Speculation: Ryzen 3000 series

[DarthKyrie]

Because they need the power efficiency and simplicity of a single die solution for 10-25w laptops.

This still doesn't make sense with AMD's whole philosophy with regard to Zen 2 and I/O not shrinking as well as the cores.

 

[zinfamous]

What makes you think the yields of an 80mm2 7nm chip are low?

People here underestimate the enormous benefit AMD has found in doing your uncore on your n-1 node: you largely increase yields on your important bit, the n node core die.

It may not be a physical yield issue, but a desire to reserve those chiplets for the much higher margin Rome chips....which actually makes sense. Again and again and again, AMD has talked about focusing on regaining significant market share in the server space, and this is how you do it.

This is why I think 16c Ryzen is not going to be on initial release for Ryzen 3. Could be wrong, of course, but I would think AMD wants to grab the higher-margin market first.

 

[Mopetar]

That's his new video:

Interesting thoughts, but the one thing that rubs me the wrong way is his thinking that R5 (8C/16T) parts will use two chiplets. It's certainly possible, but it just seems unlikely to me that AMD will have a substantial number of chiplets that need to have 4 cores disabled. The yields would have to be a lot lower than most people have speculated.

Additionally, just because Rome will have up to 64-cores, doesn't mean that customers won't want to buy server chips with lower core counts. It seems like putting such chips in an EPYC CPU makes more sense, especially if fully functional chiplets are more common and you can make an R5 with just one chiplet instead of two.

It's even more complex once you start looking at clockspeeds because now you'd need two chiplets that fall into the same bin. R5 doesn't seem like it could be a volume part if they need to use as many chiplets as the higher end 12 and 16 core parts.

 

[PeterScott]

Hey now, let's not pretend I'm the one who started needlessly leaving negative points, I'm just the one who took it to the obvious conclusion.

You chose to keep squirming your way out of any questions and then come back weeks later to leave negative reputation, I just choose to take you with me.

I gave you one downvote, you left 40 last night, then 170+ more this morning.

 

[bsp2020]

This tweet makes absolutely zero sense in regard to what AMD has stated about the I/O portion of Zen 2. Why would AMD go thru all the trouble of shrinking the I/O just for APU's?

My speculation (not a leak or rumor) is that there will be two APUs for mobile. Lower end one will be a SoC with integrated CPU and CPU cores as well as minimal IO capabilities to replace Raven Ridge for 15W or less devices. AMD should build high power APU with a CPU chiplet and a GPU chiplet to serve 35W or higher segments.

 

[mattiasnyc]

I gave you one downvote, you left 40 last night, then 170+ more this morning.

People care about votes? What is this; kindergarten?

Point us to that video you were talking about instead.

 

[PeterScott]

People care about votes? What is this; kindergarten?

Point us to that video you were talking about instead.

I don't care about normal voting, but 220+ downvotes from one person, in one day, is crazy town.

I don't like adored tv, so I am not going to link it.

 

[Kedas]

Interesting thoughts, but the one thing that rubs me the wrong way is his thinking that R5 (8C/16T) parts will use two chiplets. It's certainly possible, but it just seems unlikely to me that AMD will have a substantial number of chiplets that need to have 4 cores disabled. The yields would have to be a lot lower than most people have speculated.

Yeah I also don't see this 2 times 4 cores happening for 8 core AM4 CPUs
I assume the same ratio of yield they had with 8 core zeppelin.
For threadripper 1900X there was a reason for 2x4 cores, to get enough I/O, but this is not the case now (separate die).
AMD already did show the one 8 core chiplet AM4, I don't see them change their mind about that, maybe Adoredtv or his source is confusing CCX with chiplet or lost in translation.

 

[lobz]

I don't like adored tv, so I am not going to link it.

Now that is kindergarten You just basically prove their points about you.

 

[PeterScott]

Now that is kindergarten You just basically prove their points about you.

Hardly. If I don't like a business model (posting wild speculation for views), I am not going to support it, by posting links to it, or viewing anymore of them myself.

It isn't kindergarten, it is simply boycotting a business that I don't like.

 

[lobz]

Hardly. If I don't like a business model (posting wild speculation for views), I am not going to support it, by posting links to it, or viewing anymore of them myself.

It isn't kindergarten, it is simply boycotting a business that I don't like.

Don't pretend that you don't know you could describe exactly how and why do you say these things about him, without having to 'link' a video or 'supporting' him (with what exactly?)... I honestly don't care if you don't like his business model, the world is beautiful because we are all different. But avoiding an explanation at any cost after your wildly thrown general accusations towards him, is kindergarten for me.

 

[Tuna-Fish]

This tweet makes absolutely zero sense in regard to what AMD has stated about the I/O portion of Zen 2. Why would AMD go thru all the trouble of shrinking the I/O just for APU's?

I'm in the camp chiplet APU myself, but there is a sensible reason for that: In a GPU, there is more data traffic between the ROPs and the memory interface than there is between the ROPs and the rest of the GPU. This means that for power, it makes a lot of sense to integrate the ROPs with the memory controller. However, ROPs do benefit from shrinks and modern ones are power-hungry enough for shrinking them to have a meaningful impact of total power consumption, especially in the lower-wattage models.

And shrinking the IO chiplet does not help cost, but it does still help power a little. So especially for a laptop, a monolithic APU might make sense.

 

[PeterScott]

Don't pretend that you don't know you could describe exactly how and why do you say these things about him, without having to 'link' a video or 'supporting' him (with what exactly?)... I honestly don't care if you don't like his business model, the world is beautiful because we are all different. But avoiding an explanation at any cost after your wildly thrown general accusations towards him, is kindergarten for me.

What accusations?

 

[mattiasnyc]

I don't care about normal voting, but 220+ downvotes from one person, in one day, is crazy town.

I don't like adored tv, so I am not going to link it.

Ok, but bear with me for just one second:

I'm relatively new here, and I don't care one bit about this up/downvotes or reputation or anything. What I try to do is to look at what someone writes and then figure out if it makes sense. I'm a human, so sometimes I fail at expressing myself, and sometimes I fail at understanding people. The thing is though, your "reputation" or whatever it's called could be 100% great, but that doesn't change the fact that you toss out what appears to be accusations or attempts to smear someone, and when someone asks for the absolute lowest bar for justifying what you claim your reply is 'don't like it, won't link it'.

That's incredibly weak regardless of your "vote-count".

See what I'm getting at here?

You can spend time whining about up/down votes, or you can spend three seconds linking to that video... ya know?....

 

[Mopetar]

Yeah I also don't see this 2 times 4 cores happening for 8 core AM4 CPUs
I assume the same ratio of yield they had with 8 core zeppelin.
For threadripper 1900X there was a reason for 2x4 cores, to get enough I/O, but this is not the case now (separate die).
AMD already did show the one 8 core chiplet AM4, I don't see them change their mind about that, maybe Adoredtv or his source is confusing CCX with chiplet or lost in translation.

There is some sense to a 4+4 in that if you assume they can all hit the same high clock speeds, that it is easier to dissipate the heat if it's coming from two separate dies. It just seems unlikely to naturally have a large enough supply of chiplets with only 4 working cores for it to make sense. Depending on how everything is laid out there's probably some silicon that if defective takes out half the cores (even if the cores themselves work fine) by severing some important connection, but that would be a small area relative to other parts of the die (individual core components, cache, etc.) that are more likely to be defective.

Intentionally disabling additional cores from what could work as a 6-core chiplet seems pointless in a general sense as well. Maybe in limited circumstances if you have a chiplet where 4 of the cores can hit a really high clock speed, but the remaining two really hold it back. I don't know if AMD would evaluate each chiplet to that extent, but they might considering that these will be used across so many different product lines which different use cases and requirements.

 

[mattiasnyc]

I agree Mopetar...

 

[somethingclever]

And shrinking the IO chiplet does not help cost, but it does still help power a little. So especially for a laptop, a monolithic APU might make sense.

Any thoughts on ryzen chiplet + 7nm gpu/io die?

 

[PeterScott]

Ok, but bear with me for just one second:

I'm relatively new here, and I don't care one bit about this up/downvotes or reputation or anything. What I try to do is to look at what someone writes and then figure out if it makes sense. I'm a human, so sometimes I fail at expressing myself, and sometimes I fail at understanding people. The thing is though, your "reputation" or whatever it's called could be 100% great, but that doesn't change the fact that you toss out what appears to be accusations or attempts to smear someone, and when someone asks for the absolute lowest bar for justifying what you claim your reply is 'don't like it, won't link it'.

Again. What accusations? Where are all the terrible things I said that need evidence?

 

[mattiasnyc]

Again. What accusations? Where are all the terrible things I said that need evidence?

"He posts rumors and calls them "leaks"." In what seems to be support and or defense of "FakeTV again?"

 

[Topweasel]

"He posts rumors and calls them "leaks"." In what seems to be support and or defense of "FakeTV again?"

Or

His "contacts" are random anonymous people who send him info, and wildly inaccurate info lately, so no I don't call those leaks.

While he started the one video I watched stating, "take it with a grain of salt" disclaimer, he then spent the next 5 minutes citing why he thought it was true.

That's about as useful a disclaimer as every thing that tells you anything health related to talk to your doctor before exercising, changing your diet etc...

It's not a "allegation" so much as stating someone stating that what they think is true is what they think is true, while cautioning people it could be be note true some how makes it Fake or something weird like that. Like we haven't had "insider information" including public employees being being the leaker (anyone remember the marketing dude for AMD pre BD launch?) be wrong in the past. Second who cares about a distinction between "Leaks" and "Rumors" rumors are just leaks that nobody attaches their names too (ignoring hoaxes). In Peter's defense he isn't the one calling it "Fake news". But the callout's of the vlogger do seem a bit misguided.

 

[PeterScott]

"He posts rumors and calls them "leaks"." In what seems to be support and or defense of "FakeTV again?"

So what exactly in dispute? Are you claiming Adored TV doesn't call his information "leaks"? I thought that was pretty well established that he does.

 

[Topweasel]

So what exactly in dispute? Are you claiming Adored TV doesn't call his information "leaks"? I thought that was pretty well established that he does.

He doesn't call them leaks. He said information he gets from his sources. Same as every other tech news. They might go into slightly more detail like vendors, motherboard makers, packagers. He just says his sources.

That imho does make them leaks instead of lets say Wccftech who tends to start rumors by saying rumor has it first and then everyone quotes them about the rumors. Never suggesting they have some inside person down the chain. Just throwing it out there.

Question is does it matter if we or he call them leaks or rumors? He had a source he believed some of it backed up from other sources that made him feel more confident about the originals. He told us about it. Are we really at the point were we are going to bash a tech news and information source because they got "bad" information (outside product announcements at CES has most of his stuff fallen through yet?)? Even though he cautions us to take it with salt (FYI using any salt means this isn't to be taken as fact, the level of salt grain to Ton is just a measurement of personal confidence. Smallest is high confidence, largest lowest confidence )? This is why we can't have nice things. We have to destroy anything that isn't 100% perfect.

 

[Mopetar]

Any thoughts on ryzen chiplet + 7nm gpu/io die?

It seems like the next logical evolution, but the biggest impediment seems to be that the chiplet approach doesn't work quite as well for GPUs. There was some article where an interviewer asked someone at RTG about it and the gist was that AMD has looked into the approach, but found that it didn't work well for gaming. If you think about it, it's almost like SLI/Crossfire in a way since the chiplets would be small GPUs that all get hooked together. We know that SLI/Crossfire solutions have always had some overhead and that's if they worked at all. Unless AMD has some magic drivers (issues with Vega make this hard to believe) then this approach isn't going to work for a huge chunk of the market.

Of course that doesn't mean that they can't just produce a chiplet sized GPU for this part of the market and make bigger monolithic GPUs for the midrange and high-end. It would probably also necessitate some changes to the IO die as well, so you might need another separate part their as well. If they can get around that or they've just designed their desktop IO die with that in mind and have extra stuff that isn't enabled with Ryzen chips, it's not as much of an issue. It's just a matter of economics and how much extra space they waste on each multi-purpose IO die as opposed to the extra costs of having two separate produce lines.

If you have to make a separate GPU chiplet that doesn't get used anywhere else and a separate IO die that's necessary for an APU, why not just make a monolithic chip at that point? By the time AMD refreshes the mobile chips (Picasso is Zen+ on 12 nm) the 7nm process will be much more mature and capable of handling the larger monolithic die. Or maybe they want to start going to the chiplet route because they are trying to make chiplet Crossfire work seamlessly from a software perspective. Being able to do that has far bigger implications beyond APUs though, especially if it scales beyond 2 or 4 chiplet configurations. They might even do it anyway because apparently having a chiplet-based approach works fine for a lot of compute tasks, so we get chiplets for APUs and MI cards, but monolothic GPUs for gaming.

 

[moinmoin]

Personally I still think there is perfectly enough room for both a power optimized monolithic low end APU with up to 4 cores where the uncore is cut down to the bare necessary amount as well as a IOC/chiplet/GPUlet MCM that goes up to 8 cores and has all the standard AM4 IO to exist alongside. But it may well be the case the platform and/or IOC is not prepared for the latter.

 

[PeterScott]

He doesn't call them leaks.

IIRC the video I watched was called:

"Ryzen 3000, Radeon 3000 Series LEAKS"