Speculation: Ryzen 3000 series

[ozzy702]

So what do you think a 16-Core 5.1GHz Ryzen 3000 is going to consume, ballpark?

Full load? Probably 250W+ if we're talking 16 core 32 thread, all core 5.1ghz and that's being generous. I don't think we'll see 5ghz or higher parts at all and that these "leaks" are wishful fanboi dreams. I do think we'll see some all core turbos in the 4.7-4.8ghz range but even those will be thirsty 200W chips. Pushing clocks sucks juice, even on cutting edge processes.

 

[Shivansps]

The R3-3300 is way too much for $100, it makes sence to a point were im thinking they are probably leaving rebranded RR as Athlons at sub <$100... but why they are skipping 4C/8T and 6C/6T on $100 and going directly to 6C/12T? its too much.

If the source of all this is AMD, they are probably testing reactions.

 

[Zucker2k]

Don't take the leaks as absolute, but they will be ballpark correct from a product dev standpoint.

This is what I was going after. To be clear, I'm a skeptic of that 5.1GHz beast of a chip, hence my sarcasm laced question, which didn't come out too clearly, I might add.

 

[ArchAngel777]

I'll give AMD 6 months to come up with something. I am itching to upgrade my Haswell box and almost went with a 9700K, but naw... I'll let AMD see what it can come up with.

 

[Atari2600]

So what do you think a 16-Core 5.1GHz Ryzen 3000 is going to consume, ballpark?

A Ryzen 3000 that pushes all 16 cores at > 5 GHz won't exist.

The socket and motherboard simply won't be designed with that kind of power consumption in mind.

A Ryzen 3000 that can turbo a core or two to that kinda of speed may* exist - but its power requirements would be within socket limits (say 125W "official").

*Although I'm skeptical.



edit: For reference, a 2950 *will just about* get to ~4.0 GHz on all 16 cores at 180W. So, going to 7nm... that might become, say, 125W (guess at around 30% power improvement at iso-clock). But you'd at least double that power going from 4.0 GHz to 5.0GHz (based on power increase being approx cubic to clock) so north of 250W I'd guess.

Much guessing, little data - take with a truck load of salt.

 

[jpiniero]

Yeah, that list that Adored gave seemed to come straight from that fake 4chan post.

I'm more of the opinion that they will stick to Threadripper for 16 and above. They do have plenty of options there.

 

[inf64]

Kyle Bennet basically confirmed this (AdoredTV's leaks) in his post on [H] forum. There still might be some errors in his "leaked" roadmap but overall he hit the nail on the head, just as he did with chiplet leak and RTX leak. Next year will be very exciting when it comes to CPUs and GPUs, I think even more exciting than say 2017 (Zen1 launch).

 

[Maxima1]

This is what I was going after. To be clear, I'm a skeptic of that 5.1GHz beast of a chip, hence my sarcasm laced question, which didn't come out too clearly, I might add.

It's like the other Ryzens. The max boost isn't referring to all the cores.

 

[exquisitechar]

https://twitter.com/AdoredTV/status/1070345089300475904

Adored on the leak...

It's not fake, I know where it came from.

Note that some of the information could be complete nonsense, like the 5GHz stuff at the end. I can imagine a scenario where somebody added two more lines to the table in order to fool somebody.

The important thing is, my information came before the Reddit stuff and it's very similar but still with a lot of changes. It looks very likely that the Reddit leak is just very old, maybe 6 months or older.

I have since been told of a 64c/128T Threadripper, only about one of them though not the 6 of them on that reddit list. No word on clock speeds at all, but I've also been told that the X900X branding is dead.

Well, he acknowledges that the clocks could be made up.

 

[Topweasel]

Makes sense to kill off the naming on Thread Ripper. If its really going to be 4-16c on Consumer Ryzen and god knows the model numbers between that and the 64c Threadripper. Keeping TR within the that 100 point range in naming makes no sense. Probably will see something Like X or WX and a much wider range like Ryzen has. So for example 16C TR3 X3500 or 64c TR3 WX3900.

 

[HurleyBird]

edit: For reference, a 2950 *will just about* get to ~4.0 GHz on all 16 cores at 180W. So, going to 7nm... that might become, say, 125W (guess at around 30% power improvement at iso-clock). But you'd at least double that power going from 4.0 GHz to 5.0GHz (based on power increase being approx cubic to clock) so north of 250W I'd guess.

It's cubic to voltage. Frequency by itself is linear.

 

[rbk123]

Seems far too good to be true; makes me wonder if it was faked by someone in the other camp as there will be lots of pissing/moaning if the actual announcement ends up being much lower on the core/clocks.

 

[Tuna-Fish]

At the Next Horizon event, AMD published this slide:

This means you should expect both substantially higher clocks, and more cores (density + power), but you don't get to have those at the same time. Ryze 3xxx with twice the cores in the same TDP as current Ryzens will run at roughly the same clocks. AMD has spelled that out for us in the above slide.

However, when using less cores, they should be able to clock substantially higher. TSMC 7nm is a substantially better process than what they used for previous Ryzens, and also the first proper refinement of a CPU arch can usually pull more clocks, because now they have had enough silicon to find where the pain points are and time to smooth them out. I fully expect the best single-core boost to be >=5GHz, and would be disappointed if it was less. However, expecting that to be the base clock is going against the numbers AMD published in the above slide. Why would they not market more if they had more?

I agree with many of the posters here on the point that it doesn't really make sense for AMD to offer >8 cores for the mainstream in the AM4 socket. However, I expect they will, because in existing roadmaps they have different product code names for the APUs and the CPUs. If they went with a single design at up to 8 cores, that would likely be the APU, and there would be no two separate product code names.

 

[Zapetu]

I would see that exactly the other way around. The 6-core parts with 2 defective ones would go into consoles. I'm still of the opinion that a 4c8t Zen2@3ghz is more than enough for consoles. But I see your point. A top 12c ryzen could use 2 chiplets that have 1 or 2 defective cores but are high leakge eg. clock high so not best option for server or consoles. But if that much binning is needed, 7nm as tsmc must yield very poorly.

EDIT: In fact the next consoles most likely will be 14nm still. I could see the gpu being 7nm but even that would be an issue price wise.

AdoredTV talks about yields of the 8C chiplet in his recent video. Anyone can verify those numbers by themselves using e.g. this Die Yield Calculator. I used the same parameters as he did (300mm wafer, 7.1mm x 10.1mm = 71.71mm² chiplets, 0.4#/cm² defect density, all other parameters default) and got the following chart:

As we can see, the max number of (working) dies is 817 and 617 are good dies (with 8 working cores) and 200 are defective dies (with 6/4/2/0 working cores). Yields go up over time, so why wouldn't Sony/MS use fully working 8C chiplets (with medium clock speeds and medium/low leakage) if they have decided to go the chiplet route. It wouldn't make much sense to disable those working cores just because of power savings. They can however control how game developers are allowed to use them and keep the power usage at desired levels.

On that wafer there are 200 defective dies and if that AdoredTV leak turns out to be true, most of these partially working ones would be used for best selling (mainstream) AMD Ryzen 3000 SKUs. Fully defective dies and partial dies (#60 in that wafer, yellow ones) would be used as dummy (i.e. filler) dies.

The most important reason why it makes sense to add consoles to the mix is that then AMD would have a much larger pool of wafers of those 8C chiplets to bin and cherry pick the best ones for EPYC/TR and good ones for higher end AM4. Sure, Sony/MS could have gone with a monolithic design using only one 4C Zen2 CCX but why not benefit from this chiplet revolution by using fully working 8C chiplets?

Obviously none of this leak is still confirmed but adding almost every CPU product to the mix would allow AMD to get all the manufaturing benefits of this chiplet approach. As I see it, without consoles onboard this would miss a lot of the potential it has. Monolithic (Ryzen 3000) design would not really have a large enough pool to really cherry pick those best clocking chips in volume as AdoredTV has also many times mentioned in his videos. The more I think about this, the more sense it makes from AMDs point of view.

For a while I had my doubts about this chiplets (without IMCs) for the destop market approach (EPYC Rome made much more sense because how Naples is laid out) but now it all seems to come together almost too nicely. Sure, memory access latencies would be a little higher in some cases than using a monolithic design but plausible 32MiB of L3 cache per each chiplet will help a lot. Sure, cross CCX latencies would be lower for a monolithic chip but then there wouldn't be a large pool of chiplets to cherry pick those high clocking ones (in large enough quantaties). PS5/XBT wouldn't use anything with DDR4 controllers on-die. Sure this leak has some details that may be wrong but the overall approach looks solid to me.

If this turns out to be true, I was't really expecting AMD to be this ambitious at this point (later for sure) but if they got Sony (and MS) onboard (for PS5/XBTwo) then It would be a huge win for them. Then AMD's main goal would be to gain a lot more market share in this generation.

 

[Despoiler]

This is what I was going after. To be clear, I'm a skeptic of that 5.1GHz beast of a chip, hence my sarcasm laced question, which didn't come out too clearly, I might add.

I get where you are coming from. It's definitely not normal by the way today's products are sold, by either company, to have a 16c @ 5.1 when your 8c is only @ 4.8. One would think the lower core count would have the higher turbo because it's far more likely to be running a workload that would need it. The TDP would be inline with those workloads. That being said, my point was that all the details for all the leaks can move up or down as it gets closer to launch. We could align back to what is currently normal for a chip series or we might see AMD trying an entirely new strategy. ie a 16c with one chiplet being the best of the best for high speed 1 or 2 core turbo and all other chips being second fiddle. $500 for your halo chip, one that clocks the highest, is a very enticing way to make people want the halo chip. It's basically made for fanboying by spec.

 

[DrMrLordX]

A Ryzen 3000 that pushes all 16 cores at > 5 GHz won't exist.

The socket and motherboard simply won't be designed with that kind of power consumption in mind.

I can already push over 200W through my X370 Taichi. Maybe the barebones minimum spec for AM4 isn't designed for that kind of power output, but overbuilt boards already exist that can push that kind of power. The real question is whether the VRMs can handle the current load. Even the "good" X370 boards probably couldn't deal with a 16c monster. X570 or bust.

 

[Adul]

Kyle at hard|OCP did comment on the information that most of it was accurate, only a little of it was not.

 

[beginner99]

. It wouldn't make much sense to disable those working cores just because of power savings. They can however control how game developers are allowed to use them and keep the power usage at desired levels.

Because they are too expensive. Why would AMD sell a chiplet to a console for less than $100 (else the consoles will be too expensive) when it can sell it for more as a Ryzen, TR or Epyc CPU. 7nm is expensive. It IMHO isn't suitable for $400 consoles. Or their release date is in 2021 or even later.

 

[Zapetu]

That said, don't get hung up on the specifications — they may be preliminary, even merely estimates. The interesting thing is the mounting evidence that AMD will use chiplets in the AM4 socket, and that this means we will likely get up to 16 cores in the mainstream.

I agree we should wait for more information before concluding that AMD has achieved competitive single-thread performance.

I also wouldn't focus too much on details like TDPs, clock speeds, prices or even different SKUs. What I find most interesting about this leak, if it turns out to be true, is the chiplet design approach for the desktop. This two chiplets and one IO die on AM4 have been a popular speculation for a while and now it would seem that with dummy dies it could be solid from 6C up to 16C.

The most interesting new part is the Navi 20CU chiplet which would likely be about equal (if not exactly the same) size as the 8C chiplet. Sure it might be memory bandwidth starved with just 2xDDR4 channels but the basic idea behind it is very flexible and solid. As it wouldn't have any memory controllers on-die, it could easily use any kind of memory tech using different IO dies for different products (APU/dGPU). HBM2 is still expensive and would take a lot of space on AM4 but maybe HBM3 will use smaller packaging. Still there's the problem of interposers but it would only have to cover the area of HBM3-stack and the IO die. Still I'm not expecting HBM2 for AM4 socket at all and HBM3 is still far away..

The video claims 6+6 for 12-core, 4+4 for 8-core (except APU) and 6+dummy for 6-core (except APU), i.e. all salvaged dies.

That's why I think it would be so beneficial to have Sony/MS on-board (with 8C console chiplet) so AMD wouldn't have to waste too much working 7nm silicon for lower end SKUs. We all know that 6C and 8C Ryzen 3000s would most likely be the best selling desktop parts. And they would be using 6C and 4C salvaged dies according to leak, as you said.

Edit: Navi 12 changed to Navi 20CU chiplet.

 

[yuri69]

A local AMD overclocker, insider, etc. said this "leak" was an utter nonsense and re-confirmed that via his sources at AMD... This is the same guy who had reported a 7nm 8c ES boosting to 4.5GHz and benchmarks ran on a OCed ES@5GHz with AIO.

 

[PotatoWithEarsOnSide]

I'm not too convinced by the 7nm slide posted above. The reason being that the performance figures are likely uplifted by IPC gains, and not directly through process node gains. I'd be expecting a lower clocked 3xxxX to be hitting the same performance as a 2xxxX, so naturally it would have a lower power output irrespective of process node.
Without knowing the respective clock speeds at which those claims relate to, there's no way we can make any definitive conclusions about its power efficiency. I don't doubt that it's a bunch better, but the headline figures I suspect are rather misleading.

 

[exquisitechar]

A local AMD overclocker, insider, etc. said this "leak" was an utter nonsense and re-confirmed that via his sources at AMD... This is the same guy who had reported a 7nm 8c ES boosting to 4.5GHz and benchmarks ran on a OCed ES@5GHz with AIO.

Source?

 

[Timorous]

Because they are too expensive. Why would AMD sell a chiplet to a console for less than $100 (else the consoles will be too expensive) when it can sell it for more as a Ryzen, TR or Epyc CPU. 7nm is expensive. It IMHO isn't suitable for $400 consoles. Or their release date is in 2021 or even later.

Consoles might have much lower clock speeds which means 8c chiplets that bin poorly (below ryzen 3 spec) might be suitible for use in consoles and selling those rather than throwing them away is of benefit.

On top of that it is a guaranteed demand so it means they know they can order more wafers which may help them negotiate lower costs per wafer.

 

[Zapetu]

Because they are too expensive. Why would AMD sell a chiplet to a console for less than $100 (else the consoles will be too expensive) when it can sell it for more as a Ryzen, TR or Epyc CPU. 7nm is expensive. It IMHO isn't suitable for $400 consoles. Or their release date is in 2021 or even later.

Are they really that expensive? And as far as I know, Zen2 is designed specifically for TSMC 7nm HPC, so any 12/14nm would have to be Zen+, right? And PS5 will use Navi which is also TSMC 7nm HPC with larger die size than just 72mm². Sure PS5 could use a 7nm monolithic die with one 4C CCX but then if a critical defect hits the CXX the whole SoC would be unusable for PS5 and would be wasted. That's the beauty of the chiplet design, using the same 8C chiplet for vastly different markets.

EDIT: In fact the next consoles most likely will be 14nm still. I could see the gpu being 7nm but even that would be an issue price wise.

Sony/MS can suck up the initial costs. I'm sure that manufacturing costs will come down sooner than 2021. And we know for a fact that PS5 is going to use Navi (designed for 7nm HPC only, as far as I know) anyway so 14nm seems unlikely for MS either.

I'm just trying to justify for myself how AMD would mostly benefit from this chiplet approach and next gen consoled could be a big part of it. We can disagree about this and it really doesn't matter who's right or wrong and soon enough we'll find out anyway. It's good to have different opinions and point on views on many things. All of this is just speculation anyway.

On top of that it is a guaranteed demand so it means they know they can order more wafers which may help them negotiate lower costs per wafer.

And it's not only guaranteed demand, it would be guaranteed high volume demand. I'm sure TSMC would give some discounts for that.

 

[maddie]

Because they are too expensive. Why would AMD sell a chiplet to a console for less than $100 (else the consoles will be too expensive) when it can sell it for more as a Ryzen, TR or Epyc CPU. 7nm is expensive. It IMHO isn't suitable for $400 consoles. Or their release date is in 2021 or even later.

Because it might not be either one or the other but both at the same time, or are you saying that they can't get enough chiplets so they have to ration? Profit is profit and in a totally different market. We could also use the same argument against the desktop Ryzen as Rome has much higher prices and margins.