Speculation: Ryzen 3000 series

[dlerious]

i always thought that amd will aggresivelly increase the core count till intel finally catches up
my predictions are kinda like adored
6c/6t entry level and some navi/vega
6c/12t entry level with and without vega/navi
10/12c mid level with and without gpu
and 16 /18c at high level

now about the hedt i guess they will go either from 16c or go full crazy at 20c

HEDT already has Threadripper with 12/16/24 and 32 cores.

 

[whm1974]

HEDT already has Threadripper with 12/16/24 and 32 cores.

AMD might end just having only APUs on AM4 with 7nm only offering higher core counts with Threadripper and EPYC.

However I'm sure what would be the market for an 8c/16 APU. SFF builds?

 

[mindfury]

However I'm sure what would be the market for an 8c/16 APU. SFF builds?

Highend bussiness PC, lowend workstation & entry level server.

I think the io die on AM4 platform will have an igpu inside, thus AMD will have an easier time of breaking into highend OEM PC market.

New TR platform likely will still top at 32 cores, but runing at much higher frequency. There is chance that TR platform could get its own io die if AMD decides to make a big push towards highend workstation & lowend server market.

7nm monolithic APU won't arrival before DDR5. HBM is too costy for APU, so I think AMD will wait for DDR5.

 

[ericlp]

It's kinda funny, to hear everyone talk about "waiting for intel to catch up"... I think Intel may never catch up. Will be interesting... By the time Intel comes out with their version of the 7nm (whatever it may be)... Hopefully not 10nm. AMD will be pushing 5nm. It's going to be an interesting ride. Hopefully Intel will come out with a 7nm chip to compete and soon. In sometime 2020 AMD will be releasing a 5nm chip. They are just on fire, and it's really a good thing for all of us to finally benefit the reality that the tech is headed for great times ahead! Even if China is spending the money to fab it on a mass scale. I for one am happy to see it! I kinda was hoping Intel would have gotten it together to make this type of progress but....sadly, they did not. I should ad, that if I were AMD, I'd keep on blazing down new paths. I wouldn't wait for anyone.

 

[darkswordsman17]

I could see a blurring between platforms, where they kinda spec out something that uses the Threadripper socket, but is rated for less power, maybe uses a cheaper mechanism, and other things that would help push the costs down, but would enable features beyond AM4, but they could segment Threadripper systems from. And then have $500-800 chips. And then Threadripper becomes a bit more premium ($1000-3000 chips). Segment it so that the middle one gets 4 memory channels, but Threadripper gets 6-8, and fewer PCIe channels, and some other aspects (rated for lower TDP, maybe top it at 120-150W, while Threadripper is 150-250)

Or maybe they wait until a socket change (AM5). Keep it to 3-4 platforms (assuming mobile is kinda its own for the 4th one). AM5 grows, and covers consumer and HEDT (with 4 or maybe even up to 6 memory channels), partly so that it can also accommodate larger GPU (maybe they have it so some of the memory sockets take GDDR modules). They keep AM4 as a budget option. And then EPYC grows, with Threadripper basically being single socket version of that (with EPYC moving to support 4 or maybe even more sockets), and its got 8-12 channels of memory per socket.

 

[beginner99]

If Ryzen also uses chiplets hence the same I/O design as EPYC/TR but less channels.
doesn't that mean that the game performance of a Threadripper will be the same as a Ryzen CPU?

My thought as well unless the different IO dies have an influence. But as long as it's better than current ryzen this is only a good thing as it makes TR also a better option for gamers that do other stuff than just gaming.

That said it's kind of a perfect storm, it took Intel to completely botch 10nm for this to work.

On the other hand even if 10nm worked I doubt we would have a faster product for desktop available. Even on older intel roadmaps they cleary showed that earliest with 10nm+ they would have reached clock speeds of 14nm++. Yes we would have avx512 on client but that would mostly be useless for gaming. Even with IPC increase I doubt it would actually beat a 8700k or 9900k at ST performance simply due to the insane clocks.
So for the desktop the 14nm delay wasn't really an issue. For server however, yes it's a huge issue.

 

[epsilon84]

It's kinda funny, to hear everyone talk about "waiting for intel to catch up"... I think Intel may never catch up. Will be interesting... By the time Intel comes out with their version of the 7nm (whatever it may be)... Hopefully not 10nm. AMD will be pushing 5nm. It's going to be an interesting ride. Hopefully Intel will come out with a 7nm chip to compete and soon. In sometime 2020 AMD will be releasing a 5nm chip. They are just on fire, and it's really a good thing for all of us to finally benefit the reality that the tech is headed for great times ahead! Even if China is spending the money to fab it on a mass scale. I for one am happy to see it! I kinda was hoping Intel would have gotten it together to make this type of progress but....sadly, they did not. I should ad, that if I were AMD, I'd keep on blazing down new paths. I wouldn't wait for anyone.

Intel has been behind AMD before. Athlon 1GHz ring a bell? Or how about the Athlon 64? Tech goes in cycles, I believe Intel has made the right choice in sacking the CEO and also recruiting Jim Keller. They will be fine, even if they have to endure some short term pain.

Also, I believe using nm figures to compare progress is inaccurate and an oversimplification, because Intel 7nm != TSMC 7nm in terms of density. I believe Intel's 10nm node (at least the initial version) was roughly on par with TSMC 7nm density wise. I'll hazard a guess that Intel 7nm will have higher transistor density than TSMC 7nm and may rival TSMC 5nm in this regard. Regardless, I think 10nm is almost in write off territory for Intel and may be one of the shortest lived nodes ever. They should just cut their losses and move to 7nm ASAP.

 

[Shivansps]

AMD might end just having only APUs on AM4 with 7nm only offering higher core counts with Threadripper and EPYC.

However I'm sure what would be the market for an 8c/16 APU. SFF builds?

Every cpu on mainstream should have an IGP, most people dosent play games or need an IGP bigger than Vega 3 for the facebook games they play.

IF amd end up releasing up to 8C on 3000 series, but every one of then has an IGP that would be perfectly fine.

 

[whm1974]

Every cpu on mainstream should have an IGP, most people dosent play games or need an IGP bigger than Vega 3 for the facebook games they play.

IF amd end up releasing up to 8C on 3000 series, but every one of then has an IGP that would be perfectly fine.

I can see your point, but being om AM4 w/ DDR4 will put limits on how powerful the IGP can be, especially given that it has to share that with the CPU cores.

 

[DarthKyrie]

Intel has been behind AMD before. Athlon 1GHz ring a bell? Or how about the Athlon 64? Tech goes in cycles, I believe Intel has made the right choice in sacking the CEO and also recruiting Jim Keller. They will be fine, even if they have to endure some short term pain.

Also, I believe using nm figures to compare progress is inaccurate and an oversimplification, because Intel 7nm != TSMC 7nm in terms of density. I believe Intel's 10nm node (at least the initial version) was roughly on par with TSMC 7nm density wise. I'll hazard a guess that Intel 7nm will have higher transistor density than TSMC 7nm and may rival TSMC 5nm in this regard. Regardless, I think 10nm is almost in write off territory for Intel and may be one of the shortest lived nodes ever. They should just cut their losses and move to 7nm ASAP.

Intel needs 10nm to work because their 7nm is a shrunk version of 10nm.

 

[moinmoin]

That said it's kind of a perfect storm, it took Intel to completely botch 10nm for this to work.

I disagree. AMD created good products in their own right and keeps innovating.

Yes, Intel botched 10nm, botched Spectre/Meltdown, botched the timing for increasing the core count, botched their server lineup segmentation and marketing. But the comparison just accentuates how fundamentally decent AMD's products are, not change their quality. Now it's the customers' job to react accordingly in this "perfect storm" of misses on Intel's part.

 

[advt.naveen]

As AMD moved io in a separate die and cores only as chiplets how likely can they make ARM cores in place of x86 cores and use the available io dies with them, it may not be that easy as it sounds but it may opens way to explore some more emerging markets. X86 for compute heavy and ARM for small parallel work loads like cloud storages.

 

[Shivansps]

I can see your point, but being om AM4 w/ DDR4 will put limits on how powerful the IGP can be, especially given that it has to share that with the CPU cores.

I dont think there is much interest on AMD to make powerfull IGP on desktops anyway. They would be eating their own GPU marketshare, Intel IGP are (so far) no competition and im petty sure Sony is not going to like that.

We need to wait and see how the Navi chiplets are made, if you ask me i think they arent going to make one specially for CPUs, to me the Navi chiplets are going to be the same used on GPUs, that includes having not only the IF link, but also a built-in GDDR5/6/HBM controller, this will allow to have notebooks with powerfull IGPs, this would be the best to combat Nvidia in the mobile market.

This is not going to matter much for desktop as you cant take advantage of that whiout a socket change, unless they can also package IGP memory.
But If there is a big enoght L4 there will be some improvements.

 

[whm1974]

I dont think there is much interest on AMD to make powerfull IGP on desktops anyway. They would be eating their own GPU marketshare, Intel IGP are (so far) no competition and im petty sure Sony is not going to like that.

We need to wait and see how the Navi chiplets are made, if you ask me i think they arent going to make one specially for CPUs, to me the Navi chiplets are going to be the same used on GPUs, that includes having not only the IF link, but also a built-in GDDR5/6/HBM controller, this will allow to have notebooks with powerfull IGPs, this would be the best to combat Nvidia in the mobile market.

This is not going to matter much for desktop as you cant take advantage of that whiout a socket change, unless they can also package IGP memory.
But If there is a big enoght L4 there will be some improvements.

True, and folks like me will get a video card anyway.

 

[Atari2600]

I can see your point, but being om AM4 w/ DDR4 will put limits on how powerful the IGP can be, especially given that it has to share that with the CPU cores.

I wouldn't mind if they just incorporated a very basic IGP part.

Forget about gaming, enough to run the computer for typical 2D tasks.

It would be enough to cover the office market, and also be a very useful feature when your GPU dies!

 

[ozzy702]

Intel needs 10nm to work because their 7nm is a shrunk version of 10nm.

I don't think this is correct. Two separate teams were working on 7nm and 10nm and while there may be some overlap, it sounds like the 7nm team is on or ahead of schedule and doing fine.

 

[Shivansps]

Basic IGP would be the Vega 3, with full video encoder/decoder. You need some IGP power to play flash/html5/facebook games.

If AMD made a basic GPU chiplet like that they could also use it to create a basic dGPU to remplace all those HD5450/HD6450, pair that with a 64-bit DDR4 memory interface and off you go. Modern and basic GPU for next to nothing.

Now AMD is in a position to do wharever it want, and the cheapper APU are the ones that need more powerfull IGPs not the $400 cpu.

Hell they could even make BOTH versions, basic and big IGP. Buy wharever you need.

 

[Mopetar]

One interesting angle would be if they could find a way to have a Crossfire-like setup with an IGP and discreet card. That would probably get around issues with stronger IGPs hurting card sales (which I’m dubious about anyway) and maybe even help them sell more cards as it creates an upgrade path.

There are some issues though. First is that multi-GPU is pretty much dead in terms of support so it’s a hard sell. The second is that you need multiple markets for a chiplet strategy to work. Zen is used across their entire product line and maybe third party consoles if rumors are to be believed. I think it’s harder to do that with a small Vega chiplet.

 

[Shivansps]

They dont need to since DX12 already does this with every GPU as long as the game supports this (no one cares to support it).

I just think the GPU chiplet will be the same ones for dGPUs, there is no point in making 1 chiplet for just a few APUs, just add an IF link to the GPUs you are making for discrete cards and call it a day.

No matter from what point of view i look at it this is the best way to do it. If they can fit a Navi 12 in there.

 

[DrMrLordX]

I can see your point, but being om AM4 w/ DDR4 will put limits on how powerful the IGP can be, especially given that it has to share that with the CPU cores.

I suspect that client/desktop parts may forgo the traditional on-die APU design (HSA compliant) and go for something similar to Kabylake-G, only connecting via an IF link to a GPU/HBM chiplet on an interposer rather than using EMIB.

Mobile parts and parts in the "official" APU lineup (Raven Ridge, Picasso, etc.) will still host GPU and CPU on the same die with a single memory controller connected to system RAM. Those parts will remain fully-compliant for SVM operations (HSA, OpenCL2.0).

I don't think this is correct. Two separate teams were working on 7nm and 10nm and while there may be some overlap, it sounds like the 7nm team is on or ahead of schedule and doing fine.

If late 2021 counts as "on time" then yes, they are on time. Not sure where Intel 7nm showed up on roadmaps as far back as 2016.

 

[airfathaaaaa]

HEDT already has Threadripper with 12/16/24 and 32 cores.

i wouldnt call the wx series a true hedt but i can see them upping the wx series up untill 40-42c and the tr to 24-28-32

 

[Shivansps]

I suspect that client/desktop parts may forgo the traditional on-die APU design (HSA compliant) and go for something similar to Kabylake-G, only connecting via an IF link to a GPU/HBM chiplet on an interposer rather than using EMIB.

Mobile parts and parts in the "official" APU lineup (Raven Ridge, Picasso, etc.) will still host GPU and CPU on the same die with a single memory controller connected to system RAM. Those parts will remain fully-compliant for SVM operations (HSA, OpenCL2.0).

Why? imagine this, Navi 12 chip, the same one produced for dgpus has a IF link, then you use that chip as IGP for AM4 desktop APU, whiout direct memory the chip would run off the main memory via the IF link. Running on low clocks and cut off CU, probably harvested dies.
It makes sence, remember that 2400G IGP is almost as big as a Polaris 12. Now you only mass produce 1 chip for both APU and dgpu.

In notebooks, OEM would have the option to attach GDDR5 to use as VRAM.

For the PS5 they would use the Navi 10 in the same way, connected to the same desktop I/O chip via a IF link, the console would run on both GDDR6 and DDR4. Maybe 8GB and 8GB.
Now you only mass produce the same Navi 10 chip to supply consoles and dgpus.

 

[DrMrLordX]

Why?

Well, for starters, AMD has already got Picasso on the roadmap, on 12nm. Nothing has changed there. I expect Picasso to be Raven Ridge all over again. What you are suggesting may be an option for their 7nm APUs, but for Picasso and that generation, they won't do it.

AMD has only had one post-Construction core APU that deviates from the Raven Ridge model, and that is Kabylake-G. They demonstrated some time ago that they can run Vega + HBM connected via EMIB in a fairly low-power package. So they can easily do the same via IF/Interposer. It lets AMD neatly sidestep DDR4 bandwidth restrictions, so they can host 20 or more CUs with a 4 GB HBM stack as a part of the GPU. The only downside is that such a package is not 100% compatible with the SVM model they started with Carrizo/Bristol Ridge. That's only a downside for a small group of people, though. For them, there will be Picasso.

For their 7nm apus, who knows? Those are not due until 2020. They may follow the Raven Ridge/Picasso model to keep die counts down (a -G part with separate GPU/HBM stack would be three dice; CPU chiplet, GPU/HBM chiplet, and I/O controller). And since said 7nm APU would then be a custom part, they might forgo the I/O chiplet altogether and build those functions back into the main die. Or they might not. We will see.

 

[Shivansps]

Or the 12nm Picasso is just Athlons and maybe the lowest Ryzen APU, the R3 3200.
I would expect AMD to do that. But i would not be worried too much about picasso, its something to plug the holes until the new APUs arrive.

 

[DrMrLordX]

Picasso should be:

1). Lots of mobile parts
2). All the Athlon200GE/220GE/240GE replacement parts
3). Replacement parts for 2200G/2400G

Anything above that, if AMD chooses to make an APU of some kind, will probably be a Zen2-based knockoff of Kabylake-G.