News AMD Financial Analyst Day 2020: A Lot of Details to Digest

[moinmoin]

The anandtech article have some amd slide that say's:
The Next Five Years
OUR TECNOLOGY INVESTMENTS

However didn't see more than two years being disclosed did i miss something?
Or they are adding the previous three years?

Overall those FADs are irregular events intended for giving analysts a vague outlook into the company's future. AMD's previous FADs were in 2017, 2015, 2012...

 

[Mopetar]

AMD is starting to get into a much better position and it will be nice to see them execute as effectively with Radeon as they have with Zen.

CDNA is probably the biggest announcement and I’m excited to see what the future holds as far as that goes in terms of server products.

I hope they don’t rest on their laurels though. AMD is just as capable of stagnation as Intel.

 

[inf64]

I don't think many people realize how behind intel is right now. They are easily 3 or even 4 years behind AMD in CPU core business and probably much more than that in GPU segment. It will be a tough couple of years for intel before they have anything that resembles a competitive CPU.

 

[Atari2600]

That's one of AMD's potential problems. AMD needed to face off against Granite Rapids with Zen5, not Zen4.

Whats that opinion based on?

 

[DrMrLordX]

Whats that opinion based on?

Based on their need to

a). maintain cadence
b). win market share

We have no idea what Granite Rapids' performance will be like, but AMD need to err on the side of complete annihilation. Right now, Rome devastates Cascade Lake-SP and Cascade Lake-AP in nearly every possible server room metric, yet their market share hasn't moved by all that much. AMD can't really deal with all the other factors causing people to stubbornly side with Intel. Their only choice is to excel.

When in doubt, crush the opposition.

Right now, it looks like Milan will face Cooper Lake and (in limited runs) Ice Lake-SP. AMD should want Zen4/Genoa to face off against Sapphire Rapids and Zen5 to face off against Granite Rapids.

 

[moinmoin]

Based on their need to

a). maintain cadence
b). win market share

We have no idea what Granite Rapids' performance will be like, but AMD need to err on the side of complete annihilation. Right now, Rome devastates Cascade Lake-SP and Cascade Lake-AP in nearly every possible server room metric, yet their market share hasn't moved by all that much. AMD can't really deal with all the other factors causing people to stubbornly side with Intel. Their only choice is to excel.

When in doubt, crush the opposition.

Right now, it looks like Milan will face Cooper Lake and (in limited runs) Ice Lake-SP. AMD should want Zen4/Genoa to face off against Sapphire Rapids and Zen5 to face off against Granite Rapids.

Personally I think execution is of far bigger importance than launch timing. Since the launch of Zen AMD has shown to make steady improvements while not stagnating in any area they touch upon. Core, uncore, package and platform all have seen significant changes within the 3 years all while keeping compatibility. Them keeping going this way will make a much more impactful impression than just following the schedule of the competitor (which is an interesting story of its own considering how big the repackaging of existing tech has been part of "new" launches in the last couple years).

 

[DrMrLordX]

Personally I think execution is of far bigger importance than launch timing.

Execution is launch timing. AMD has the plans on the table to continue improvements, yes, but their current status was won through a fairly-constant drumbeat of upgrades to their core product. The only thing they didn't do was release an EPYC update to coincide with Pinnacle Ridge/Zen+, which is somewhat understandable. AMD "got away with it" when Intel couldn't do any better than move from Skylake-SP to Cascade Lake-SP. AMD can't rely on Intel screwing up that badly again. It's safe to say that Milan will enter the market largely-unopposed since Intel seems to be struggling with Ice Lake-SP (and Cooper Lake is an unremarkable product with horrid TDPs). They can't just assume that Sapphire or Granite will suffer the same fate as Ice Lake-SP. Anything less than a yearly cadence from here on out is risky business.

 

[moinmoin]

Execution is launch timing. AMD has the plans on the table to continue improvements, yes, but their current status was won through a fairly-constant drumbeat of upgrades to their core product. The only thing they didn't do was release an EPYC update to coincide with Pinnacle Ridge/Zen+, which is somewhat understandable. AMD "got away with it" when Intel couldn't do any better than move from Skylake-SP to Cascade Lake-SP. AMD can't rely on Intel screwing up that badly again. It's safe to say that Milan will enter the market largely-unopposed since Intel seems to be struggling with Ice Lake-SP (and Cooper Lake is an unremarkable product with horrid TDPs). They can't just assume that Sapphire or Granite will suffer the same fate as Ice Lake-SP. Anything less than a yearly cadence from here on out is risky business.

As Intel has repeatedly shown launch timing is more about putting lipstick on a pig, managing launches with new model names and categories, new segmentation, new prices and a couple mhz here and there. That's from where Intel first has to move away again and show that they can constantly introduce features again. AMD on the other hand has been wasteful with the cards they could have kept in the hand: No holding back features for segmentation and future pseudo-updates, no keeping core count lower for later increases. So by all accounts they already try to "crush the opposition". With this precedence it seems odd to me to think that AMD may intend to stop doing so in the future.

You may well be right that the launch timing will put AMD at disadvantage. But I'd expect them to try harder with the actual products then.

 

[DrMrLordX]

But I'd expect them to try harder with the actual products then.

That's also an unknown. With the recent announcements that Milan may have . . . unexpected features like 3D stacking, and with the previous revelation that Rome wound up with enhancements not initially intended for Zen2 (like TAGE), Zen4/Genoa may wind up being an absolute beast. In the end, the product has to perform. I do not mean to state that AMD intends NOT to "crush the opposition"; there's simply the fear that they might not be able to do so, which in turn may sink their chances of gaining permanent market share via recognition (which is something they failed to do the last time they were competitive).

 

[moinmoin]

I wouldn't "worry" so much. From where I stand AMD is already doing fine as is, everything else is a bonus for them. And if AMD keeps up the good work they did the last couple years they won't go down again anytime soon.

And if Intel comes back sooner with actual competitive products the better for us consumers! Currently it just looks more likely that they can only serve very specific markets with actual new products.

 

[awesomedeluxe]

Good: Zen 4 is officially 5nm. Not sure this was confirmed before; I think there was an outside chance it was would be 6nm.

Bad: Zen 4 in 2022? IIRC Zen 4 has the major changes in power management. This is what AMD needs to win in mobile. Tiger Lake / Rocket Lake likely take the crowns in the laptop space thanks to lower idle draw.

2022 will be a great year for notebooks if Intel and AMD stick to their roadmaps. I was really hoping AMD could bring real advancements to the mobile space sooner, though.

 

[yuri69]

Not saying this is easy to do, but with the CPU people there for some time, these techniques appear to begin showing up in the GPU circuitry.

This is so mind-blogging.

* 2006 ATi was acquired by AMD
* 2007 AMD presents GPU-CPU design synergy with M-Space fluff
* 2011 Bulldozer development gets presented by C.Maier - detailed doom and gloom that was brought by the GPU-CPU design methodology merge

Now AMD presents RDNA benefiting from GPU-CPU design synergy... After like 14 years they finally realized they needed to optimize their GPU designs using CPU know-how?

 

[amrnuke]

But to do so you have to create bigger (more expensive) die and clock it lower. It's higher cost for AMD, which wants to improve their profit margin.

Is that not what Nvidia did with the 1080 Ti -> 2080 Ti transition as well? Throw 50% more transistors and 5-10% more W, and clock it lower, and magically 30-35% performance gains, solid efficiency, etc. Also MSRP'd for $300 more. But plenty of people still bought (and still are buying) it.

 

[Hitman928]

Good: Zen 4 is officially 5nm. Not sure this was confirmed before; I think there was an outside chance it was would be 6nm.

Bad: Zen 4 in 2022? IIRC Zen 4 has the major changes in power management. This is what AMD needs to win in mobile. Tiger Lake / Rocket Lake likely take the crowns in the laptop space thanks to lower idle draw.

2022 will be a great year for notebooks if Intel and AMD stick to their roadmaps. I was really hoping AMD could bring real advancements to the mobile space sooner, though.

If this ends up being true, I don't think AMD has to worry about getting bested in the mobile space:

Ryzen 4000 mobile laptops could hit “18 hours” of battery life, AMD says

Calling AMD’s upcoming 7nm Ryzen 4000 laptop CPUs a “watershed” moment for AMD, the company hinted that at least one model could hit “18 hours” of runtime.

www.pcworld.com

 

[moinmoin]

This is so mind-blogging.

* 2006 ATi was acquired by AMD
* 2007 AMD presents GPU-CPU design synergy with M-Space fluff
* 2011 Bulldozer development gets presented by C.Maier - detailed doom and gloom that was brought by the GPU-CPU design methodology merge

Now AMD presents RDNA benefiting from GPU-CPU design synergy... After like 14 years they finally realized they needed to optimize their GPU designs using CPU know-how?

We have to keep in mind that AMD cut much of its graphics workforce during the down time. Until 2012 they had the TeraScale line which was originally developed at ATi before the merger. The first graphics uarch developed within AMD was GCN which was used until very recently. RDNA appears to be AMD's first graphics uarch developed in a stable environment, now with plenty financial backing to boot.

For a long time knowledge went the other way: AMD still had to learn synthesizing and modularizing CPU designs (something already applied to GPUs, this allowed GCN to live as long as it did), how to port them between nodes etc. That all was knowledge originally from ATi, which got a further boost in the semi custom business efforts. These all were preconditions to the successful execution AMD now showcases with Zen and tries to apply to RDNA as well.

 

[amrnuke]

Execution is launch timing. AMD has the plans on the table to continue improvements, yes, but their current status was won through a fairly-constant drumbeat of upgrades to their core product. The only thing they didn't do was release an EPYC update to coincide with Pinnacle Ridge/Zen+, which is somewhat understandable. AMD "got away with it" when Intel couldn't do any better than move from Skylake-SP to Cascade Lake-SP. AMD can't rely on Intel screwing up that badly again. It's safe to say that Milan will enter the market largely-unopposed since Intel seems to be struggling with Ice Lake-SP (and Cooper Lake is an unremarkable product with horrid TDPs). They can't just assume that Sapphire or Granite will suffer the same fate as Ice Lake-SP. Anything less than a yearly cadence from here on out is risky business.

I don't want to get into any type of ticky-tacky argument, but this seems like your opinion about the market more than anything that's actually being demanded.

I don't think AMD (or AMD's interpretation of market demands) agree with you all that much. "Mark Papermaster: We’re on a 12-18 month cadence, and we believe that is sustainable. It’s what the industry demands from us."

Zen: Mar 2017. Zen+: Apr 2018 (13 months). Zen2: July 2019 (15 months).

If we assume a perpetual 13-15 month cadence, on the low end of what Papermaster said is their real cadence, we would expect Zen3 in August-October. Adding another 13-15 month cadence and that's September 2021 to January 2022 for Zen4. January 2022 is entirely in line with history and with what AMD have said. However, given that he said 12-18 months, it could be the case that we get Zen3 in very late 2020 and Zen4 doesn't arrive until early-mid 2022.

Kind of moot, since Intel's cadence is... well, there isn't one on the desktop side. At this point, I'll believe in Intel's desktop 10nm, or any other future node, when I actually see it on desktop. And it seems market analysts agree with that sentiment.

I'm not saying that AMD slowing down would be smart, but I also think as long as they're making steady progress over time, the rest is irrelevant.

 

[DrMrLordX]

I don't want to get into any type of ticky-tacky argument, but this seems like your opinion about the market more than anything that's actually being demanded.

If you say so. Many of us watched AMD go spiraling into the trashcan years ago. Dealing with an entrenched foe like Intel is no easy task, and there are other competitors waiting to nip at AMD's heels as well. They really can't afford to slow down unless each generational step proves to be a bigger advancement than initially expected. AMD can't really afford to flame out again.

I'm not saying that AMD slowing down would be smart, but I also think as long as they're making steady progress over time, the rest is irrelevant.

That has failed them in the past. K8->Phenom->Phenom II represented steady progress, but where did it get them?

 

[amrnuke]

If you say so. Many of us watched AMD go spiraling into the trashcan years ago. Dealing with an entrenched foe like Intel is no easy task, and there are other competitors waiting to nip at AMD's heels as well. They really can't afford to slow down unless each generational step proves to be a bigger advancement than initially expected. AMD can't really afford to flame out again.

As in my post, keeping the current 13-15 month cadence, Zen4 still could be January 2022, even without slowing down.

If you're implying that they need to accelerate the cadence to release a new chip every year in July (you haven't said that, but it sounds like you think that would be the goal), that's going to push them into making Zen3.5 types of chips, which would be antithetical to continued rapid advancement in many ways, including by wasting resources on meeting a deadline instead of devoting all the resources possible to the true next generation.

 

[Thunder 57]

That has failed them in the past. K8->Phenom->Phenom II represented steady progress, but where did it get them?

That really isn't a fair comparison as K8 was 2003-2007. Phenom in 2007 was broken and really not any good until Phenom II in early 2009. AMD has executed well and on schedule this time around.

 

[Atari2600]

That really isn't a fair comparison as K8 was 2003-2007. Phenom in 2007 was broken and really not any good until Phenom II in early 2009. AMD has executed well and on schedule this time around.

Also, that was around the time that AMD were hitting the financial buffers in terms of supporting their own foundries and offloading to Global Foundries.

 

[Gideon]

Also, that was around the time that AMD were hitting the financial buffers in terms of supporting their own foundries and offloading to Global Foundries.

And in the middle of ATI acquisition (completed in 2006).

To be fair in 2005 they did go Dual Core, which was somewhat of an innovation (and a considerable amount of work on the software/os side).

But yeah, they were stuck on the exact same architecture from 2003-2007 on 130nm, 90nm and 65nm processes. At some pointt Hector Ruiz in his infinite wisdom even postponed Phenom and future node development.

 

[NostaSeronx]

Actually, 2004 they killed K9 which would be Opteron2 and dead for good and K10 which would be redeveloped as Bulldozer after 2006.

K9 is the big core(Bulldozer spot) and K10 was the small core(Bobcat spot). With 2006+ re-doing this w/ K10 renamed to Bulldozer as the big core spot and Bobcat(in-house Geode) as the small core spot.

Hector Ruiz killed AMD's competitive roadmap and Dirk Meyer had to focus server when BoD wanted tablets. Intel has core+atom and we will respond with atom+atom!

K9 is pretty close to Zen, even using the earliest form of the tools which Zen was crafted with. However, hand-custom Bulldozer won somehow.

//
K9 65nm Project: Project was cancelled after the Macros were completed.
Implementation of K9 Trace Cache blocks (not tapeout out)
R.I.P. They had a Core 2 competitor and they killed it.

 

[DrMrLordX]

If you're implying that they need to accelerate the cadence to release a new chip every year in July (you haven't said that, but it sounds like you think that would be the goal), that's going to push them into making Zen3.5 types of chips, which would be antithetical to continued rapid advancement in many ways, including by wasting resources on meeting a deadline instead of devoting all the resources possible to the true next generation.

That's what previous roadmaps had them on, and now the roadmaps are changed. That's the concern. Not necessarily July either; EPYC launches have thus far been August affairs (with consumer chips available in March, April, and July). AMD clearly isn't interested in "half step" releases in the server room or else there would have been an EPYC+.

Look, not trying to make a mountain out of a molehill here, but AMD gets a lot of praise for what they've done right, so I think it's only fair to note when something looks a bit off. Previously, it looked like Milan was going to be widely-available in August of this year while Genoa would be widely-available in August/September of 2021. Now it's not quite the same. Something has changed, and it's fair for people to wonder what.

That really isn't a fair comparison as K8 was 2003-2007. Phenom in 2007 was broken and really not any good until Phenom II in early 2009. AMD has executed well and on schedule this time around.

It is unfair, but at the same time it isn't. If you went back to 2004/2005 and said to me that AMD would surely mess up Phenom as badly as they did, I would not have believed you. Their execution from the original Athlon all the way up to the first x2 chips ranged from good to great. Then 65nm showed up and everything went to hell. At least this time around, TSMC is in control of the nodes (more or less) so AMD can't get that part wrong. And, you know, no Hector.

 

[Gideon]

Look, not trying to make a mountain out of a molehill here, but AMD gets a lot of praise for what they've done right, so I think it's only fair to note when something looks a bit off. Previously, it looked like Milan was going to be widely-available in August of this year while Genoa would be widely-available in August/September of 2021. Now it's not quite the same. Something has changed, and it's fair for people to wonder what.

While I understand your concern and agree to an extent, AMD has always mentioned 12-18 month cadence for zen, so it's not like they changed anything.

People expecting what you state, have always been unrealistic, just denying what AMD has been explicitly saying for years.

As far as Desktop chips go:

• Zen was released on 2017-03-02.
• Zen+ was released 2018-04-18 (almost a 2 month shift for essentially a half-node shrink).
• Zen2 was 2019-07-07.
• Zen3 looks to be September at best (if everything is perfect), Q4 is much more likely.

Anandtech's article even mentions this. Regarding EPYC:

Zen 4 based Genoa has already been announced as the CPU to power the El Capitan supercomputer, and in this roadmap AMD has put it as coming out by 2022. We asked AMD for clarification, and they stated that in this sort of graph, we should interpret it as the full stack of Genoa should be formally launched by the end of 2022

And Zen 3 Ryzen:

We asked AMD to clarify, and were told that we should interpret this as that the range of Zen 3 consumer products, such as desktop CPUs, HEDT CPUs, mobile APUs, and consumer APUs, should all be available by the end of 2021

Genoa was never going to be released before 2021Q4 and that might still be the case. The might just be covering their backside with the "2022" number in case an additional respin is required (~3month delay). You don't promise such things that far in advance with no leeway (or AMD used to during the Phenom/Bulldozer times and constantly deliver 1-2Quarters late).

It's certainly isn't ideal if it instead comes 2022Q1 but won't change the overall picture much.

 

[eek2121]

While I understand your concern and agree to an extent, AMD has always mentioned 12-18 month cadence for zen, so it's not like they changed anything.

People expecting what you state, have always been unrealistic, just denying what AMD has been explicitly saying for years.

As far as Desktop chips go:

• Zen was released on 2017-03-02.
• Zen+ was released 2018-04-18 (almost a 2 month shift for essentially a half-node shrink).
• Zen2 was 2019-07-07.
• Zen3 looks to be September at best (if everything is perfect), Q4 is much more likely.

Anandtech's article even mentions this. Regarding EPYC:

And Zen 3 Ryzen:


Genoa was never going to be released before 2021Q4 and that might still be the case. The might just be covering their backside with the "2022" number in case an additional respin is required (~3month delay). You don't promise such things that far in advance with no leeway (or AMD used to during the Phenom/Bulldozer times and constantly deliver 1-2Quarters late).

It's certainly isn't ideal if it instead comes 2022Q1 but won't change the overall picture much.

I feel the need to point this out: Genoa is a server chip, not the code name for the Zen 4 architecture. We will see desktop chips before the server chips drop. As to launch timing? Nobody except AMD knows, of course. If I had to speculate, I would say Zen 3 launches in July and Zen 4 will launch later rather than sooner. However, I expect Vermeer to launch in 2021 along with some variants of Genoa. The cheaper chips for both Vermeer and Genoa won’t land until 2022.