Speculation: Ryzen 3000 series

[DrMrLordX]

What happens if you dont adapt and is 2 month late? Then people start to whine blame spread fear what not.

So . . . you want roadmaps with no dates on them now? Adaptability is nice and all, but consistency is far better. Whining and fear are irrelevant. Actual product on actual shelves is much, much more important.

If AMD wants to get out of this alive, they need Zen4 out on the streets within 3-6 months of Intel hitting with 7nm product on desktop or server (whichever comes first). AMD did well with Rome, which presumably is already plugging away in some select partner's server room (Amazon, Google, etc). Good on them for that. Rome was always planned for TSMC anyway. Pretty sure Matisse was a GF product, explaining part of the delay.

If AMD can't get Zen2 out before July 2019 and can't get Zen3 out before Q4 2020, then they are actually in some trouble. And AMD does not want to launch Zen2 and only give is a 9-month product window (ditto for Zen3), so now the earliest we can expect Zen3 is July 2020 and Zen4 July 2021. Intel could have something more sophisticated than Willow Cove out on their 7nm in 1H 2021. But again, those release dates for Zen2 and Zen3 are desktop . . . potentially, server could come up to 6 months earlier to select buyers.

I don't think Matisse was ever planned to be on GF's 7LP. After GF abandoned 7LP, wasn't it said by Papermaster or someone else that only one product was planned to be on that node and that they didn't expect roadmaps to change?

Yeah, pretty sure Matisse was a GF product originally.

I think it's to be expected that AMD will be a quarter off here and there in their "yearly cadence". On the plus side, we can hopefully expect a better launch without too many issues for early adopters.

Look, the cadence wasn't meant to be investor fud. It was supposed to be an actual representation of what they could, and would, do. We're all too accustomed to Intel not delivering on their schedules.

 

[krumme]

Considering the 14nm IO chip I disagree. GF dropping 14nm probably had more to do with AMD's choice not move to GF's 7nm in future and them saying they will only be ordering legacy chips and the new IO chips from them in the future. AMD didn't move to TSMC because GF dropped 7nm, GF dropped 7nm because AMD went to TSMC for 7nm.

14lpp was on a lease plan from Samsung so not even their own. 7nm was never remotely close. Look at the funding necessary. Gf was not near. See how much trouble Intel is having on more or less same density with their 10nm node.

 

[krumme]

Ph.D in electrical engineering. Not in chemistry or physics.

Naa you need to read longer in Wiki:

Yes she graduated with her PhD in electrical engineering from MIT in 1994. But her PhD thesis was titled Extreme-submicrometer silicon-on-insulator (SOI) MOSFETs.

Edit after her phd:
"
In June 1994, Su became a member of the technical staff at Texas Instruments,[12] working in the company's Semiconductor Process and Device Center (SPDC)[11] until February 1995.[12] That month[8], IBM hired Su as a research staff member specializing in device physics[20], and she was appointed vice president of IBM's semiconductor research and development center.[8] During her time at IBM,[6] Su played a "critical role"[7] in developing the "recipe"[2] to make copper connections work with semiconductor chips instead of aluminum, "solving the problem of preventing copper impurities from contaminating the devices during production".[7]Working with various IBM design teams on the details of the device, Su explained, "my specialty was not in copper, but I migrated to where the problems were".[6] "

 

[krumme]

So . . . you want roadmaps with no dates on them now? Adaptability is nice and all, but consistency is far better. Whining and fear are irrelevant. Actual product on actual shelves is much, much more important.

If AMD wants to get out of this alive, they need Zen4 out on the streets within 3-6 months of Intel hitting with 7nm product on desktop or server (whichever comes first). AMD did well with Rome, which presumably is already plugging away in some select partner's server room (Amazon, Google, etc). Good on them for that. Rome was always planned for TSMC anyway. Pretty sure Matisse was a GF product, explaining part of the delay.

If AMD can't get Zen2 out before July 2019 and can't get Zen3 out before Q4 2020, then they are actually in some trouble. And AMD does not want to launch Zen2 and only give is a 9-month product window (ditto for Zen3), so now the earliest we can expect Zen3 is July 2020 and Zen4 July 2021. Intel could have something more sophisticated than Willow Cove out on their 7nm in 1H 2021. But again, those release dates for Zen2 and Zen3 are desktop . . . potentially, server could come up to 6 months earlier to select buyers.



Yeah, pretty sure Matisse was a GF product originally.



Look, the cadence wasn't meant to be investor fud. It was supposed to be an actual representation of what they could, and would, do. We're all too accustomed to Intel not delivering on their schedules.

I dont want fairy tale road maps or 10nm 2c junk cpu with the only purpose of giving some bosses a bonus to say 10nm was released.

Try to read the management principles I give. Normally we think control is detailed planning. It just fits badly when complexity explodes. I say control is adaptability by effective feedback and lots of areas and ways to change direction. And product development that matches it.

The separation of io and core is a perfect example of product development that gives flexibility and adaptability.

Heck having a CEO that have a PhD in soi gives adaptability because she can go the the lead engineers have a talk, understand them and make a decision without going through the entire polit bureau. Fast change.

 

[DrMrLordX]

I dont want fairy tale road maps or 10nm 2c junk cpu with the only purpose of giving some bosses a bonus to say 10nm was released.

Maybe now you are seeing my point, then? Maybe? AMD just told us earlier this year that Q2 was the launch window for Matisse. They whiffed.

Execution is the foundation for everything. The roadmap is meant to be representative of what is actually happening in the labs, not just fatty meat for the investors. We need something we can believe. Adaptability or no, if all you do is adapt by delaying everything interminably, then . . . that's bad. Maybe "horrible" is too strong a word. But I can see these delays hurting AMD.

AMD told us 2019 for Zen2, 2020 for Zen3 and 2021 for Zen4. Intel has 3 core designs backlogged already that they can't launch thanks to the 10nm fiasco. I expect a 4th or 5th core step beyond Skylake to show up on 7nm in 2021, assuming they can execute on their "roadmap" (or whatever you call it). Which, granted, is quite a leap of faith. That core could annihilate Zen3. Whatever it is, assume it is going to be at least a step above Willow Cove. If Zen3 is all AMD has to show in 2021 when it comes out, and Zen4 starts looking more like a 2022 product, then bad things happen. Bad. Things. Matisse showing up in July 2019 is the first hairline crack in AMD's sterling record starting back in 2017 of compelling and timely releases of CPU product.

 

[krumme]

Maybe now you are seeing my point, then? Maybe? AMD just told us earlier this year that Q2 was the launch window for Matisse. They whiffed.

Execution is the foundation for everything. The roadmap is meant to be representative of what is actually happening in the labs, not just fatty meat for the investors. We need something we can believe. Adaptability or no, if all you do is adapt by delaying everything interminably, then . . . that's bad. Maybe "horrible" is too strong a word. But I can see these delays hurting AMD.

AMD told us 2019 for Zen2, 2020 for Zen3 and 2021 for Zen4. Intel has 3 core designs backlogged already that they can't launch thanks to the 10nm fiasco. I expect a 4th or 5th core step beyond Skylake to show up on 7nm in 2021, assuming they can execute on their "roadmap" (or whatever you call it). Which, granted, is quite a leap of faith. That core could annihilate Zen3. Whatever it is, assume it is going to be at least a step above Willow Cove. If Zen3 is all AMD has to show in 2021 when it comes out, and Zen4 starts looking more like a 2022 product, then bad things happen. Bad. Things. Matisse showing up in July 2019 is the first hairline crack in AMD's sterling record starting back in 2017 of compelling and timely releases of CPU product.

I dont put so much trust in roadmaps and look at ppt and try to derive if its q3 or q4. So I havnt used time on it like you. Especially on Intel. Frankly I stopped reading their bs. I hope they get back and will be glad if they do.

As said it's a matter of a single metal layer spin to eg get fmax a tad higher or take an error out. 7nm+ and zen3 is not delayed because of that.

It's no big deal unless it's a persistent pattern and the performance disappoint. What actually matters for zen2 is imo fmax and a wide variety of latency numbers.

 

[moinmoin]

Intel excactly shows why this cadence planning and idea of execution is something from hell.

Cadence is not necessarily from hell if it is safe-guarded. Advancements in chip design and process nodes can be safe guarded by doing iterative steps.

TSMC is highly iterative with their process nodes, there are so many comparably tiny steps it's easy to get lost in them (you can go from 7nm to 6nm, 7nm+ to 5nm etc.). AMD with Zen is also iterative, but on multiple levels (core design, uncore design, on chip interconnect design) at once.

Intel's huge issue is that they linked their iterative chip designs to their process nodes, but with 10nm they tried too many changes at once so they got stuck. Their progress on process nodes was defined by pure greed, doing iterative steps there like TSMC does would have been much more expensive, so their strong self-belief in being "process leader" (honestly unbelievable they still spout that) made them bite off more than they could chew. It will be very interesting how their 7nm will fare. Considering their process node cadence until 14nm Intel's 7nm already is late as well.

 

[krumme]

Cadence is not necessarily from hell if it is safe-guarded. Advancements in chip design and process nodes can be safe guarded by doing iterative steps.

TSMC is highly iterative with their process nodes, there are so many comparably tiny steps it's easy to get lost in them (you can go from 7nm to 6nm, 7nm+ to 5nm etc.). AMD with Zen is also iterative, but on multiple levels (core design, uncore design, on chip interconnect design) at once.

Intel's huge issue is that they linked their iterative chip designs to their process nodes, but with 10nm they tried too many changes at once so they got stuck. Their progress on process nodes was defined by pure greed, doing iterative steps there like TSMC does would have been much more expensive, so their strong self-belief in being "process leader" (honestly unbelievable they still spout that) made them bite off more than they could chew. It will be very interesting how their 7nm will fare. Considering their process node cadence until 14nm Intel's 7nm already is late as well.

I agree. And establishing a firm cadence is best option if you can brake down the parts and steps. Be it either in design or process whatever. But something cant just be devided enough or the consequences makes far less sense. Eg zen as an basic arch needs a certain level of parts to be released while some minor parts that didn't get into zen 2 can just be pushed to zen3. Same goes for process development I guess.

It's also important that what you establish with a firm cadence is also often shorter loops that enables you to adjust plan. So in that sense it enables adaptability.

 

[Ajay]

Naa you need to read longer in Wiki:

Yes she graduated with her PhD in electrical engineering from MIT in 1994. But her PhD thesis was titled Extreme-submicrometer silicon-on-insulator (SOI) MOSFETs.

Thanks! Didn’t use wiki - my bad.

 

[PotatoWithEarsOnSide]

I'm not sure how a July launch is inconsistent with AMD having said "mid-year"...

 

[coercitiv]

AMD just told us earlier this year that Q2 was the launch window for Matisse. They whiffed.

They said "mid-year" though.

 

[DrMrLordX]

like you. Especially on Intel. Frankly I stopped reading their bs. I hope they get back and will be glad if they do.

If Intel doesn't make good on their 7nm promises then even worse things happen. Sorry if I'm overly-optimistic about Intel's future. I have to give them a little credit for being able to launch 7nm in a timely fashion.

7nm+ and zen3 is not delayed because of that.

No, but it's clear that AMD wants more than 9-10 months market time for Matisse, which is what they would get if they launched Zen3 in Apr-May 2020. By default, Zen3 (for desktop) can't launch before July 2020.

They said "mid-year" though.

Most of the reports on Matisse I saw earlier in the year were either "Q2 or Q3" or just "Q2". AMD left themselves some wiggle room (I guess).

 

[B-Riz]

I'm guessing the Zen2 delay is to make sure motherboards are available and are rock solid.

No cluster fudge like the Zen launch.

If chip yields are as good as Zen was after a certain point, they may delay pushing Zen3 if Intel has nothing really beating them strongly.

 

[tomatosummit]

If Intel doesn't make good on their 7nm promises then even worse things happen. Sorry if I'm overly-optimistic about Intel's future. I have to give them a little credit for being able to launch 7nm in a timely fashion.
No, but it's clear that AMD wants more than 9-10 months market time for Matisse, which is what they would get if they launched Zen3 in Apr-May 2020. By default, Zen3 (for desktop) can't launch before July 2020.
Most of the reports on Matisse I saw earlier in the year were either "Q2 or Q3" or just "Q2". AMD left themselves some wiggle room (I guess).

I don't know if you made a typo there, but intel certainly hasn't "launched 7nm in a timely fashion". If the rocket lake leaks are to be believed then 10nm will never reach performance parity for desktop and high performance laptops and high performance servers. The roadmap shows 4core <35w parts into late 2020 and 10nm igpu chiplets paired with 14nm core chiplets. For servers as icelake is launching alongside cooper lake you can assume it's a low volume part where cooper lake is the high availability part to go along side. So when 10nm does launch it's not going to be that great.

I don't think we've heard anything about Intel's 7nm apart from that one slide have we where it essentially says it has more realistic design goals but again we won't know about it's specific performance until we get close to launch. It'll be dense that's for sure but I don't know if we can discern anything else.

Likewise the zen 2 delays are rumoured to be from the 570 chip set and there have been suspiciously few leaks from that until just the last couple of weeks, normally asmedia leaks the details well in advance (and we know they're not making this one, which is great as x370/470 are a bit naff). Ryzen2000 was also 13.5 months after ryzen1000 and a July launch for ryzen3000 isn't that much of a delay either.

 

[B-Riz]

I don't know if you made a typo there, but intel certainly hasn't "launched 7nm in a timely fashion". If the rocket lake leaks are to be believed then 10nm will never reach performance parity for desktop and high performance laptops and high performance servers. The roadmap shows 4core <35w parts into late 2020 and 10nm igpu chiplets paired with 14nm core chiplets. For servers as icelake is launching alongside cooper lake you can assume it's a low volume part where cooper lake is the high availability part to go along side. So when 10nm does launch it's not going to be that great.

I don't think we've heard anything about Intel's 7nm apart from that one slide have we where it essentially says it has more realistic design goals but again we won't know about it's specific performance until we get close to launch. It'll be dense that's for sure but I don't know if we can discern anything else.

Likewise the zen 2 delays are rumoured to be from the 570 chip set and there have been suspiciously few leaks from that until just the last couple of weeks, normally asmedia leaks the details well in advance (and we know they're not making this one, which is great as x370/470 are a bit naff). Ryzen2000 was also 13.5 months after ryzen1000 and a July launch for ryzen3000 isn't that much of a delay either.

My thought is, if this is a near perfectly executed launch with no major bugs and high availability, it will be a near legendary moment in computing history. (Beating Intel on process, etc.)

 

[B-Riz]

Lol, remember Sandy Bridge P67 recall? So glad I waited to buy the 2500K and board, lool.

 

[Veradun]

Most of the reports on Matisse I saw earlier in the year were either "Q2 or Q3" or just "Q2". AMD left themselves some wiggle room (I guess).

Reports are meaningless. AMD said 2019 first, then mid-year, and now Q3.

 

[Justinbaileyman]

Any chance the 3000 series 16 core part will be beating a Threadripper 2950x part since there will be an ipc improvement?
Kinda on the fence weather or not to just upgrade now to 2950x since there is a deal going on or should I just wait it out for 2 more months.
Personally thinking the 2950x will remain top dog considering because of quad channel memory, would I be wrong in thinking this??

 

[beginner99]

Any chance the 3000 series 16 core part will be beating a Threadripper 2950x part since there will be an ipc improvement?
Kinda on the fence weather or not to just upgrade now to 2950x since there is a deal going on or should I just wait it out for 2 more months.
Personally thinking the 2950x will remain top dog considering because of quad channel memory, would I be wrong in thinking this??

For memory bound workloads, 2950x will be faster due to quad channel RAM as you say but not all loads are memory bound. anything compute bound should be faster on ryzen 3000 16-core, if that part ever gets released. Rumors are only 12-core part at launch.

 

[Fanatical Meat]

Another X570 “leak” and another chipset fan

https://wccftech.com/msi-x570-gaming-motherboards-amd-ryzen-3000-cpus-leak/

 

[fleshconsumed]

Well, I hope ASUS releases a fanless X570 board. Otherwise I'll be sticking with X470.

 

[Ajay]

My thought is, if this is a near perfectly executed launch with no major bugs and high availability, it will be a near legendary moment in computing history. (Beating Intel on process, etc.)

Well there are some indications that Zen2 got a B0 re-spin this year, so I’m guessing AMD really wanted top notch release. I think clocks will be lower than some of the extreme suggestions in this thread, but that won’t matter with an additional throughput of 10-15% on a clock normalized basis.

 

[tomatosummit]

Well, I hope ASUS releases a fanless X570 board. Otherwise I'll be sticking with X470.

The fans do suck but we'll see. I suspect you can mod them a bit and combine them with bigger heatsinks or case airflow/custom fans. It's be nice if the mobos came with standard mouning for a 40mm fan or something. I use a noctua 40mm in one of my machines and it's all but silent.

Well there are some indications that Zen2 got a B0 re-spin this year, so I’m guessing AMD really wanted top notch release. I think clocks will be lower than some of the extreme suggestions in this thread, but that won’t matter with an additional throughput of 10-15% on a clock normalized basis.

A respin might not only be clock speeds, didn't the zepplin b0 fix the segfault that ryzen1000 had?
I agree about the clock speed optimism though. 'If' a part can hit a single core 5ghz it doesn't mean getting 16 or even 8 cores to 5ghz is easy. It's very likely to be hot and very power hungry at best. I don't think there's been a specific leak denying that the process can hit 5ghz, it's just that it might not be easy for everyone.

 

[Justinbaileyman]

For memory bound workloads, 2950x will be faster due to quad channel RAM as you say but not all loads are memory bound. anything compute bound should be faster on ryzen 3000 16-core, if that part ever gets released. Rumors are only 12-core part at launch.

Ok, thanks for the reply.. I just went ahead and bought the 2950x for now.
If Ryzen 3000 ends up having a 16c part when released and ends up being faster then the 2950x then I will probably snag one up as well.
I just couldn't pass up the combo deal on newegg, only $739 + free shipping for the 2950x and a 21.5 inch monitor.
I have been wanting this bad boy since it launch and now I finally got it.
I am very excited!!
I think this is a crazy good deal considering the 2950x alone sells for $819.99 on sale right now, well plus tax that is.
So basically you get a huge discount on the 2950x and its kinda like you get the monitor for free..

 

[Fanatical Meat]

The fans do suck but we'll see. I suspect you can mod them a bit and combine them with bigger heatsinks or case airflow/custom fans. It's be nice if the mobos came with standard mouning for a 40mm fan or something. I use a noctua 40mm in one of my machines and it's all but silent.


A respin might not only be clock speeds, didn't the zepplin b0 fix the segfault that ryzen1000 had?
I agree about the clock speed optimism though. 'If' a part can hit a single core 5ghz it doesn't mean getting 16 or even 8 cores to 5ghz is easy. It's very likely to be hot and very power hungry at best. I don't think there's been a specific leak denying that the process can hit 5ghz, it's just that it might not be easy for everyone.

Yeah, I’m thinking maybe the chipset fan isn’t a big deal. Fans seem to work pretty good today, also could be a fun nerd project changing or modifying the fan.

**Ive never owned a PC with a chipset fan so my opinion may change if the fan sucks**