2.6x the performance of 3090 to match Navi 31 at 4K:

ZOTAC GeForce RTX 3090 AMP Extreme Holo Review

The ZOTAC RTX 3090 AMP Extreme Holo is the longest graphics card we ever tested. At more than 350 mm long, this triple-fan, triple-slot monster has enough power to handle all the latest games at 4K resolution. The RTX 3090 GPU is highly overclocked, and the power limits have been raised, too.

www.techpowerup.com

I guess many people they don't realize that Nvidia (if rumors are true also and we are talking about a 5nm TSMC based design with 18432 Cuda cores etc) can do a Maxwell2->Pascal minor features facelift add 96MB of NV cache lol and still be at 471mm2 size EASILY at 5nm TSMC. And also the 2.6X is quiet doable for NV depending the clocks their designs can hit.
My prediction is 2 chip designs in 5nm TSMC and 3 designs with 5nm Samsung probably all with NV cache (for TSMC is certain imo) and we are talking about minor die area anyway.
1)TSMC 5nm 192 ROPs/18432CC 384bit memory bus with 96MB NV cache
2)TSMC 5nm 128 ROPs/12288CC 256bit memory bus with 64MB NV cache
etc...
The 3rd (Samsung based) will have higher than 3090 performance at $499 and the 2nd TSMC based will have around 1.7X higher performance according to my napkin calculations. I'm a little bit worried about the size of cache but if AMD can support 1080p with 16MB for Navi24 and with 32MB 1440p with only 2-3% hit on average then at 4K the 64MB will be enough with similar minor hit (also Nvidia will have much higher additional memory bandwidth through the regular bus vs AMD designs so maybe 96MB for 2880p and 64MB for 2160p will be enough with minor hit) So similar performance in the end at maybe half the die size? (or with sizeable features gains otherwise) is this a success for AMD?

 

[leoneazzurro]

If the rumors are true for a 440mm2 128bit memory bus etc based Navi 33, then I think is very difficult for Navi 32 (2x) to be below 690mm2 and Navi 31 below 940mm2 (including everything, infinity cache etc). In David Wang presentation there was a slide shown indicating +50% performance/watt initial target for RDNA3 vs RDNA2 (not present in Tech ARP video with the static split screen slides...) and also Rich Bergman indicated similar performance/watt improvements as with RDNA2 vs 1

AMD is expecting the next-gen RDNA 3 GPUs scheduled for 2021 to deliver 50% improved performance-per-watt over the new RX 6000 series

We have not even got a chance to explore the limits of the new RX 6000 GPUs based on the RDNA 2 architecture, and AMD is already planning some impressive performance gains for the upcoming RDNA 3 cards that should release in late 2021. The performance uplift should be facilitated through MCM...

www.notebookcheck.net

I think AMD will be able to double (+100%) the performance/watt vs RDNA2 for some of the MCM designs so at 420W TBP (which I think is the limit for AMD reference designs with 3x 8pins) we will have 2.8X max increase in performance vs 6900XT.
So according to TPU Nvidia will need 2.6x the performance of 3090 to match Navi 31 at 4K:

ZOTAC GeForce RTX 3090 AMP Extreme Holo Review

The ZOTAC RTX 3090 AMP Extreme Holo is the longest graphics card we ever tested. At more than 350 mm long, this triple-fan, triple-slot monster has enough power to handle all the latest games at 4K resolution. The RTX 3090 GPU is highly overclocked, and the power limits have been raised, too.

www.techpowerup.com

I guess many people they don't realize that Nvidia (if rumors are true also and we are talking about a 5nm TSMC based design with 18432 Cuda cores etc) can do a Maxwell2->Pascal minor features facelift add 96MB of NV cache lol and still be at 471mm2 size EASILY at 5nm TSMC. And also the 2.6X is quiet doable for NV depending the clocks their designs can hit.
My prediction is 2 chip designs in 5nm TSMC and 3 designs with 5nm Samsung probably all with NV cache (for TSMC is certain imo) and we are talking about minor die area anyway.
1)TSMC 5nm 192 ROPs/18432CC 384bit memory bus with 96MB NV cache
2)TSMC 5nm 128 ROPs/12288CC 256bit memory bus with 64MB NV cache
etc...
The 3rd (Samsung based) will have higher than 3090 performance at $499 and the 2nd TSMC based will have around 1.7X higher performance according to my napkin calculations. I'm a little bit worried about the size of cache but if AMD can support 1080p with 16MB for Navi24 and with 32MB 1440p with only 2-3% hit on average then at 4K the 64MB will be enough with similar minor hit (also Nvidia will have much higher additional memory bandwidth through the regular bus vs AMD designs so maybe 96MB for 2880p and 64MB for 2160p will be enough with minor hit) So similar performance in the end at maybe half the die size? (or with sizeable features gains otherwise) is this a success for AMD?

Except that your speculation is quite likely to be wrong on several levels. Starting with die sizes and ending with performances. So let's take this apart one by one.

1) Die sizes and costs

First, rumors for the top end NV part speak about 144 Ampere-like SM. 82 of such SM took 626mm^2 on 8nm Samsung process with 28,3 Billion transistors with a density of 45,3 Million transistors per mm^2. Navi 21 on 7nm TSMC gets 28.6 billion transistors on a 520 mm^2 die, with a density of 51.2 million transistors for mm^2. Let's say TSMC 5nm will scale with the promised 1,8x compared to 7nm. That is a density of 93 Million transistors per mm^2. So if we consider everything scaling linearly, you get around 50 Billion transistors and a die size of around 540 mm^2. Without adding any cache. Problem is, not everything scales linearly, especially the memory intefaces. And internal buses do not scale linearly. And 96 Mbytes of cache is quite LOW for a 4K part (even Navi21 has 128, Navi22 has 96 for 1440p and NAvi 23 32Mbytes for 1080p). If you add also 96 Mbytes of cache to the mix and consider the not linear scaling you'll end up easily on more than 600 mm^2 on TSMC 5nm process. Where you took your 471mm^2 figure is beyond me. Another problem with your cost estimations is that you do not consider that a 5nm wafer from TSMC costs way more than a 7nm wafer ( around 1,5x) and that a 7nm wafer from TSMC costs way more than a 8nm wafer from Samsung (it is almost double according to some leaks). It does not take much for understanding that card prices will not go down. Same for a "5nm" Samsung process, considering that by area and power it seems to be roughly on par with TSMC's 7nm https://www.anandtech.com/show/16463/snapdragon-888-vs-exynos-2100-galaxy-s21-ultra/2. So 3090 performance for 499$ seems to be a wishful thinking. I'd more than welcome that, for sure. But realistically it's not happening. On AMD's side, rumors say only the Graphic compute dies will be on 5nm, while cache will be on 6nm (cheaper process) and 30 WGP/"120CU" dies are rumored to be less than 250 mm^2. Which means way better yields and lower die costs. Adding to this, we have some "fixed" costs that are not depending on the die, like RAM, PCB, cooling system. As next-gen parts are rumored to be even more power-hungry than current ones, thus needing beefier power sections and coolers, I'd say this does not bode well for getting cheaper cards, on both sides. For sure, AMD packaging for N31 and N32 will cost more. But it will have quite probably less 5nm silicon. So it's likely to cost more, but on production costs it's not a black and white situation.

2) Performances
At this moment, in rasterization, a RDNA2 CU is only slightly less powerful than an Ampere SM (80 RDNA2 CUs in the 6900XT vs 82 SM in the 3090). Yes, a Nvidia SM will likely gain a bit more than AMD going for the new process, but rumored configurations of the top cards say 120 "CU" per die (240 "CU" total) for Navi31 and 144 SM for Nvidia. You'll need a 60% improvement (by means of clocks, IPC, and more) only for getting a tie in that department. That supposing RDNA3 will not bring any improvement either. And that supposing that we'll have no other limiting factor on both sides. So we don't know performances but on the brute force approach AMD configuration seems having the upper hand, even the second top card is rumored to have 160 "CU". On RT sides, we know nothing except that in the past Nvidia performed way better per SM than RDNA2 per CU. We don't know if with RDNA3 it will be the same or not.

 

[Mopetar]

$1999 minimum.

Navi31 will have the mining perf of the RTX 3060Ti, maybe *slightly* higher. It will probably be the worst MH/$ GPU.

What're you expecting the memory bandwidth to be for the card. I don't know what Navi 31 will ultimately be, but a two or three die MCM approach seem to be bandied about as the most likely.

To limit mining performance to that of a 3060 Ti essentially means no more than a 256-bit memory bus. The 6600 XT does surprisingly well with only a 128-bit bus, but you'd need a lot more cache to keep that GPU from starving with the other rumored specs.

Even if the cost of the GPU is high, as long as it gets a certain MH/s per dollar it will float to that price. The only good news is that price may just be a smaller fraction of the initial MSRP. If it went from $2,000 to something like $2,400 then no one would really complain given some AIB models would probably encroach on that price anyways.

If it had three dies, each with a 128-bit bus the total memory bandwidth would be equivalent to a 3080 Ti and those currently go for around $2,200 or more so it certainly is possible for AMD to launch a card that wouldn't inflate in price too much.