Re-litigating Radeon market performance *spawn

[tunafish]

AMD wanted to do a major laptop push with both RDNA2 and RDNA3, and, well, crickets. Geforce is not just plain better in a limited power envelope, but also has a such massive brand power that laptop makers don't want to tie their expensive products to the brand that the market feels is inferior.

I think that outside APUs, laptop GPU penetration will significantly lag desktop GPU market penetration. Which is probably why Strix Halo, it lets AMD build a single product that's both the best CPU on the market and the best GPU AMD can put in the power budget.

 

[Bondrewd]

AMD wanted to do a major laptop push with both RDNA2 and RDNA3, and, well, crickets.

RDNA2 did very well in laptops.
3, though. Oops!
Ouch. But happens.

Which is probably why Strix Halo, it lets AMD build a single product that's both the best CPU on the market and the best GPU AMD can put in the power budget.

Nah it's just that Apple proven big APUs are workable/viable in laptops and now certain OEMs get major stiffies for any IHV emulating them.
Also why LNL exists!

 

[Dictator]

Why does AMD historically have such a record with cancelled or botched high end units? I swear I read similar stories every few years about the high-end GPU underperforming or being cancelled.

Kind of starting with 7000 series

 

[Qesa]

Why does AMD historically have such a record with cancelled or botched high end units? I swear I read similar stories every few years about the high-end GPU underperforming or being cancelled.

Kind of starting with 7000 series

Do they actually underperform internal simulations or is it just the rumour mill? "AMD will leapfrog Nvidia" gets more clicks than "the status quo will continue", then when it inevitably doesn't happen the rumour mongers backtrack with "oh we weren't making stuff up, no, it missed their targets".

The same happened with Fiji, Vega and RDNA1 as well, so it's not a pattern starting with RDNA3.

It's at least somewhat plausible for RDNA4's high end to be cancelled if it relied on CoWoS packaging that suddenly became scarce with the post-chatGPT AI hype.

 

[Dictator]

Do they actually underperform internal simulations or is it just the rumour mill? "AMD will leapfrog Nvidia" gets more clicks than "the status quo will continue", then when it inevitably doesn't happen the rumour mongers backtrack with "oh we weren't making stuff up, no, it missed their targets".

The same happened with Fiji, Vega and RDNA1 as well, so it's not a pattern starting with RDNA3.

It's at least somewhat plausible for RDNA4's high end to be cancelled if it relied on CoWoS packaging that suddenly became scarce with the post-chatGPT AI hype.

Yes from my perspective it is just the rumour mill. I do not think there is evidence to the contrary. So like you said, it could just be the clicks and the hype cycle which contracts after reality sets in and all clicks have been gotten.

 

[trinibwoy]

It's at least somewhat plausible for RDNA4's high end to be cancelled if it relied on CoWoS packaging that suddenly became scarce with the post-chatGPT AI hype.

Wouldn’t RDNA 4 high end have significantly higher volumes than AI accelerators? CoWoS packaging should have been a concern even without the AI hype. What packaging tech is RDNA 3 using?

I agree a lot of AMD disappointment is a result of pre-launch hype which is mostly not AMD’s fault. People love an underdog comeback story.

 

[arandomguy]

What packaging tech is RDNA 3 using?

Info_os

 

[tunafish]

Do they actually underperform internal simulations or is it just the rumour mill? "AMD will leapfrog Nvidia" gets more clicks than "the status quo will continue", then when it inevitably doesn't happen the rumour mongers backtrack with "oh we weren't making stuff up, no, it missed their targets".

There exist AMD slides to AIB vendors from well before release that claim that RDNA3 is a 3GHz+ architecture. Then when they had actual physical samples and knew what the chips were capable of, that messaging went away.

 

[Qesa]

There exist AMD slides to AIB vendors from well before release that claim that RDNA3 is a 3GHz+ architecture. Then when they had actual physical samples and knew what the chips were capable of, that messaging went away.

It didn't go away once they had samples, "architected to exceed 3 GHz" was still in their launch slide deck. AMD marketing fluffing numbers isn't the same as missing targets.

EDIT: See here: https://www.tomshardware.com/news/a...-for-gpus#section-amd-rdna-3-closing-thoughts

Embargoed 14 November 2022, 3 mentions of 3 GHz (plus a 30% higher clocks than RDNA2 at iso power)

 

[Bondrewd]

Why does AMD historically have such a record with cancelled or botched high end units

It's Vega and?
R600 too I guess.

 

[LordEC911]

I thought Ibiza got confirmed in one of those AMD AAR interviews.

It is good to keep in mind that a potential future design could be trying to implement many new technologies and/or features that aren't 100% owned/dependent on AMD. Hitting correctly on edge case timelines is extremely hard and expensive.

 

[Bondrewd]

I thought Ibiza got confirmed in one of those AMD AAR interviews.

And the 32nm GPU lot was nuked from orbit ca 2009 but that's external (TSM) factors, not AMD.

 

[trinibwoy]

There exist AMD slides to AIB vendors from well before release that claim that RDNA3 is a 3GHz+ architecture. Then when they had actual physical samples and knew what the chips were capable of, that messaging went away.

It seems bizarre that they would share those targets externally before they confirmed it was happening. Don’t these guys usually run years/months of internal simulation and testing before saying anything?

 

[Bondrewd]

Don’t these guys usually run years/months of internal simulation and testing before saying anything?

bingo.
That's where RDNA3 went wrong.
Missed the pre-Si viz by quite some margin.

 

[mboeller]

It's Vega and?
R600 too I guess.

Fiji? only 4GB of HBM memory

 

[raytracingfan]

Why does AMD historically have such a record with cancelled or botched high end units? I swear I read similar stories every few years about the high-end GPU underperforming or being cancelled.

Kind of starting with 7000 series

The HD 7000 series or the RX 7000 series?

 

[Bondrewd]

Fiji? only 4GB of HBM memory

It was mid but nothing anywhere close to high-profile failures of R600 or Vega.

 

[Seanspeed]

I always wonder how the 7000/GCN1 series would have looked if Nvidia didn't split up Kepler like they did and just released the full lineup in 2012. I think Nvidia made a gift to AMD here cuz it allowed the 7970 and whatnot to look competitive against Nvidia, all while the 680 was really only the upper midrange for Kepler. Similarly, if GK100 was released in 2012, it would have meant that AMD wouldn't have a response for that til late 2013, making it look much later to the party. Instead, it made Hawaii look quite good cuz Nvidia delayed releasing GK100 til 2013 themselves.

I actually think Nvidia learned from that, cuz they never did it again and it exactly had the effect of making AMD look less competitive.

 

[trinibwoy]

bingo.
That's where RDNA3 went wrong.
Missed the pre-Si viz by quite some margin.

Ok but why make promises before seeing what the silicon can do? Do AIBs really need that much advance notice on stuff like clock speeds?

 

[Bondrewd]

Do AIBs really need that much advance notice on stuff like clock speeds?

Not an AIB thing, more laptop-related really.
They have to give perf guidance in quite some advance to court OEM design wins.

Instead, it made Hawaii look quite good cuz Nvidia delayed releasing GK100 til 2013 themselves.

The funny bit of history is that Hawaii itself looks bad because it was GDDRified late in the development.

and it exactly had the effect of making AMD look less competitive.

that's all on AMD.