Speculation: Ryzen 3000 series

[PotatoWithEarsOnSide]

Too many commas.

 

[ozzy702]

Intel is hot overpriced garbage and should be ignored/avoided until 2020/2021 when they will hopefully figure out how to run a modern day CPU business, fire their whole marketing team, streamline their product offerings, and lower their prices

I hope that with Zen2 AMD knocks it out of the park so this actually happens and so we no longer have to choose betweewn moar cores and performance/price vs outright top tier performance. I hate making compromises and would love to spend less and get more.

Intel needs a massive restructuring to refocus/innovate and the best part about all this is the consumer wins big time! I'd love to see AMD do the same on the GPU side but that defect is even larger than CPU so it will probably be more like 2021+ before AMD brings us solid competition on that front.

 

[ub4ty]

Too many commas.

What's that you said? Oh and Momma Su is bout dat lyfe.
Going hard in the paint on Intel

 

[Adawy]

Martin Brundle didn't recognize Momma Su, shame on him.

/s

 

[IEC]

If these "leaked" clocks are anywhere near reality, I'll be buying several CPUs.

But I'm not holding my breath. Every time the hype goes stratospheric people set themselves up to be deflated when reality brings them back down to earth. Vega being the worst offender there.

Don't get me wrong, I'm hoping that the clockspeeds are true as it means more performance at every price point for everyone - but I really dislike it when hyped-up rumors become the expectation.

 

[ub4ty]

If these "leaked" clocks are anywhere near reality, I'll be buying several CPUs.

But I'm not holding my breath. Every time the hype goes stratospheric people set themselves up to be deflated when reality brings them back down to earth. Vega being the worst offender there.

Don't get me wrong, I'm hoping that the clockspeeds are true as it means more performance at every price point for everyone - but I really dislike it when hyped-up rumors become the expectation.

Agreed. I was super salty after buying a Vega into the launch hype. Flipped it some time after I realized it was a let down for my purposes.. The only consolation was that I charged some miner $850 for it and paid retail $399.

The only expectation I have is that power utilization drops and clocks get a nice bump @ next week's announcement. I'm on first gen. So, this is where I make my move as I alway have : when gate size is cut in half. So, there's nothing for me to be let down about.

Bonus points for :
- PCIE 4.0
- New Mobo x570 that has PCIE 4.0 and maybe some more I/O
- Consumer paired GPU w/ PCIE 4.0 w/ ability CCIX over PCIE 4.0 straight to CPU
- If AM4 socket has the I/O chiplet paradigm Rome has
- If IPC sees a pickup
- If core count for AM4 sees a bump
And of course if Momma Su allows threadripper to be a cut down Rome .. (Life will be complete if this happens). Of course, the price has to be right as well or I'm just going to ride my current CPUs a little longer until I feel the price is right.

So, my expectations are going to be satisfied and, if any of the bonus bullets are hit, I'm going to be very excited. That's how you go into these things IMO. Expectations set at mid level based on factual data that's already been released w/ some sensible nice haves in bonus category. Absolutely no fine grain details are in my head on clocks/power util/price.

 

[jpiniero]

And of course if Momma Su allows threadripper to be a cut down Rome .. (Life will be complete if this happens). Of course, the price has to be right as well or I'm just going to ride my current CPUs a little longer until I feel the price is right.

I wouldn't rule out an full blown unlocked Rome model actually but it wouldn't be cheap.

 

[Markfw]

I wouldn't rule out an full blown unlocked Rome model actually but it wouldn't be cheap.

It could not be a full blown Rome, as that is 8 channel memory, and I don't think any socket TR4 are being announced anyway at CES, just AM4.

 

[jpiniero]

It could not be a full blown Rome, as that is 8 channel memory

It would be on SP3.

 

[Markfw]

It would be on SP3.

Why do that ? Thats a regular Rome socket, but that also is not supposed to be announced I don't think, just AM4.

At this point, too much speculation, we should just wait until CES.

 

[jpiniero]

Why do that ? Thats a regular Rome socket, but that also is not supposed to be announced I don't think, just AM4.

Make money obviously. I think there's a market for it although I won't say AMD would do it and it sure wouldn't be cheap.

 

[IEC]

Make money obviously. I think there's a market for it although I won't say AMD would do it and it sure wouldn't be cheap.

I think it'd be so niche no one is going to bother making boards for it.

There might be a enthusiast-class ASRock server board for 7nm SP3 Epyc chips if demand is high enough (are you listening, ASRock?), but we're talking full-blown Epyc, not Threadripper. For product segmentation reasons I just don't see them selling chips that could be sold at Epyc margins for the consumer.

 

[scannall]

Intel is hot overpriced garbage and should be ignored/avoided until 2020/2021 when they will hopefully figure out how to run a modern day CPU business, fire their whole marketing team, streamline their product offerings, and lower their prices

A big part of their problem is they have lost a lot of senior engineers over the last 5-7 years. Both from poaching, and 'early retirement' to get the expensive guys off the payrolls. Saving a few nickels for a short term stock bump. And risking big dollars if they can't get the talent they need now that the fire is hot.

 

[ub4ty]

I think it'd be so niche no one is going to bother making boards for it.

There might be a enthusiast-class ASRock server board for 7nm SP3 Epyc chips if demand is high enough (are you listening, ASRock?), but we're talking full-blown Epyc, not Threadripper. For product segmentation reasons I just don't see them selling chips that could be sold at Epyc margins for the consumer.

Asrock !!!! Do it

 

[darkswordsman17]

A big part of their problem is they have lost a lot of senior engineers over the last 5-7 years. Both from poaching, and 'early retirement' to get the expensive guys off the payrolls. Saving a few nickels for a short term stock bump. And risking big dollars if they can't get the talent they need now that the fire is hot.

I'm not following that logic at all? So they lost talent when things were lean, why would they not be able to hire it when things are going well?

Not only that, but considering where things are versus where they were, it seems like they might be better off without them? Speaking from an engineering and successful implementation of product, they don't seem to have suffered for it.

 

[DrMrLordX]

I see nothing regarding 'tile of tile' : https://en.wikichip.org/wiki/intel/microarchitectures/cascade_lake

That was my terminology, not theirs. But if you look at actual Cascade Lake-AP which is two dice "glued" together, they've basically made a big, slow mesh of two tiles. And each tile is a faster mesh of 24 smaller tiles. That was just my comparison. Obviously were UPI faster/lower latency, it would work more like a tile/subtile system instead of what they have now.

Well, everyone has interposers. It seems to me just an attempt to create a proprietary tech.

In all fairness to Intel, the silicon bridges used for EMIB should be a lot cheaper to implement than full interposers, especially in configurations with a small number of dice. What we do not know are the performance differences between a full interposer and an EMIB silicon bridge. Thanks to Kaby-G we should be able to see how PCIe-over-EMIB works vs. traditional implementations. I just haven't seen anyone do the testing yet.

 

[amd6502]

If the ES leak is legit, it seems these are very early crude ES of a Ryzen 3000, with maybe some stepping(s) to go.

Videocardz is claiming leaks on 12nm APU Ryzen 3000 series, also feels 8c/16t ryzen will be 7nm AM4 flagship cpu in this gen.

https://twitter.com/x/status/1081246710976917505

See comments in link above.

Given the 12c ES leak, it seems AMD is experimenting with bringing >8c to AM4. However, perhaps they do this with a mixed product line, using two Pinnacle or Summit dies glued together; it would make better sense to me than a complex IO hub for consumer DT. If so, it would be a good complement to offer some high thread high MT with more moderate frequencies alongside the high frequency high sparse thread performance 8c/16t 7nm flagship CPUs.

Another possibility is what Adoredtv at some point suggested; they use a gated off sawed off IO hub from epyc to fit two chiplets and hub onto AM4 footprint.

I think Zen2 is still in early ES testing meaning it'll be a while.

So this may be the order of launches for 3rd gen:

12nm mobile APU (ryzen 3 through 7) up to 4c/8t
12nm desktop APU (ryzen 3 through 5) up to 4c/8t
7nm Ryzen CPU
7nm Epyc (2nd gen, with 3rd gen core)
7nm Threatripper (assuming it is based on die salvaged Epyc IO hub).

Not sure where Navi sits in this order.

 

[NostaSeronx]

imho...

Matisse2 1H2019 => GlobalFoundries' 12HPC/HD (9-track/4Fin/78CPP High Performance/High Density FinFET) => Low-number 3000s
^-- 7LP-port up
Matisse 2H2019 => TSMC's 7HPC (7.5-track/3Fin/57CPP High Performance FinFET) => High-number 3000s (Lower defect density from less area means less salvaged needed)
^-- Paper launches with X570 at Computex.

Also, imho TSMC's N7 is a short-lived node. N7+ is much better than N7. Which leads me to think Matisse Prime is getting respinned before launch to 7HPC+.

 

[beginner99]

gaming performance like it is today is mostly dictated by your graphics card

Wrong. It'd dictated by your screen resolution and refresh rate. 1080p >=120hz gaming heavily relies on CPU and if you arent' that keen on ultra settings a 580/1060 is good enough. In contrast here you take a big hit with a Ryzen CPU. High refresh rate gaming = Intel and more budget towards CPU than GPU. end of discussion.

of course if you game at 4k, then the CPU becomes much less important (in most games but not all, Startcraft...) and you should go with the best bang for the buck and put the money in the GPU

 

[beginner99]

Thanks to Kaby-G we should be able to see how PCIe-over-EMIB works vs. traditional implementations. I just haven't seen anyone do the testing yet.

AFAIK only the HBM is connected to GPU via EMIB (instead of interposer like on vega). CPU-GPU pcie is just an on-package organic pcie link.

 

[Adawy]

imho...

Matisse2 1H2019 => GlobalFoundries' 12HPC/HD (9-track/4Fin/78CPP High Performance/High Density FinFET) => Low-number 3000s
^-- 7LP-port up
Matisse 2H2019 => TSMC's 7HPC (7.5-track/3Fin/57CPP High Performance FinFET) => High-number 3000s (Lower defect density from less area means less salvaged needed)
^-- Paper launches with X570 at Computex.

Also, imho TSMC's N7 is a short-lived node. N7+ is much better than N7. Which leads me to think Matisse Prime is getting respinned before launch to 7HPC+.

Are you saying that Zen 3 will offer a decent leap over Zen 2 and not just the usual ~5% from optimization?

 

[NostaSeronx]

Are you saying that Zen 3 will offer a decent leap over Zen 2 and not just the usual ~5% from optimization?

Probably, N7 is pretty bad. So, Zen2 will be the Agena and Zen3 will be the Deneb. A better comparison is same node generation: Zen2 is to Kaveri in 15W and Zen3 is to Carrizo/Bristol in 15W.
https://images.anandtech.com/doci/9319/Slide 12 - 15W Optimized.png

That but replace it with 105W and reverse it. Lower performance boost at low TDP, and higher performance boost at high TDP.

If you want farther projection. Zen4 will be on Samsung's 3GAAE and Zen5 will be on 3GAAP.

 

[moinmoin]

Are you saying that Zen 3 will offer a decent leap over Zen 2 and not just the usual ~5% from optimization?

Papermaster was quoted last November that 7nm+ is mostly used for increasing efficiency:
Looking ahead, a 7-nm-plus node using extreme ultraviolet lithography (EUV) will “primarily leverage efficiency with some modest device performance opportunities,” he said in the interview.

 

[AtenRa]

Wrong. It'd dictated by your screen resolution and refresh rate. 1080p >=120hz gaming heavily relies on CPU and if you arent' that keen on ultra settings a 580/1060 is good enough. In contrast here you take a big hit with a Ryzen CPU. High refresh rate gaming = Intel and more budget towards CPU than GPU. end of discussion.

of course if you game at 4k, then the CPU becomes much less important (in most games but not all, Startcraft...) and you should go with the best bang for the buck and put the money in the GPU

Actually its dictated by the budget available. If you have an infinite budget to spend, you will need both a fast CPU and a fast GPU. The faster the GPU you are going to use, the faster the CPU you are going to need in order to get 100% of the GPU performance.

 

[PotatoWithEarsOnSide]

With an infinite budget you likely also buy the highest resolution monitor available, again shifting you to the GPU.
The bottleneck at the top end is always shifting IMO.