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

 

[Thunder 57]

That's literally the point, every ATi/AMD attempt to cultivate brand loyalty with unbeatable value offerings ended with NV pricecuts. And so the status quo was maintained.

According to the graph above, even when Nvidia has released hot garbage like the FX or Fermi AMD never got 50% market share. What should they do, stop making GPU's? Obviously they can't now but I bet it's come across their mind. At least in the past they could maintain ~40%.

 

[GodisanAtheist]

You claimed that AMD cards don't sell because they don't have a halo product. That chart directly contradicts that notion.

- AMD really doesn't do Halo's though. Last one where they really went for broke was the FURY X and it got curb stomped by the 980Ti.

AMD just seems deeply insecure about big GPU dies and I find it hard to blame them.

- 2900XT got it's teeth kicked in by 8800GTX.
- Then a bunch of smaller n-1 dies until Hawaii and the 290X which got dumped on because of its garbage cooler and "AMD Fine Wine" which really only saw it reach its full potential vs GK102 near the end of the gen.
- Fury X got wiped by 980Ti, then it was back out in the wilderness with rebadged thin margin parts like Vega and Polaris and RDNA 1.
- RDNA 2 came out swinging at the same time NV mis-stepped with the Samsung node, but still lagged in features and RT performance which was becoming more and more relevant at that point and mining just completely torched the whole scene top to bottom in terms of gaming and goodwill.
- RDNA 3 was another flub, clearly intended to go Halo but was another expensive n-1 part.
- RDNA 4 looks like AMD is back out in the wilderness.

His argument is AMD isn't NV and never seems to try and pull out all the stops for a top end chip (Halo). AMD gets gun-shy around the 500mm2 mark with it's consumer GPUs. NV just squeezed out a 750mm2 monster just to make sure they're already unassailable lead stays that way.

 

[Timorous]

These didn't win. Not a real halo.

So the last real Halo card that has a proper advantage was arguably the 295X2 since there was no sane priced dual GPU 700 series.

Or if you want single GPU then the last halo was probably the 9700pro / 9800XT. AMD have completed at the high end in other gens but those are the last cards I remember that just absolutely demolished the competition at the high end.

 

[Heartbreaker]

Is he? AMD's two best selling generations in the past 10 years were Polaris/Vega and RDNA1, and those generations were the least competitive in the high end. If anything, it looks like there's a negative correlation between AMD competing in the high end and them gaining market-share.

View attachment 117631

article

It's hard to get the whole picture what was driving shifts.

Polaris/Vega were released in 2017 was released in a crypto boom, and the graph looks more like AMD's market share for Polaris/Vega was primarily increasing because of Crypto. It's rising right until Crypto bubble pops, then it collapses hard after, then bounces back a bit. You can see the same effect for the second crypto boom. It's a period when people will buy all the GPUs, and AMD rises in market share while crypto buying is going on, and dips hard right after, both times.

For RDNA 1, I'm not seeing much movement at all. That post crypto dip was already bouncing back before RDNA 1, and It dips on the release quarter then levels off the same as the quarter before.

 

[GTracing]

- AMD really doesn't do Halo's though. Last one where they really went for broke was the FURY X and it got curb stomped by the 980Ti.

The 6950 XT was closer to being a halo card than the fury x was.

 

[Timorous]

Do you know what a loss leader is? A loss leader is a product that loses money. AMD isn't losing money on every GPU sold.

If it cuts into CPU wafer allocation then arguably it does given the margin difference between a $500 9800X3D and a $500 9070 for instance.

 

[Heartbreaker]

Also I looked at the graph to see what AMD/ATI was selling the last time they actually had higher market share than NVidia:

It was the X800 series, that actual matched or exceeded NVidias best:

ATI's New Radeon X850 and X800 Lines: A Smorgasbord of GPUs

www.anandtech.com

Be sure to read the top comment below the Review. The more things change...

 

[Hail The Brain Slug]

If it cuts into CPU wafer allocation then arguably it does given the margin difference between a $500 9800X3D and a $500 9070 for instance.

Making less profit is not losing money. Keeping a diversified product portfolio is important to a healthy business that can withstand market changes.

 

[GodisanAtheist]

Also I looked at the graph to see what AMD/ATI was selling the last time they actually had higher market share than NVidia:

It was the X800 series, that actual matched or exceeded NVidias best:

ATI's New Radeon X850 and X800 Lines: A Smorgasbord of GPUs

www.anandtech.com

- And also, ironically, had a feature deficit vs NV (no DX9.0c).

 

[Thunder 57]

- And also, ironically, had a feature deficit vs NV (no DX9.0c).

Yet previouslly the 9700 Pro beat the FX while using a single slot and it still didn't pass 50%. In DX9 is was a bloodbath.

I guess people saw the DX9 numbers and reacted by buying ATi which explains the gains.

 

[Heartbreaker]

Yet previouslly the 9700 Pro beat the FX while using a single slot and it still didn't pass 50%. In DX9 is was a bloodbath.

Look where they were starting from though. They had only 26% market share when the first released 9000 series and that took them up to 45%, which is a MASSIVE jump in share.

9700Pro was my last Radeon...

 

[gaav87]

Yet previouslly the 9700 Pro beat the FX while using a single slot and it still didn't pass 50%. In DX9 is was a bloodbath.

I guess people saw the DX9 numbers and reacted by buying ATi which explains the gains.

Imagine, a 96CU 24gb 384bit rdna4 for 1k$ ...
Would be, a bloodbath like this but for whole 4080super dupers ti+.

 

[yuri69]

The 6950 XT was closer to being a halo card than the fury x was.

Yes, Fury X was not a Halo SKU since it was severely memory capacity and b/w (!) bound. It was more like a HBM1 prototype sold to public.

 

[soresu]

HEVC 4:2:2 on rx7000 and newer or am i reading this wrong ?​

https://twitter.com/x/status/1892303189602394430

Agility SDK 1.716.0-preview: New D3D12 Video Encode Features - DirectX Developer Blog

Today DirectX 12 provides APIs to support GPU video encode acceleration for several applications, as detailed in D3D12 Video Encoding – Windows drivers | Microsoft Learn previous blog posts such as Announcing new DirectX 12 feature – Video Encoding! – DirectX Developer Blog. In this blog post...

devblogs.microsoft.com

No, only that it supports new features in the SDK from DX12 and video encoding.

Nothing in the AMD specific PR mentions 4:2:2 at all.

 

[soresu]

Yes, Fury X was not a Halo SKU since it was severely memory capacity and b/w (!) bound. It was more like a HBM1 prototype sold to public.

More like HBM1 itself was the prototype, and Fury the launch platform.

 

[Hitman928]

According to the graph above, even when Nvidia has released hot garbage like the FX or Fermi AMD never got 50% market share. What should they do, stop making GPU's? Obviously they can't now but I bet it's come across their mind. At least in the past they could maintain ~40%.

AMD/ATI's main problem isn't halo/no halo or pricing, it's consistency. NV has had much more consistent execution from generation to generation. Every time AMD starts to gain a little momentum, they fall flat on their face. That's why they were successful on the CPU side, it's not like AMD jumped to majority market share right away, they gave customers consistent product improvements while Intel failed to execute. Had AMD had a Bulldozer moment with Zen 4 and Intel released a winning lineup, AMD would have given up a ton of share it worked years to claw away from Intel. When you are coming at a clear market leader, you need to be offering consistently good options for generations before you are able to build a truly solid base. Having a halo product that is a clear market leader is basically a short cut to build your base (assuming no supply issues and such) but it won't last long if you turn around and have another fail generation right after.

 

[GodisanAtheist]

AMD/ATI's main problem isn't halo/no halo or pricing, it's consistency. NV has had much more consistent execution from generation to generation. Every time AMD starts to gain a little momentum, they fall flat on their face. That's why they were successful on the CPU side, it's not like AMD jumped to majority market share right away, they gave customers consistent product improvements while Intel failed to execute. Had AMD had a Bulldozer moment with Zen 4 and Intel released a winning lineup, AMD would have given up a ton of share it worked years to claw away from Intel. When you are coming at a clear market leader, you need to be offering consistently good options for generations before you are able to build a truly solid base. Having a halo product that is a clear market leader is basically a short cut to build your base (assuming no supply issues and such) but it won't last long if you turn around and have another fail generation right after.

- Yep. NV is an execution machine. In fact Blackwell is probably their first real stumble since Fermi, and before Fermi their last real stumble was NV35/5900. They just don't really mess up all that often, and they always ALWAYS offer a top to bottom stack. Every. Single. Time.

NV owners more or less always have an NV card to upgrade to in the next gen. AMD owners often find themselves in a situation where they also only have an NV card to upgrade to in a subsequent gen. Do that for 20+ years and the market is inevitably going to end up badly lopsided.

 

[maddie]

AMD/ATI's main problem isn't halo/no halo or pricing, it's consistency. NV has had much more consistent execution from generation to generation. Every time AMD starts to gain a little momentum, they fall flat on their face. That's why they were successful on the CPU side, it's not like AMD jumped to majority market share right away, they gave customers consistent product improvements while Intel failed to execute. Had AMD had a Bulldozer moment with Zen 4 and Intel released a winning lineup, AMD would have given up a ton of share it worked years to claw away from Intel. When you are coming at a clear market leader, you need to be offering consistently good options for generations before you are able to build a truly solid base. Having a halo product that is a clear market leader is basically a short cut to build your base (assuming no supply issues and such) but it won't last long if you turn around and have another fail generation right after.

Agreed. This argument shows how rudderless AMD graphics division has become. They communicated a rational path to competitiveness and at the first opportunity to make a few short term $s, they threw it all away.

I guess the thinking is, "let's really start in the next generation".

 

[eek2121]

It's $749.

Why would it need to be anything but $749.

They don't have to. 749 it is.

I don’t normally attack posts like this, but uh, it isn’t $749. 😉

 

[jpiniero]

I don’t normally attack posts like this, but uh, it isn’t $749. 😉

$899 then?

 

[Thunder 57]

AMD/ATI's main problem isn't halo/no halo or pricing, it's consistency. NV has had much more consistent execution from generation to generation. Every time AMD starts to gain a little momentum, they fall flat on their face. That's why they were successful on the CPU side, it's not like AMD jumped to majority market share right away, they gave customers consistent product improvements while Intel failed to execute. Had AMD had a Bulldozer moment with Zen 4 and Intel released a winning lineup, AMD would have given up a ton of share it worked years to claw away from Intel. When you are coming at a clear market leader, you need to be offering consistently good options for generations before you are able to build a truly solid base. Having a halo product that is a clear market leader is basically a short cut to build your base (assuming no supply issues and such) but it won't last long if you turn around and have another fail generation right after.

That's the most reasonable take on all of this I've read yet.

 

[IEC]

$899 then?

"Look at me, I'm the scalper now." -Retailer

Low price of $1599, with this $500 PSU. What a deal!

 

[Thunder 57]

"Look at me, I'm the scalper now." -Retailer

Low price of $1599, with this $500 PSU. What a deal!

Everyone's found a way to get in on the action. Nvidia, just raise prices and hand out scraps. AIB's, we'll set "fair market" prices based on what people will pay. Retailers, "Want that fancy new video card? We'll give it to you if you a bunch of other crap we would never be able to sell in a bundle". What a deal!

The problem is that there is still enough demand out there so they can get away with this. Until that changes, don't expect these practices to go away. My guess is that there is really tiny supply or there are too many people out there with no self control. Probably a mix of both.

 

[marees]

I think people who expect AMD to 'compete' in DGPU market are in for a shock

If AMD wanted to really compete then they would have released RDNA 4, 6 months ago, at whatever price they could

 

[Dezii90]

I think people who expect AMD to 'compete' in DGPU market are in for a shock

If AMD wanted to really compete then they would have released RDNA 4, 6 months ago, at whatever price they could

What stops Jensen from launching 6 months ago as well?