Speculation: Ryzen 3000 series

[PotatoWithEarsOnSide]

I figure that it is probably just the PCIe4. Doubtful there's enough room for anything significant like L4 cache. It's unlikely to be dead space though.

 

[trollspotter]

Well, like you said, the same amount of uncore as RR plus maybe a tiny extra for modernization like pcie 4, plus more lanes of pcie than RR since PR and SR have more.

If that leaves space then that includes the graphics related portions of RR minus most of the 11CU's.

Maybe stare at this some more. For size comparison, I think the CCX is like 40 some mm2.



My opinion is this: to not include about 35mm2 of area of graphics in the ~120mm2 controller would be a mistake. Fallback graphics are very useful for consumers, it'd be yet another feature upgrade in the 3000 series, and OEMs would find it irresistible.

After staring at this RR die shot some more and doing some quick measurements, I'm convinced there has to be something in there. I subtracted the area of 8 of the 11 Vega CUs and the CCX0 and came up with ~123mm2, the same size as the IOC. The additional eight PCIe lanes to get it from 16 to 24 don't take up very much die space, and since this is a desktop-specific IOC there may be other things they didn't need to include that free up a little bit of space.

3CUs is more than enough graphics power for a backup GPU or office tasks. Look at how closely the Athlon 200GE trails the 2200G in this test. If the 200GE had the same number of CPU cores and same CPU/IGP clocks this score would be even higher, and it only has 3 CUs compared to 8 in the 2200G. The 3 CUs don't seem to have the same memory bandwidth problems as 8 or 11 CUs.

 

[somethingclever]

View attachment 2458
After staring at this RR die shot some more and doing some quick measurements, I'm convinced there has to be something in there. I subtracted the area of 8 of the 11 Vega CUs and the CCX0 and came up with ~123mm2, the same size as the IOC. The additional eight PCIe lanes to get it from 16 to 24 don't take up very much die space, and since this is a desktop-specific IOC there may be other things they didn't need to include that free up a little bit of space.

3CUs is more than enough graphics power for a backup GPU or office tasks. Look at how closely the Athlon 200GE trails the 2200G in this test. If the 200GE had the same number of CPU cores and same CPU/IGP clocks this score would be even higher, and it only has 3 CUs compared to 8 in the 2200G. The 3 CUs don't seem to have the same memory bandwidth problems as 8 or 11 CUs.

View attachment 2456

I would expect what's on the IOC and taking up most of the space is most likely infinity fabric and it's supporting logic for up to 4CCXs.

We know from Zen 1 and + that IF has a significant power overhead and thus probably takes up a significant amount of die area. Scaling that logic to allow for up to 4 CCXs would probably take up even more area.

 

[HurleyBird]

I'm moderately skeptical that each CCX is still four cores (or that there is anything in there that looks like a traditional CCX in the first place), and massively doubtful that even if that were the case communication between CCXs in the same die would travel through the IO die or that each CCX would get its own data path rather than using a crossbar, since either would be stupidly wasteful.

There's enough room on there for a small IGP, and a small IGP would make sense, but there are also a million other possible explanations for why the IO die is the size it is. We should know either way on or closer to launch.

 

[Shivansps]

There's enough room on there for a small IGP, and a small IGP would make sense, but there are also a million other possible explanations for why the IO die is the size it is. We should know either way on or closer to launch.

It would make sence, specially for Ryzen PRO, and most people dosent need a big IGP for gaming. It would get a bit messy, explaning that a 2400G has better IGP than the Ryzen 7 3000... But it is acceptable. You know what im using right now for the people who wants a 2600/2700 only for work? a GT710, a Vega 3 would be a huge improvement.

 

[dnavas]

I know it's not as interesting, but couldn't the size difference just be down to using GF12nm vs TSMC16nm or better power distribution for higher clocking on desktop, or something like that?
An IGPU sounds like a lot of design effort for negligible gain.

 

[NostaSeronx]

[x]GMI2 supports I/O & CPU chiplet to GPU communication though.

 

[scannall]

It would make sence, specially for Ryzen PRO, and most people dosent need a big IGP for gaming. It would get a bit messy, explaning that a 2400G has better IGP than the Ryzen 7 3000... But it is acceptable. You know what im using right now for the people who wants a 2600/2700 only for work? a GT710, a Vega 3 would be a huge improvement.

Toss 1 CU into the chipset and be done with it. That's plenty for 2D work laptops with no gaming.

 

[Shivansps]

Toss 1 CU into the chipset and be done with it. That's plenty for 2D work laptops with no gaming.

And thats enoght to run Windows 10 desktop at high resolutions? im not convinced.

At any rate 1CU would be a waste considering how big the display and multimedia blocks are, 3CU would be enoght. Thats already faster than anything Intel has as well.

 

[amd6502]

Doesn't RR support 3 or 4 displays?

That's kind of overkill for bare basics fallback graphics. So maybe the display engine can be downsized. 2CU would be enough; maybe even 1CU. But each CU isn't that much, so 3 is not a bad number.

If they stuck 1 to 3 CU's in there I think it's understood that these aren't "real" APUs, but in fact CPUs that have bare essentials graphics capability.

It would be nice to believe, but realistically (given how little is known right now) basic video may or may not be a new feature.

I would expect what's on the IOC and taking up most of the space is most likely infinity fabric and it's supporting logic for up to 4CCXs.

We know from Zen 1 and + that IF has a significant power overhead and thus probably takes up a significant amount of die area. Scaling that logic to allow for up to 4 CCXs would probably take up even more area.

There's enough room on there for a small IGP, and a small IGP would make sense, but there are also a million other possible explanations for why the IO die is the size it is. We should know either way on or closer to launch.

 

[Thunder 57]

Since we're on this iGPU thing, it could be interesting. If AMD designed this for desktop use than it might make sense to include a small iGPU. I think the OEM's all want it. I mean when AMD bought ATI, "The Future is Fusion".

Yet on their top tier desktop chips still require a dGPU? I don't think AMD likes that. What if every new chip now comes with a basic iGPU?

And I was thinking about this:

It would make sence, specially for Ryzen PRO, and most people dosent need a big IGP for gaming. It would get a bit messy, explaning that a 2400G has better IGP than the Ryzen 7 3000... But it is acceptable. You know what im using right now for the people who wants a 2600/2700 only for work? a GT710, a Vega 3 would be a huge improvement.

Maybe they won't event advertise the basic iGPU. That way it still makes sense that a 2400G has way better graphics than a basic Ryzen 3000 because it doesn't have the magic Vega sauce.

 

[Shivansps]

Doesn't RR support 3 or 4 displays?

That's kind of overkill for bare basics fallback graphics. So maybe the display engine can be downsized. 2CU would be enough; maybe even 1CU. But each CU isn't that much, so 3 is not a bad number.

If they stuck 1 to 3 CU's in there I think it's understood that these aren't "real" APUs, but in fact CPUs that have bare essentials graphics capability.

It would be nice to believe, but realistically (given how little is known right now) basic video may or may not be a new feature.

And the multimedia block can be cut down as well, remove the encoder, Zen2 has enoght cores to do software encoding anyway.

 

[amd6502]

And the multimedia block can be cut down as well, remove the encoder, Zen2 has enoght cores to do software encoding anyway.

Software encoding would be a very good idea. Plenty of threads for that.

So, I've been pondering how a 16c/32t might be run without badly choking the dual channel memory. The answer is adding a mode to SMT, namely aSMT or asymmetric multi threading.

Strong thread+weak thread. It accomplishes what big.little core arrangements do in the acorn world.

The OS tasksets low priority tasks (niced processes) to "little" threads, which yield resources when another thread is running on the physical core.

 

[Mockingbird]

Let me make it clear why AdoredTV's list is very very likely fake. (If, for whatever reason, you are still unconvinced otherwise.)

Games, by design, are hard to parallel across many cores.

You would see a huge benefit going from 2C to 4C, but a significant (but smaller) benefit going from 4C to 6C.

You hardly see a benefit going from 6C to 8C. (a few games can benefit, but the majority cannot)

That is hence the reason why the the 6C Ryzen 5 is AMD's best selling product.

Now, according to AdoredTV's list, AMD would cut its 6C processor's MSRP in half, from $199 to $99.

Someone would point out that Ryzen 5 2600 is selling for ~$160, which is below MSRP.

This is true, but the hypothetical Ryzen 3 3300 would likely also go on sale after being released (possibly to $80).

So let's think about that for a moment: AMD's new best selling processor is selling for $80. (down from $160)

That would significantly hurt AMD's bottom line.
_______________________________________________________________________________________

More likely than not, AMD is going to keep the number of cores in the Ryzen 3, Ryzen 5, and Ryzen 7 the same.

Ryzen 9 would target new price points.

12C/24T Ryzen 9 3800X at $500 to $600 would target Core i9-9900K

16C/32T Ryzen 9 3900X at $700 to $800 would target Intel's unreleased 10-core Coffee Lake Refresh

There's some overlap with Threadripper, but the people who buy Threadripper probably need the extra PCIe lanes that the Ryzen 9 doesn't have.

 

[DuronBurgerMan]

I'd buy a 12c/24t AM4 chip at $500 or less if it dropped into my X370 board without issue.

 

[somethingclever]

Games, by design, are hard to parallel across many cores.

You would see a huge benefit going from 2C to 4C, but a significant (but smaller) benefit going from 4C to 6C.

For the last time, AMD did not design zen2 with gaming in mind! Yes you're right that they do not scale very well, but I can assure you that that isn't a deciding factor in how AMD decided it's lineup.

There are significantly larger market forces in play, freelance and professional content creators, OEM companies, and many more gamers that simply don't care about or can't afford top performance.

We won't know if AdoredTV is wrong until later this year, but gaming won't be the reason why his predictions aren't accurate.

So let's think about that for a moment: AMD's new best selling processor is selling for $80. (down from $160)

I think you're significantly overestimating the cost of AMDs chiplets. The entire basis of AMD using chiplets is to cut costs in two ways:

Only 1 7nm die saves them money on layout, test runs, and validation. (A significant amount of money)

Having a smaller die gives higher yields.

Depending on how successful those two strategies play out and how well TSMC yields are, it could cost AMD next to nothing for a chiplet it can't sell as Rome or TR.

I'm willing to bet that every if it doesnt end up being 80$ it'll still be lower than the current 160.

 

[Mockingbird]

For the last time, AMD did not design zen2 with gaming in mind! Yes you're right that they do not scale very well, but I can assure you that that isn't a deciding factor in how AMD decided it's lineup.

There are significantly larger market forces in play, freelance and professional content creators, OEM companies, and many more gamers that simply don't care about or can't afford top performance.

We won't know if AdoredTV is wrong until later this year, but gaming won't be the reason why his predictions aren't accurate.



I think you're significantly overestimating the cost of AMDs chiplets. The entire basis of AMD using chiplets is to cut costs in two we ays:

Only 1 7nm die saves them money on layout, test runs, and validation. (A significant amount of money)

Having a smaller die gives higher yields.

Depending on how successful those two strategies play out and how well TSMC yields are, it could cost AMD next to nothing for a chiplet it can't sell as Rome or TR.

I'm willing to bet that every if it doesnt end up being 80$ it'll still be lower than the current 160.

You forget that it isn't just about the how much it costs to make the processors and ship them to the store, but also, development cost.

Also, it's pretty obvious that gamers make up a big chuck of AMD's customers unlike Intel who mostly ship to OEMs.

 

[Atari2600]

Also, it's pretty obvious that gamers make up a big chuck of AMD's customers unlike Intel who mostly ship to OEMs.

But AMD don't want them to make up a big chunk of their revenue.

Yes they want to keep selling to gamers - but they'd much rather a massive chunk of their revenue was coming from server/workstation markets - as that is where the real profits are.

 

[somethingclever]

You forget that it isn't just about the how much it costs to make the processors and ship them to the store, but also, development cost.

Also, it's pretty obvious that gamers make up a big chuck of AMD's customers unlike Intel who mostly ship to OEMs.

I literally was talking about development costs in my first point.

Gamers may make up a significant fraction of AMDs marketshare at the moment, but like i said and which has been mentioned earlier in the thread, there are much larger market segments that AMD is in a prime position to gain a share in; AMD realizes this as well and they'd be fools not to take advantage of it. Intel is having supply issues with 14nm, a stagnant architecture, and higher price.

How do they further incentivize that market to buy their cpus instead of intels?

Offer 50% more cores at the same or lower price, at comparable or greater performance, and with greater power efficiency (as we saw with the ES demo).

Even if AMD has lower margins with those prices, it's probably a trade-off they're willing to take to gain market share. They can improve their margin by improving yield or in future products.

 

[PotatoWithEarsOnSide]

Let me make it clear why AdoredTV's list is very very likely fake. (If, for whatever reason, you are still unconvinced otherwise.)

Games, by design, are hard to parallel across many cores.

You would see a huge benefit going from 2C to 4C, but a significant (but smaller) benefit going from 4C to 6C.

You hardly see a benefit going from 6C to 8C. (a few games can benefit, but the majority cannot)

That is hence the reason why the the 6C Ryzen 5 is AMD's best selling product.

Now, according to AdoredTV's list, AMD would cut its 6C processor's MSRP in half, from $199 to $99.

Someone would point out that Ryzen 5 2600 is selling for ~$160, which is below MSRP.

This is true, but the hypothetical Ryzen 3 3300 would likely also go on sale after being released (possibly to $80).

So let's think about that for a moment: AMD's new best selling processor is selling for $80. (down from $160)

That would significantly hurt AMD's bottom line.
_______________________________________________________________________________________

More likely than not, AMD is going to keep the number of cores in the Ryzen 3, Ryzen 5, and Ryzen 7 the same.

Ryzen 9 would target new price points.

12C/24T Ryzen 9 3800X at $500 to $600 would target Core i9-9900K

16C/32T Ryzen 9 3900X at $700 to $800 would target Intel's unreleased 10-core Coffee Lake Refresh

There's some overlap with Threadripper, but the people who buy Threadripper probably need the extra PCIe lanes that the Ryzen 9 doesn't have.

I'm not sure that incorrect pricing necessarily invalidates the entirety of the leak.
I'm also not convinced that their existing best seller would continue to be their best seller, though I can see your logic in that it could potentially lower the ASP which might in turn "hurt" AMD. However, reduced margins on increased sales can still produce significantly more profit, both short and long term, so even that "hurt" is relative; reduced margins on increased volume actually hurts Intel more than it would AMD, as they are the ones losing sales and margins.
In relation to the second part of your post, I can see this happening if the performance gain from one generation to the next is above and beyond what we've become used to. However, I suspect that your R9 pricing may be a little on the high side; the goal won't be to maximise margins, rather to decimate Intel's own high margin sales, and this is best done by beating Intel on all three metrics by a distance (price, performance, and performance per watt).

 

[Spartak]

1. TBA dates were incorrect
2. G-versions are incorrect (if they arrive it will be next year and they will be Renoir based, not chiplets)
3. extended TDP range was incorrect
4. Wishful pricing at low end

Case closed.

 

[somethingclever]

1. TBA dates were incorrect
2. G-versions are incorrect (if they arrive it will be next year and they will be Renoir based, not chiplets)
3. extended TDP range was incorrect
4. Wishful pricing at low end

Case closed.

FFS can you stop ranting about how AdoredTV was wrong? This is a SPECULATION THREAD on what is or isn't POSSIBLE and any TECHNICAL ARGUMENT for or against those possibilities.

Anyone legitimately participating could give a rats ass about whether AdoredTV is right or wrong. The reason why he's relevant in this discussion at all is because he offered good TECHNICAL ARGUMENTS that supported a SPECULATED POSSIBILITY. From the ES demo at CES we know one of his speculations was true: AMD is extending the chiplet/IO die strategy to Ryzen, and there is the potential for up to 16C/32T Ryzen. For the rest, not even you can say he was wrong yet because Ryzen hasn't been released.

Participate in the discussion or shut up, read, and actually put some effort into learning more about the industry. If you want to discuss AdoredTV's credibility or merits, start a new thread or something.

 

[Spartak]

Don't get your undies twisted just because some leak is proven wrong.

FYI the chiplet speculation was his own and completely unrelated to the leak.

 

[somethingclever]

Don't get your undies twisted just because some leak is proven wrong.

FYI the chiplet speculation was his own and completely unrelated to the leak.

That's my point; he was right about the chiplets!

You have any technical insight/arguments against his "leaks" (whatever that means to you)? I'd love to hear them because I don't know enough about TSMC's 7nm process, and AMD's metrics on power/frequency/feature scaling to argue for or against them myself. I'm just here to talk and learn more about the architecture.

 

[Shivansps]

To me, and this is just my opinion, the simple fact that the CPUs are H2 2019 makes the whole table false just by the fact it has way too much information too soon. And the 240GE at $75 kinda makes the $99 6C/12T very very unlikely.

But still it may end up correct in a few cases, especialy SKU names and core configuration.