Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

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...

forums.anandtech.com

 

[Mahboi]

No.

no.

Use words

 

[Aapje]

Lots of laptops are bought by light gamers, who play a bit of Minecraft or The Sims in the evening. Lots of women too, who don't see themselves as hardcore gamers and don't want to be the kind of person who spends lots on gaming stuff, as they see that as childish/male (same thing, really).

So lots of 'I need a laptop anyway, but why not get one that is a little better.'

 

[adroc_thurston]

Use words

I think "no" means "no" but I digress.

 

[Mahboi]

I think "no" means "no" but I digress.

Use words

 

[adroc_thurston]

Use words

'no' is a word.

 

[Thunder 57]

Lots of laptops are bought by light gamers, who play a bit of Minecraft or The Sims in the evening. Lots of women too, who don't see themselves as hardcore gamers and don't want to be the kind of person who spends lots on gaming stuff, as they see that as childish/male (same thing, really).

So lots of 'I need a laptop anyway, but why not get one that is a little better.'

So simply being male means one is childish? That's the stupidest thing I've read lately. Maybe you've been watching too much TV where all the commercials depict dads as being buffoons.

 

[Joe NYC]

Same as desktop GPU where the >=80 class is not even 25% of the market.
Arguably worse even since a ton of gamers don't wanna bother themselves with a laptop with a "strong" GPU.
You can probably bet ~90% of units is 70 class down to APU and you won't be far off.

Yeah, that's what I was thinking too as far a share.

But on the first go of a strong APU? How's that supposed to work out to good margins?
You're cramming 40 CUs into a compact die alongside 4 Z5 cores, 8 Z5c cores, 4 Z5LP cores. That is a lot of silicon to feed in a laptop even if it's a fat one, especially all under one fan.
I don't picture this thing being cheap to make, I don't picture the packaging being cheap, nor do I picture the cooling system being cheap. Maybe over time you can streamline and have a ton of reused designs, but I can't picture it being high margin on the first go, more like an expensive experimentation that'll get cheaper as time goes on.

Strix Halo is chiplet based, so the Zen 5 cores are on a separate die(s). But the SoC die is likely going to be quite sizeable. I have seen estimates of 200-250 mm2 and (N3E based on some rumors).

So not a cheap die. To design and manufacture. AMD is trying to "seed" this market, and it is going to have costs...

Follow up versions will then likely go for increased efficiency... Maybe use N6 for all the IO, SRAM etc.

 

[gdansk]

So simply being male means one is childish? That's the stupidest thing I've read lately. Maybe you've been watching too much TV where all the commercials depict dads as being buffoons.

At times like this it's best not to be offended. Read it as charitably as one can: most people don't want gamer laptops.

 

[Mahboi]

So not a cheap die. To design and manufacture. AMD is trying to "seed" this market, and it is going to have costs...

Follow up versions will then likely go for increased efficiency... Maybe use N6 for all the IO, SRAM etc.

Yes that's also what I'm thinking. I don't see the first chip to be cheap to make, and I don't see it being so much better that it'll BTFO the competition, so I don't see how you could have high margins with this.
Later down the line, sure, but I don't see it for Halo itself with the info we have here.

 

[Thunder 57]

At times like this it's best not to be offended. Read it as charitably as one can: most people don't want gamer laptops.

I just found it to be an odd statement. I agree about the gamer laptop part. I don't want the word "gaming" on my stuff, or RGB anything. I've said before how I hate how companies throw gaming on everything, like "gaming memory", or "gaming SSD". And I criticized AMD when they called their L3 cache "GameCache". Thankfully that term didn't stick around long.

 

[Joe NYC]

Yes that's also what I'm thinking. I don't see the first chip to be cheap to make, and I don't see it being so much better that it'll BTFO the competition, so I don't see how you could have high margins with this.
Later down the line, sure, but I don't see it for Halo itself with the info we have here.

I think a better way to look at the cost comparison is to start with a baseline, Strix Point, and then calculating what it takes to turn it into Strix Halo, as far as functionality. Vs. adding a 4060 or 4070 as a separate chip.

From the POV of GPU performance:

Strix Halo = Strix Point (16 CU) + 24 CU + 128 bit memory width

Since 4060 already has 128 bit memory access, the difference on the die is 24 CUs, which could be 50-60 mm2.

On 4060 side, if we subtract memory controllers, of, let's say 20-30 mm2 and 50-60 mm2 of additional graphics resources (added to Strix), and the full 4060 is ~160 mm2, the duplication ends up being:
160 - 20 - 50 = 90 mm2 on the high end
160 - 30 - 60 = 70 mm2 on the low end

Strix Halo may have other goodies, but those are beside the point.

Back to Strix Halo Margin - AMD margin, if Strix Halo sells at the same price as Strix Point + 4060, the margin would be:

StrixPoint Margin + NVidia 4060 margin + ~80 mm2 die area savings

On the negative side of the ledger would be cost differential between N4p (of 4060) and N3E of Strix Halo, if that's what it is.

So, it does not look so bad, and might be accretive to AMD margins if AMD sells a lot of them.

 

[adroc_thurston]

So, it does not look so bad, and might be accretive to AMD margins if AMD sells a lot of them.

they're not gonna sell a lot of them since it's a pilot product.
but yes it'll be margin accretive.

 

[branch_suggestion]

MI300A TOP500 submissions were interesting.
Still no 300X, hopefully we get some updated benchmarks at Computex.

 

[Kepler_L2]

MI300A TOP500 submissions were interesting.
Still no 300X, hopefully we get some updated benchmarks at Computex.

MLPerf numbers should be public soon.

 

[TESKATLIPOKA]

What percentage of mobile dGPUs are <= 4070 in your opinion?

My guess is about ~75% units, if not more.

So there is more than enough market potential for Strix Halo to compete for. No point to even think about higher tier of dGPU market before enough of this ( <= 4070) market is gained.

Strix Halo will likely have to compete against Blackwell and not Ada considering It's coming next year.
So the question is how good is that level of performance against Blackwell.
I don't expect miracles from Blackwell, but It still should perform better.

Then there is the question of price.

 

[del42sa]

I guess we will find soon and then we talk

 

[Joe NYC]

Strix Halo will likely have to compete against Blackwell and not Ada considering It's coming next year.
So the question is how good is that level of performance against Blackwell.
I don't expect miracles from Blackwell, but It still should perform better.

Then there is the question of price.

Strix Halo should be enough above 4060 that whatever performance increment 5060 offers, Strix Halo should still be in the same ballpark.

Since you mention 5060, the price is likely going to go up...

But from the rumors, NVidia is only launching high end initially. Who knows when the rest of the line-up is released.

 

[Mahboi]

StrixPoint Margin + NVidia 4060 margin + ~80 mm2 die area savings

On the negative side of the ledger would be cost differential between N4p (of 4060) and N3E of Strix Halo, if that's what it is.

So, it does not look so bad, and might be accretive to AMD margins if AMD sells a lot of them.

Very interesting, but your calculation focuses on area, and I'm curious about three elements:
- costs of packaging of Halo
- R&D, which I expect has been particularly extensive (LP core, advanced packaging, big APU in a laptop, etc)
- volume assessment

As always I'm a little skeptical of the AMBeliebers that scream that "chipletz are de wey" since they focus on the gains and ignore the extra costs. Which aren't big for cheaper stuff (fan-outs) but get higher pretty fast.
Halo is chiplet-based, Strix is not. To me, that sounds like an extra layer of problems to deal with in a supposed low power laptop. I suspect it may be their (first?) implementation of a silicon bridge, which is less money than MTL's Silicon interposer, but is still more than traces in PCB, as Halo is definitely the first part to leave AMD's Zen 2 era "just do the cheapest" approach. It's a testbed for Zen 6 and later, so no point in going cheap.

I expect both packaging costs to be relatively high, and R&D to have been heck of a lot.
Ofc volume is supposed to quash that, but I doubt that it'll sell as much as people believe. NV sticker is still worth a lot, Halo completely negates the value of putting a dGPU in a laptop. So I expect that Halo laptops will have to be more price competitive, I.E cheaper than full laptops with dGPUs. I also can easily expect NV to come out blasting about how "real laptops have dGPUs" and all the obvious jazz they'll do to impede Halo.

If Halo sells for cheaper perf to perf than Point + 4050/60/70, and its situation is "less area, less complexity to integrate, more packaging costs, has to sell for less of a price", I really really am doubtful about any margins or large scale sales.
Of course I'm willingly ignoring Halo's ability to be a Windows M1, a really low power stable chip, but that one is still out there, we don't know a thing about Z5 LP except that it's meant to exist. And we have to be doubtful about the market size of people who want a really low power full AMD laptop, but will be told "by the way, if you want it, you gotta pay for the midrange gaming sized GPU". Market of windows low power users? Huge. Market of those willing to pay another $100/200 for a gaming GPU? Tiny.

Halo will be a hard sell, it'll have cost a lot to develop, and its advantage is not that high in area, it's there for sure but we're talking 100m² if we're being ultra generous.
So I'm going to keep my original stance that adroc is full of copium on this and that Halo will be a net loss for AMD. Which is fine because it's an experimental, high risk product, that's meant to open the way to Z6 and beyond, and that's really all I want out of AMD, to take risks. But they're not making their money back on this one. I think we should expect Halo to open a lot of doors, but to be a bust financially speaking.
And I'm glad about it. AMD's been playing the poor man's game for too long, Halo is a real hard leap forward, so do it Lisa, lose money and come back with most of the hard lifting done for Zen 6.

 

[Mahboi]

Strix Halo will likely have to compete against Blackwell and not Ada considering It's coming next year.

So the question is how good is that level of performance against Blackwell.
I don't expect miracles from Blackwell, but It still should perform better.

Then there is the question of price.

Ada was already suffering at giving a decent low-midrange. Arguably outside of the "lucky" 3060, even Ampere wasn't doing toooo good. NV just likes selling big things a lot more than AMD, so I'm really expecting AMD to be pushing NV out of the competition in low to midrange, and Blackwell being really poor at it.

 

[branch_suggestion]

Very interesting, but your calculation focuses on area, and I'm curious about three elements:
- costs of packaging of Halo

~$10 over standard.

- R&D, which I expect has been particularly extensive (LP core, advanced packaging, big APU in a laptop, etc)

That R&D gets amortized over time as it applies to many products in the coming years.

- volume assessment

Proof of concept pipecleaner initially, but mainline after so it is okay.

As always I'm a little skeptical of the AMBeliebers that scream that "chipletz are de wey" since they focus on the gains and ignore the extra costs. Which aren't big for cheaper stuff (fan-outs) but get higher pretty fast.
Halo is chiplet-based, Strix is not. To me, that sounds like an extra layer of problems to deal with in a supposed low power laptop. I suspect it may be their (first?) implementation of a silicon bridge, which is less money than MTL's Silicon interposer, but is still more than traces in PCB, as Halo is definitely the first part to leave AMD's Zen 2 era "just do the cheapest" approach. It's a testbed for Zen 6 and later, so no point in going cheap.

-Halo and LNL are the same concept with varying tech underlying (LNL does save extra power with MoP). And LNL targets tablet level TDPs so power isn't an issue.

I expect both packaging costs to be relatively high, and R&D to have been heck of a lot.
Ofc volume is supposed to quash that, but I doubt that it'll sell as much as people believe. NV sticker is still worth a lot, Halo completely negates the value of putting a dGPU in a laptop. So I expect that Halo laptops will have to be more price competitive, I.E cheaper than full laptops with dGPUs. I also can easily expect NV to come out blasting about how "real laptops have dGPUs" and all the obvious jazz they'll do to impede Halo.

The parts Halo is built for is PC MBP clones, stuff that never has a dGPU.
That will need to change for the Z6 equivalent to take more TAM.

If Halo sells for cheaper perf to perf than Point + 4050/60/70, and its situation is "less area, less complexity to integrate, more packaging costs, has to sell for less of a price", I really really am doubtful about any margins or large scale sales.
Of course I'm willingly ignoring Halo's ability to be a Windows M1, a really low power stable chip, but that one is still out there, we don't know a thing about Z5 LP except that it's meant to exist. And we have to be doubtful about the market size of people who want a really low power full AMD laptop, but will be told "by the way, if you want it, you gotta pay for the midrange gaming sized GPU". Market of windows low power users? Huge. Market of those willing to pay another $100/200 for a gaming GPU? Tiny.

It does save OEMs the additional mobo/cooling costs of a typical 2 chip laptop, and is far easier to design and give ample room for batteries, cooling and IO.
While attractive, it is going to rely on some strong MDF investment beyond just the MS rebate to gain real traction vs the green sticker (NV MDF in laptops will be immense as they have nearly infinite cash to burn). RDNA3 being poison in the eyes of OEMs is not easy to ignore but maybe AMD marketing can get good all of a sudden.

Halo will be a hard sell, it'll have cost a lot to develop, and its advantage is not that high in area, it's there for sure but we're talking 100m² if we're being ultra generous.
So I'm going to keep my original stance that adroc is full of copium on this and that Halo will be a net loss for AMD. Which is fine because it's an experimental, high risk product, that's meant to open the way to Z6 and beyond, and that's really all I want out of AMD, to take risks. But they're not making their money back on this one. I think we should expect Halo to open a lot of doors, but to be a bust financially speaking.
And I'm glad about it. AMD's been playing the poor man's game for too long, Halo is a real hard leap forward, so do it Lisa, lose money and come back with most of the hard lifting done for Zen 6.

AMD has shown with HWK that they are now willing to lower the barrier of entry to solid APUs quite sharply.
Halo could be offered in rather cheap machines if they had the volume to sell, but as it looks volumes and wins will probably be quite low and focused on MBP clones, so strong margins but low TAM but does offer a gateway to future expansion.

 

[Mahboi]

~$10 over standard.

Ah, interesting, thank you. For a silicon bridge? Or are you assuming it'll be InFO?

That R&D gets amortized over time as it applies to many products in the coming years.
Proof of concept pipecleaner initially, but mainline after so it is okay.

Thanks, you just confirmed my point: Halo is a waste of money but will open doors, great.

-Halo and LNL are the same concept with varying tech underlying (LNL does save extra power with MoP). And LNL targets tablet level TDPs so power isn't an issue.

LNL doesn't get a 6700 xt strapped on its back.

The parts Halo is built for is PC MBP clones, stuff that never has a dGPU.
That will need to change for the Z6 equivalent to take more TAM.

MBP?
And yes to Z6 being the real result of Halo.

It does save OEMs the additional mobo/cooling costs of a typical 2 chip laptop, and is far easier to design and give ample room for batteries, cooling and IO.

So the biggest issue with laptop design is finding enough room to cram all the stuff?
Room for heatsinks is a bigger problem than having enough cooling for the chips?

While attractive, it is going to rely on some strong MDF investment beyond just the MS rebate to gain real traction vs the green sticker (NV MDF in laptops will be immense as they have nearly infinite cash to burn).

MDF? (there should be a damn website with all the acronyms...)

RDNA3 being poison in the eyes of OEMs is not easy to ignore but maybe AMD marketing can get good all of a sudden.

No.

AMD has shown with HWK thy they are now willing to lower the barrier of entry to solid APUs quite sharply.

Hawk but not Phoenix? What's changed in a year?

Halo could be offered in rather cheap machines if they had the volume to sell, but as it looks volumes and wins will probably be quite low and focused on MBP clones, so strong margins but low TAM but does offer a gateway to future expansion.

MBP again...

But yeah that's also my assessment.
Halo will be a net loss, but it will cover all bases: low power, gaming capable APU, and have Big Chip Energy. Then you can split its parts into low power stuff, big chip with small GPU or big GPU in Zen 6 and beyond.
It's a long term investment product, it's not going to be a high return product.

 

[ToTTenTranz]

they're not gonna sell a lot of them since it's a pilot product.

IMO it's a shame that Halo isn’t getting high volume for gaming laptops and larger handhelds.
Even 128bit + 8c + 10WGP cutdown versions of Halo would run circles around Strix Point thanks to 32MB LLC.

MBP?

Macbook Pro. Halo is a M3 Max competitor.

MDF? (

Marketing development funds. Something AMD traditionally struggles with.

 

[branch_suggestion]

Ah, interesting, thank you. For a silicon bridge? Or are you assuming it'll be InFO?

Yes, same as N31/32.

Thanks, you just confirmed my point: Halo is a waste of money but will open doors, great.

Then it isn't a waste of money.

LNL doesn't get a 6700 xt strapped on its back.

Yes but the packaging and overall design is sound, unlike the far more complex and expensive MTL.

MBP?

MacBook Pro

And yes to Z6 being the real result of Halo.

Medusa in its lovely forms will be pretty cool.

So the biggest issue with laptop design is finding enough room to cram all the stuff?
Room for heatsinks is a bigger problem than having enough cooling for the chips?

Well the room is okay, what people really want is lightness. Halo is the most perf for the least weight, basically.

MDF? (there should be a damn website with all the acronyms...)

Market Development Funds.

No.

We can dream.

Hawk but not Phoenix? What's changed in a year?

Hawk is in <$700 parts and came out in force, Phoenix was never that cheap and was a mess overall.

MBP again...

We all learn.

But yeah that's also my assessment.
Halo will be a net loss, but it will cover all bases: low power, gaming capable APU, and have Big Chip Energy. Then you can split its parts into low power stuff, big chip with small GPU or big GPU in Zen 6 and beyond.
It's a long term investment product, it's not going to be a high return product.

You can thank the semi-custom guys who have made these exact things in consoles for over a decade.

 

[branch_suggestion]

IMO it's a shame that Halo isn’t getting high volume for gaming laptops and larger handhelds.
Even 128bit + 8c + 10WGP cutdown versions of Halo would run circles around Strix Point thanks to 32MB LLC.

Say thanks to Nutella Man for his generosity.

Macbook Pro. Halo is a M3 Max competitor.

M4 Max really, both are mainly N3E and out around the same time.

 

[Mahboi]

Yes, same as N31/32.

I expect they'll go for something less power consuming and more advanced/expensive.
InFO made sense for N31, but for a low power laptop, I'm fully expecting a silicon bridge.

Then it isn't a waste of money.

Sure, it's just a manner of speaking. It's a "future investment oriented product". Lose now, gain later.
Again this is all I've been asking out of AMD/Radeon for years.

Yes but the packaging and overall design is sound, unlike the far more complex and expensive MTL.

Ponte Vecchio Redux Reloaded

Medusa in its lovely forms will be pretty cool.
Well the room is okay, what people really want is lightness. Halo is the most perf for the least weight, basically.

I see.

We can dream.

I can dream of Z6 being amazing, but I can't dream of AMD marketing getting their heads out of the toilet bowl.

Hawk is in <$700 parts and came out in force, Phoenix was never that cheap and was a mess overall.

Palpi voice:
Strix Point will become more powerful than either of us!

We all learn.

I'm just doing my classes on Anandtech.

You can thank the semi-custom guys who have made these exact things in consoles for over a decade.

Yet again the Invisible Hand of Cerny virtuously pushes us towards good engineering.
Gaben and Cerny better be given good champagne when they visit AMD.