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

 

[darkswordsman17]

Last we heard (rather recently), the parts hadn't even been taped out yet. So Q3 2022 at the earliest, I'd say.

When I was hitting some of the links people dropped about the greater than 2.5x perf stuff, looked like the rumors were saying Q4 2022. Right now, I have a hunch RDNA3 is gonna end up like Ampere. Good, but I think the reality is gonna temper people a lot. I expect them to be expensive and sure they might offer 2.5x perf but it isn't likely to live up to the 3x that the SP count would indicate and the price and other (multi-chiplet, 300-400W board) to achieve that is not gonna usher in some new era like say the VLIW4 days or R300.

Which that means the launch is likely gonna be holiday season, so prices and demand are gonna be stupid again, it might be 2 years from now before people can get 2x the perf of 6700XT for $500 (I'm honestly not even sure of that, it might end up being 50-75% perf boost, and while you might be able to easily get more perf it'll cost you a bunch extra to get it). I'd say I hope that we'll see a PS5 Pro pushing north of 15TF (rumors saying dual GPU? Which I'd expect lower clocks, but they should be able to get 15-20TF but it'll probably be even more expensive $750-800 MSRP with it offering roughly double the base PS5 perf for roughly 2x the price with a PS5 Slim with no UHD drive being $399) and hopefully the next PS VR headset is 4K. But guessing that hardware (either, let alone both) will be very difficult to come by as well and it won't be cheap either (I could see $1500), so that's probably not gonna be an easy alternative (I bought a Series X instead of trying to get an RDNA2 card last year, ended up just giving it to my nephew as there wasn't any games that interested me).

 

[JujuFish]

hopefully the next PS VR headset is 4K

The latest rumor I've seen is 2000x2040 per eye.

 

[Mopetar]

So let me get this straight, Navi 23 is ~23x mm^2 on TSMC N7, and now it's nominal successor is supposed to be nearly twice that size on more expensive TSMC N6? This is supposedly the "affordable" gpu for more mainstream gamers.

So either AMD struck the most lopsided contract alive with TSMC for wafers or that AMD isn't planning on selling any performance GPUs for less than $500 MSRP next generation...

It's a matter of economics and products as well. Once AMD releases APUs with RDNA2 graphics that use DDR5 memory the performance on those APUs will increase significantly.

If mining continues to be viable over that time span then there are no affordable mainstream cards because you either have to get lucky enough to snag one at MSRP (competing against all the scalpers and miners for these cards) or pay more than the MSRP due to markup by scalpers or merchants.

Releasing a $200 desktop card makes less and less sense every year.

 

[soresu]

that is not gonna usher in some new era like say the VLIW4 days

VLIW4 had an extremely limited run because of process woes, hardly a new era.

In fact Northern Islands gen only had the one high end discrete chip IIRC?

Trinity/Richland APUs gave it some further time in the spotlight, but it was otherwise really just a short lived mid point between VLIW5 and GCN/SI.

 

[soresu]

It's a matter of economics and products as well. Once AMD releases APUs with RDNA2 graphics that use DDR5 memory the performance on those APUs will increase significantly.

Rembrandt alone isn't nearly enough to fill that niche.

So Raphael better be a real monster of an APU to match that shortfall.

I'm hoping for at least 24 CU to double Rembrandt's rumoured raw silicon.

If the GCD also has V cache it could really fly with that.

Coupled with FSR it could do fairly well for gamers on a budget.

 

[Kepler_L2]

Rembrandt alone isn't nearly enough to fill that niche.

So Raphael better be a real monster of an APU to match that shortfall.

I'm hoping for at least 24 CU to double Rembrandt's rumoured raw silicon.

If the GCD also has V cache it could really fly with that.

Coupled with FSR it could do fairly well for gamers on a budget.

Raphael iGPU will be weaker than RMB.

 

[Leeea]

it might be 2 years from now before people can get 2x the perf of 6700XT for $500

You are way to optimistic.

Ryzen 3600 MSRP at release: $200
Ryzen 5600 MSRP at release: $299
+50% price for +20% performance

AMD is going charge a premium price on it for premium performance. rx6700 is MSRP $479, so that puts the next gen rx7600 at $718.

Ouch.

 

[beginner99]

I'm hoping for at least 24 CU to double Rembrandt's rumoured raw silicon.

I still fail to see the point for such powerful iGPUs. Almost all office PCs and laptops are more than fine with what we have currently. They do not need more 3D gpu power. In fact much more important are dedicated hardware pieces like in phone SOCs for hardware encode and decode and other accelerators. Think M1. But the actually problem in that is also on the software side, speak windows.
Anyway back to my initial point, large iGPU is a waste of silicon and will simply make the product too expensive for your mass market laptops which it needs to be sold in to make it financially viable.

You are way to optimistic.

Ryzen 3600 MSRP at release: $200
Ryzen 5600 MSRP at release: $299
+50% price for +20% performance

AMD is going charge a premium price on it for premium performance. rx6700 is MSRP $479, so that puts the next gen rx7600 at $718.

Ouch.

True but this is also mostly about supply and demand related to general chip shortage and not AMD milking us. Right now I'm glad if we can get rx6700 performance at msrp in one years time.

 

[moinmoin]

VLIW4 had an extremely limited run because of process woes, hardly a new era.

In fact Northern Islands gen only had the one high end discrete chip IIRC?

Trinity/Richland APUs gave it some further time in the spotlight, but it was otherwise really just a short lived mid point between VLIW5 and GCN/SI.

So it was kind of like Vega. Just that Vega got a pipecleaner bonus release that was even more quickly discarded.

 

[Zepp]

Ryzen 3600 MSRP at release: $200
Ryzen 5600 MSRP at release: $299
+50% price for +20% performance

AMD is going charge a premium price on it for premium performance. rx6700 is MSRP $479, so that puts the next gen rx7600 at $718.

3600X is a more direct comparison since there isn't a non-X 5600. It released at $250 which is a 20% difference.

 

[jpiniero]

Raphael iGPU will be weaker than RMB.

No question. Raphael is going to be whatever they think is the min CUs possible. But certinately more than good enough for anything other than gaming.

 

[Leeea]

3600X is a more direct comparison since there isn't a non-X 5600. It released at $250 which is a 20% difference.

On release lowest end 5000 part was the 5600x. (and it remained this way for nearly a year*)
On release lowest end 3000 part was the 3600.

I am comparing the cheapest 5000 to the cheapest 3000, which saw a 50% price increase generation over generation on release.

Yes, on GPUs they typically do not release the skews together. But it is even worse in GPU land, with the low end skews taking forever to come out.


*they just released the new cheapest 5000 skew, the 5600G, which comes in at $259, which is a 30% price 12% performance increase over the previous gen 3600 nearly a year later.


I love amd's products, I bought a amd gpu & cpu this generation. But they are not the value* leader they once were. As they take the lead they are also taking the premium on price. It is not wrong of them, but the trend indicates the next gen will cost considerably more then the current generation.

*but they are still far more consumer friendly, avoiding Intel's and Nvidia's consumer unfriendly product segmentation.

 

[Zepp]

On release lowest end 5000 part was the 5600x. (and it remained this way for nearly a year*)
On release lowest end 3000 part was the 3600.

I am comparing the cheapest 5000 to the cheapest 3000, which saw a 50% price increase generation over generation on release.

The 3600X is the most comparable SKU to the 5600X. They launched 1.4 years apart with the same price/performance.

 

[darkswordsman17]

The latest rumor I've seen is 2000x2040 per eye.

Ooh, nice. I was hoping they'd go that far. Any rumor about refresh rate? Even if the PS5 Pro brings a big boost (the talk of dual GPU makes me guess it'll be in the 15-18TF range), they'll still have to rely on upscaling/FSR to do even 4K/60 well. I would kinda love if Sony goes for broke and we get dual RDNA3 GPUs (~25TF) paired with 12core Zen3 V-Cache (since it seems good for games), and is $1000, with a very high end headset with that ~4K res at 120Hz and a good FOV and great tracking that potentially enables hand tracking. I'd pay $2000 for a system offering that if it came with some worthwhile content.

VLIW4 had an extremely limited run because of process woes, hardly a new era.

In fact Northern Islands gen only had the one high end discrete chip IIRC?

Trinity/Richland APUs gave it some further time in the spotlight, but it was otherwise really just a short lived mid point between VLIW5 and GCN/SI.

Er, so was it VLIW5? The 4000 and 5000 series, where we got a big step up in the 4000 series, and then doubled that shortly after. Of course Crossfire and SLI used to be worthwhile back then so you could get even another step up.

You are way to optimistic.

Ryzen 3600 MSRP at release: $200
Ryzen 5600 MSRP at release: $299
+50% price for +20% performance

AMD is going charge a premium price on it for premium performance. rx6700 is MSRP $479, so that puts the next gen rx7600 at $718.

Ouch.

We'll see. GPUs have been different. Generally we never see doubling of CPU performance within what, 5 years outside of some outliers. Whereas we are quite likely to see a doubling of the base processor units in GPUs over a couple of gens (although that's slowing down, we used to see that in a single gen). I don't see the RX7600 being over $500. But I won't be surprised a bit if the 7700 is in that range. Which that seems to be more the typical now, where you might get a big boost but it'll be a tier or two higher. And yeah, probably am being optimistic that we'll see that since I doubt we'll have RDNA 4 mid-range GPUs out by then. But competition should be in full swing by then, and I think we're gonna see a wild-card in Apple who seems like they might plan on disrupting things by packing a powerful GPU in their future M-series line (which sure isn't a direct competition to PC gaming GPUs, but sounds like Apple is gonna throw another big wrench in things).

 

[soresu]

So it was kind of like Vega. Just that Vega got a pipecleaner bonus release that was even more quickly discarded.

Nooooooo, Vega was not like VLIW4, process wasn't the main reason it had a limited SKU range.

Unlike VLIW4 Vega 14nm was definitely super b0rked.

The µArch and I think to some extent the driver too.

I think that was the real reason there wasn't a full line up for that µArch.

It seems from benchmarks like Vega 7nm fixed some of V14's problems but by then the damage was already done and it was indeed mostly a pipecleaner while RDNA1 was cooking (RDNA1 itself little more than a bridge to RDNA2).

 

[soresu]

Er, so was it VLIW5? The 4000 and 5000 series, where we got a big step up in the 4000 series, and then doubled that shortly after. Of course Crossfire and SLI used to be worthwhile back then so you could get even another step up.

No, VLIW5 and VLIW4 are related but separate and a significant change at the silicon/µArch level.

VLIW5 had 1 larger ALU and 4 smaller ALU (per SIMD?), whereas VLIW4 was 4 ALU's of equal performance/capability.

I think the big ALU in VLIW5 didn't get so much of a workout, but the smaller ALU's were perhaps a bit limited - so VLIW4 had a more balanced approach (for greater ALU utilisation?).

 

[Ajay]

We'll see. GPUs have been different. Generally we never see doubling of CPU performance within what, 5 years outside of some outliers. Whereas we are quite likely to see a doubling of the base processor units in GPUs over a couple of gens (although that's slowing down, we used to see that in a single gen). I don't see the RX7600 being over $500. But I won't be surprised a bit if the 7700 is in that range. Which that seems to be more the typical now, where you might get a big boost but it'll be a tier or two higher. And yeah, probably am being optimistic that we'll see that since I doubt we'll have RDNA 4 mid-range GPUs out by then. But competition should be in full swing by then, and I think we're gonna see a wild-card in Apple who seems like they might plan on disrupting things by packing a powerful GPU in their future M-series line (which sure isn't a direct competition to PC gaming GPUs, but sounds like Apple is gonna throw another big wrench in things).

I am wondering about pricing as well (since I've written off this generation). The issue is that AMD is a for profit company - I'm sure investors are not thrilled watching retailers soaking up all the extra $$ that the market will currently bear because of mining sales. If mining sales drop off soon (proof of state, etc.), then I expect reasonable pricing. If GPU mining doesn't wind down, I expect fairly expensive GFX cards - like +$200-$300 on a 7700XT. And, yes, that would make me very sad.

 

[Joe NYC]

Wow, didn’t realize it hasn’t taped out yet! I thought it was typically 12-15 months to reach volume production after tapeout + time to fill the channel. Maybe my memory is betraying me.

Look at the bright side. RDNA is going to completely leapfrog NVidia, with chiplet partitioning, 3D stacked L3, hybrid bonded communication link between GCP dies:

WildCracks on Twitter: "@nerdtechgasm For Navi 31 and 32 (MCM RDNA 3): 2 5nm GCDs with Command Processor, Shader Engines, I/O, Memory PHYs, L2$, etc. 1 (or multiple?) 6nm MCD with Infinity V-Cache, 3D Stacked on top of both GCDs like in the image Navi 33 is Monolithic 6nm Aldebaran CDNA 2 is 2 7nm die apparently https://t.co/nLx7akKbOd" / Twitter

 

[Ajay]

Look at the bright side. RDNA is going to completely leapfrog NVidia, with chiplet partitioning, 3D stacked L3, hybrid bonded communication link between GCP dies:

WildCracks on Twitter: "@nerdtechgasm For Navi 31 and 32 (MCM RDNA 3): 2 5nm GCDs with Command Processor, Shader Engines, I/O, Memory PHYs, L2$, etc. 1 (or multiple?) 6nm MCD with Infinity V-Cache, 3D Stacked on top of both GCDs like in the image Navi 33 is Monolithic 6nm Aldebaran CDNA 2 is 2 7nm die apparently https://t.co/nLx7akKbOd" / Twitter

Impressive, if true! Reading the Beyond3d thread on Lovelace, Nvidia isn’t sitting on its buttocks either. https://forum.beyond3d.com/threads/...ace-speculation-rumours-and-discussion.62474/

The RDNA3 thread is totally blowing up though!

 

[beginner99]

Impressive, if true! Reading the Beyond3d thread on Lovelace, Nvidia isn’t sitting on its buttocks either. https://forum.beyond3d.com/threads/...ace-speculation-rumours-and-discussion.62474/

The RDNA3 thread is totally blowing up though!

I mean it makes sense. AMD is a tech/hardware company while Nvidia is a solutions/software company.

 

[Ajay]

I mean it makes sense. AMD is a tech/hardware company while Nvidia is a solutions/software company.

As AMD grows, it needs to grow into the solution/software space. [which is part of why I think the $4B stock buyback plan was moronic**]


** Unless you own some of AMD stock

 

[Joe NYC]

As AMD grows, it need to grow in the solution/software space. [which is part of why I think the $4B stock buyback plan was moronic**]


** Unless you own some of AMD stock

As far as the stock buyback, that's how much was authorized. 4 billion. But AMD does not have to spend all the money.

Someone posted the ranges of purchases, and they were all near 52 week low of the stock. The stock is now trading at the opposite end of the spectrum, near 52 weak high. I don't expect AMD to be buying any shares at this level.

 

[Ajay]

As far as the stock buyback, that's how much was authorized. 4 billion. But AMD does not have to spend all the money.

Someone posted the ranges of purchases, and they were all near 52 week low of the stock. The stock is now trading at the opposite end of the spectrum, near 52 weak high. I don't expect AMD to be buying any shares at this level.

In that case, it was a brilliant move. Pump up the stock, but only buy it back in drips and drabs citing the high stock price as an impediment.

 

[Mopetar]

In that case, it was a brilliant move. Pump up the stock, but only buy it back in drips and drabs citing the high stock price as an impediment.

I mean that's kind of the purpose of a buyback. There's no point in doing it when the price is high unless you have good reason to believe the entire market is wrong and undervaluing the company.

Otherwise if you wanted to return money to shareholders a dividend makes more sense.

 

[Joe NYC]

I mean that's kind of the purpose of a buyback. There's no point in doing it when the price is high unless you have good reason to believe the entire market is wrong and undervaluing the company.

Otherwise if you wanted to return money to shareholders a dividend makes more sense.

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