Speculation: Ryzen 3000 series

[LightningZ71]

I would imagine that TR may be where AMD dumps their chiplets that only have 4 working cores, or, more specifically, where they can only get 2 cores on a single CCX to work properly, but those don't have any issues with clocking. One of the issues that it APPEARS that we are seeing is the difficulty with extracting the heat from the 7nm chiplets. What if AMD focuses on clock speeds for Threadripper at the cost of maximum core count, and continues the WX line for the devices that instead focus on cores over clocks? Just an example of the idea:

TR 3900x
4 dies, 16 active cores (2 per CCX) Boost to 4.1 Ghz
TR 3920x
6 dies, 24 active cores (2 per CCX) Boost to 4.2 Ghz
TR 3950x
8 dies, 32 active cores (2 per CCX) Boost to 4.3 Ghz
TR 3900WX
4 Dies, 24 active cores (3 per CCX) Boost to 4 Ghz
TR 3920WX
6 Dies, 36 active cores (3 per CCX) Boost to 4 Ghz
TR 3950WX
8 Dies, 48 active cores (3 per CCX) Boost to 4 Ghz
TR 3990WX
8 Dies, 64 active cores (4 per CCX) Boost to 3.6 Ghz

The idea is that the lower active core count chiplets can be better binned for clocking potential, and then pushed harder for clock speed. I believe that my above boost numbers might even be rather conservative, given what we see from the Ryzen 3900. Either way, the resulting TR processors would be a significant performance uplift from the existing models in many regards, and, with the new memory controller proving to be better capable of higher throughputs, I don't see the 4 channel memory as a significant bottleneck for the 48 and 64 core products.

 

[Hans de Vries]

This image here is from 9 months ago. It was dismissed as the mother of all fakes and nobody ever looked at it again.

Now it looks more like an initial idea of an SKU list. Many of these parts are already on the market or are coming soon. The numbers as well as the frequencies have shifted somewhat. The 7nm prices became less disruptive.

The 300GE and 320GE are coming soon. The Picasso's became the 3200G and the 3400G. What is now the 3900x is here listed with a max boost of 4.65 GHz with a higher base clock corresponding to a higher TDP (125W?)

What is now the 3950X has a max boost here of 4.9 GHz (that would be with the extra 200MHz from PBO2)

Note that AMD can produce higher frequency parts by etching higher FinFets. This increases the transistor's drive and the CPU frequencies but at the cost of significantly lower yields and higher variability.

The Threadrippers as shown here would need TDP's of 250W to 400W. The last two Threadrippers would need the same 1500W aquarium cooler as Intel's "5GHz" 28 core Xeon W-3175X and shouldn't be taken seriously, maybe just a marketing response to the former.

Note also the two Renoir's with eight Zen 2 cores and 14CU / 20CU graphics. 7nm is absolutely ideal for mobile at lower voltages. Renoir uses a new FP6 package for mobile. These APU's require on package memory and they may be the first APU's to include a 4GB HBM2e stack working with a HBCC High Bandwidth Cache Controller while the HBM2e stack is packaged using TSMC's InFO-MS fine line organic substrates without the need for a silicon imposer. Effectively the 4GB would be more like 8GB in game performance.

 

[dlerious]

I think it would be on similar or same as 3000 Epyc series, SP4, or SP4r2 or something; these are non-socketed BGA embedded. If you had a 32c/64t beast, the upgrade inability with BGA boards would be a non-issue; it's going to remain capable for well over a decade, as we're seeing the end of Moore's law and diminishing returns on tech; the platform features are going to outdate before the processor.

I don't know if/when those would move to 7nm. They're maxed out at 16 cores for a 2-die package with 2.15/3 GHz speeds at 100W TDP for around $880. The upgrade inability could be a big deal for me - there might be a new spec like DDR5, PCIe 5, or a new storage spec that I need, requiring a purchase of both CPU and motherboard.

I wonder how the DIY market will be handled with BGA. Are you going to grab a CPU and motherboard and solder yourself or are you going to be stuck grabbing a pre-assembled combo from a limited SKU range from a motherboard manufacturer.

 

[dlerious]

Gamers Nexus checking effect of temps on clocks.

 

[amd6502]

This image here is from 9 months ago. It was dismissed as the mother of all fakes and nobody ever looked at it again.

Now it looks more like an initial idea of an SKU list. Many of these parts are already on the market or are coming soon. The numbers as well as the frequencies have shifted somewhat. The 7nm prices became less disruptive.

The 300GE and 320GE are coming soon. The Picasso's became the 3200G and the 3400G. What is now the 3900x is here listed with a max boost of 4.65 GHz with a higher base clock corresponding to a higher TDP (125W?)

What is now the 3950X has a max boost here of 4.9 GHz (that would be with the extra 200MHz from PBO2)

Note that AMD can produce higher frequency parts by etching higher FinFets. This increases the transistor's drive and the CPU frequencies but at the cost of significantly lower yields and higher variability.

The Threadrippers as shown here would need TDP's of 250W to 400W. The last two Threadrippers would need the same 1500W aquarium cooler as Intel's "5GHz" 28 core Xeon W-3175X and shouldn't be taken seriously, maybe just a marketing response to the former.

Note also the two Renoir's with eight Zen 2 cores and 14CU / 20CU graphics. 7nm is absolutely ideal for mobile at lower voltages. Renoir uses a new FP6 package for mobile. These APU's require on package memory and they may be the first APU's to include a 4GB HBM2e stack working with a HBCC High Bandwidth Cache Controller while the HBM2e stack is packaged using TSMC's InFO-MS fine line organic substrates without the need for a silicon imposer. Effectively the 4GB would be more like 8GB in game performance.


View attachment 8654

That's really interesting. 9 month is not that long ago tough, so it could very well be a regular fanfiction "leak". The other next likeliest possibility coudl be that it's a fanfiction or speculation leak from an AMD engineer (or even other employee like a marketing intern), based on some management rumors floating about, and facts he knew, including the capabilities of the dies.

Is raven2 (or whatever the native 2c APU confirmed to be 12nm rather than 14nm?)

Assuming it's the leak of the second kind, the potential 8c APU seems to me to likelier be 8c chiplet based, with either a GPU chiplet, or with an integrated iGPU on the IOX, with the 3600G being the die salvaged derivative of the full 20CU 3600GX. If it is this higher quality "leak" then I also look forward to what seems to be an 28nm FDSOI version of XV without iGPU. Heck, bargain-bin ($30=) cut down to ~110mm2 gpu-less excavator would be absolutely brilliant and a great thing to happen to AM4.

 

[VirtualLarry]

Heck, bargain-bin ($30=) cut down to ~110mm2 gpu-less excavator would be absolutely brilliant and a great thing to happen to AM4.

Isn't the 200GE already available in places for $40-50? WITH a decent (desktop usage, 4K UHD even) iGPU. What good would a slow-at-any-Mhz Excavator-based AM4 iGPU-less CPU do for AM4? Why? Where is the profit potential for AMD? Sounds like it would lower ASPs, not raise them.

I get that you're a Nosta "follower" / fan, but still. I don't recall anything that he's said, actually come true.

This is kind of like saying, why doesn't Intel just come out with a RTL-level version of Core2Quad (*or Sandy, perhaps), and give it to TSMC to mfg for Intel, it would be brilliant and cheap! Well, technology has moved on from there. Likewise, with Excavator.

 

[moinmoin]

Is raven2 (or whatever the native 2c APU confirmed to be 12nm rather than 14nm?)

Chances are it's the same die as the Ryzen Embedded R1000 series which is 14nm.

 

[amd6502]

Chances are it's the same die as the Ryzen Embedded R1000 series which is 14nm.

Not sure but it kind of looks that way (I was hoping it were a 12nm one).

Isn't the 200GE already available in places for $40-50? WITH a decent (desktop usage, 4K UHD even) iGPU. What good would a slow-at-any-Mhz Excavator-based AM4 iGPU-less CPU do for AM4? Why? Where is the profit potential for AMD? Sounds like it would lower ASPs, not raise them.

I get that you're a Nosta "follower" / fan, but still. I don't recall anything that he's said, actually come true.

This is kind of like saying, why doesn't Intel just come out with a RTL-level version of Core2Quad (*or Sandy, perhaps), and give it to TSMC to mfg for Intel, it would be brilliant and cheap! Well, technology has moved on from there. Likewise, with Excavator.

Both Sandy and C2Q are power hogs in today's terms, so I would say no way. C2Q also lacks the performance versus a 5ghz XV quad. If Sandy were ported to 22FDX then it could be efficient enough. But they would have to work on security fixes (possibly redesign) as well as the port which sounds pricey.

Athlon 200ge is around $57 and stock will dry out. 300ge (or whatever) should replace the Athlon around $50 (or low-mid 50s).

If it were me I'd rather have XV near 5ghz than athlon 300ge (and I would end up saving $20 on top of that). But I have old video cards laying around that are much more capable than Vega 3. If I had no cards laying around and if I were building an office machine with minor casual gaming ability I would go with the 300ge, but for a gaming build with a discrete $100-150 card I might very well pick the XV near 5ghz.

As for margins it'd be a 1 Billion transistor product at a low cost node versus a ~2.5 Billion transistor product at a medium-low cost node. I could see margins being similar for a $30 product versus $50 product, but the project cost on the CPU might be hard to justify (since the 2c/4t mobile APU is almost a must-do).

However, if the XV project could be justified (eg maybe piggy backing off of embedded x86 demand, possibly in automotive), then why not? nothing could really beat it in manufacturing cost for its useful lifespan of ~ 5+ years. more choice, and a wider price range for the whole product line should be a good thing.

Paging Nosta: did you make this "leak"?

 

[Thunder 57]

Not sure but it kind of looks that way (I was hoping it were a 12nm one).



Both Sandy and C2Q are power hogs in today's terms, so I would say no way. C2Q also lacks the performance versus a 5ghz XV quad. If Sandy were ported to 22FDX then it could be efficient enough. But they would have to work on security fixes (possibly redesign) as well as the port which sounds pricey.

Athlon 200ge is around $57 and stock will dry out. 300ge (or whatever) should replace the Athlon around $50 (or low-mid 50s).

If it were me I'd rather have XV near 5ghz than athlon 300ge (and I would end up saving $20 on top of that). But I have old video cards laying around that are much more capable than Vega 3. If I had no cards laying around and if I were building an office machine with minor casual gaming ability I would go with the 300ge, but for a gaming build with a discrete $100-150 card I might very well pick the XV near 5ghz.

As for margins it'd be a 1 Billion transistor product at a low cost node versus a ~2.5 Billion transistor product at a medium-low cost node. I could see margins being similar for a $30 product versus $50 product, but the project cost on the CPU might be hard to justify (since the 2c/4t mobile APU is almost a must-do).

However, if the XV project could be justified (eg maybe piggy backing off of embedded x86 demand, possibly in automotive), then why not? nothing could really beat it in manufacturing cost for its useful lifespan of ~ 5+ years. more choice, and a wider price range for the whole product line should be a good thing.

Paging Nosta: did you make this "leak"?

Bulldozer is dead, it's not coming back. And certainly not on some FD-SOI process. Just stop.

 

[lopri]

Even if there is no core count increase Zen2 Threadripper is going to be a significant upgrade over Zen+ Threadripper. I do think core count increase will come, but I have my doubt on a 64 core SKU.

 

[VirtualLarry]

Both Sandy and C2Q are power hogs in today's terms, so I would say no way. C2Q also lacks the performance versus a 5ghz XV quad. If Sandy were ported to 22FDX then it could be efficient enough. But they would have to work on security fixes (possibly redesign) as well as the port which sounds pricey.

I meant to add, on TSMC's current process, so a 7nm re-design (from RTL) of a Core2 / Sandy-era CPU.

 

[VirtualLarry]

However, if the XV project could be justified (eg maybe piggy backing off of embedded x86 demand, possibly in automotive), then why not? nothing could really beat it in manufacturing cost for its useful lifespan of ~ 5+ years. more choice, and a wider price range for the whole product line should be a good thing.

Paging Nosta: did you make this "leak"?

I could see it possibly happening for automotive, but not so much consumer.

 

[amd6502]

I could see it possibly happening for automotive, but not so much consumer.

With options like Risc-V it's a lot of competition. But if you're going to design a die anyways (eg for embedded) you branch it to various markets; so there would be consumer derivative as well. If they were to design it to be CCD compatible with some new AM4 graphics capable IOX then that would be even more product derivatives.

Bulldozer is dead, it's not coming back. And certainly not on some FD-SOI process. Just stop.

Probably; the leak does seem more fake than real.

 

[Slaughterem]

Even if there is no core count increase Zen2 Threadripper is going to be a significant upgrade over Zen+ Threadripper. I do think core count increase will come, but I have my doubt on a 64 core SKU.

I agree, If I was making decisions at amd I would think that 1920x 1950x 2920x and 2950x are end of sales. I would come out with 3920x 12 core 4 mem ch 64 Pcie 4 at base clock of 3.8 boost 4.6 125 watts and a price of $699. 3970x 16 core 4 mem 64 Pcie 4 at base of 3.5 boost 4.7 125 watts $899. This would provide a decent upgrade path for the 4 other processors. I would then have a 3980wx at 24 core (you know the platform) base of 3.4 boost 4.5 180 watt $1299 and 3990wx 32 core base of 3.4 boost to 4.5 at 210 watts $1799. They may want to come out with a 36 and 48 core with 6 mem channels on a new socket and IO chip. I don't believe they will make a 64 core for threadripper at this time.

 

[HurleyBird]

I don't believe they will make a 64 core for threadripper at this time.

I'm guessing that they will, but towards the end of the generation.

 

[tamz_msc]

Somebody over at Reddit discovered that the Ryzen 9 3900X will only boost to the advertised max boost clock under extremely specific circumstances while executing very basic code.

https://www.reddit.com/r/Amd/comments/cfli2n

This calls into question the rationale behind advertising these chips as being capable of hitting the boost clocks, especially on the higher end SKUs, and it looks like something that may not be fixed with a BIOS/AGESA update.

 

[inf64]

Zen3K SKUs are marketed with "Up to " X.XX boost. There is nothing wrong with that. It was always going to depend on the temperature/heat and naturally the high power instructions will result in lower boost clocks.

 

[beginner99]

Zen3K SKUs are marketed with "Up to " X.XX boost. There is nothing wrong with that. It was always going to depend on the temperature/heat and naturally the high power instructions will result in lower boost clocks

Read that thread. 4.6 ghz is only reached when running the op NOP which means exactly that. do nothing. even a simple if will reduce clocks to below 4.6 ghz. This isn't any better that intels fake tdp but at least that tdp can be fixed and it was obvious a 9900k will use a lot more power than a 7700k. This feels more like cheating.

I was puzzeled why it's barley possible to reach 4.3 ghz all core even with manual oc. I mean intel proccessors usually do all-core on the advertised turbo if you have the right cooling. With ryzen3000k it's a brick wall. It also makes no sense marketing wise. just set boost to 100mhz lower on all skus and it would probably have been fine.

 

[Drazick]

Read that thread. 4.6 ghz is only reached when running the op NOP which means exactly that. do nothing. even a simple if will reduce clocks to below 4.6 ghz. This isn't any better that intels fake tdp but at least that tdp can be fixed and it was obvious a 9900k will use a lot more power than a 7700k. This feels more like cheating.

I was puzzeled why it's barley possible to reach 4.3 ghz all core even with manual oc. I mean intel proccessors usually do all-core on the advertised turbo if you have the right cooling. With ryzen3000k it's a brick wall. It also makes no sense marketing wise. just set boost to 100mhz lower on all skus and it would probably have been fine.

You may have a different outlook on this.
AMD indeed designed the CPU to run at the highest possible speed while taking into account the real world load and temperature.
So, in comparison to Intel which uses pre defined rules (AVX for instance) AMD uses a policy which adapts according to real world data per CPU.

As users, we can not ask for more (Policy wise).
We can wish for better manufacturing, but regarding the OC policy, this is great.
AMD extracts the maximum from the core the user bought.

 

[Asterox]

This image here is from 9 months ago. It was dismissed as the mother of all fakes and nobody ever looked at it again.

Now it looks more like an initial idea of an SKU list. Many of these parts are already on the market or are coming soon. The numbers as well as the frequencies have shifted somewhat. The 7nm prices became less disruptive.

The 300GE and 320GE are coming soon. The Picasso's became the 3200G and the 3400G. What is now the 3900x is here listed with a max boost of 4.65 GHz with a higher base clock corresponding to a higher TDP (125W?)

What is now the 3950X has a max boost here of 4.9 GHz (that would be with the extra 200MHz from PBO2)

Note that AMD can produce higher frequency parts by etching higher FinFets. This increases the transistor's drive and the CPU frequencies but at the cost of significantly lower yields and higher variability.

The Threadrippers as shown here would need TDP's of 250W to 400W. The last two Threadrippers would need the same 1500W aquarium cooler as Intel's "5GHz" 28 core Xeon W-3175X and shouldn't be taken seriously, maybe just a marketing response to the former.

Note also the two Renoir's with eight Zen 2 cores and 14CU / 20CU graphics. 7nm is absolutely ideal for mobile at lower voltages. Renoir uses a new FP6 package for mobile. These APU's require on package memory and they may be the first APU's to include a 4GB HBM2e stack working with a HBCC High Bandwidth Cache Controller while the HBM2e stack is packaged using TSMC's InFO-MS fine line organic substrates without the need for a silicon imposer. Effectively the 4GB would be more like 8GB in game performance.


View attachment 8654

For me the most "deliberately" part is, on that list there is not a single one 6/12 Ryzen 3000 CPU.

It's a classic trap, "that was just AMD overhyping"(in a situation where hype is not important at all) future new CPU-s.It is nice idea, for the experiment we will hide the best seling(in future) 6/12 CPU model.

 

[Space Tyrant]

You may have a different outlook on this.
AMD indeed designed the CPU to run at the highest possible speed while taking into account the real world load and temperature.
So, in comparison to Intel which uses pre defined rules (AVX for instance) AMD uses a policy which adapts according to real world data per CPU.

As users, we can not ask for more (Policy wise).
We can wish for better manufacturing, but regarding the OC policy, this is great.
AMD extracts the maximum from the core the user bought.

I don't think there's anything wrong with the OC policy. I think the problem (based on several comments from disappointed users) is the highlighting of a number that isn't relevant to most users actual workloads.

 

[ClockHound]

Somebody over at Reddit discovered that the Ryzen 9 3900X will only boost to the advertised max boost clock under extremely specific circumstances while executing very basic code.

https://www.reddit.com/r/Amd/comments/cfli2n

This calls into question the rationale behind advertising these chips as being capable of hitting the boost clocks, especially on the higher end SKUs, and it looks like something that may not be fixed with a BIOS/AGESA update.

Not at all. AMD optimized the Z2 boost to finish an infinite loop in the shortest amount of infinite time possible. That's some long term customer-focus right there. ;-)

 

[DrMrLordX]

Not at all. AMD optimized the Z2 boost to finish an infinite loop in the shortest amount of infinite time possible. That's some long term customer-focus right there. ;-)

I know when I choose CPUs, I go for the one that can retire the most NOPs per second and then let er rip ad infinitum. Good job AMD!

 

[misuspita]

So, I saw that picture from 9 months ago that Hans De Vries posted in these threads and in there appears to be a 3600G in there. Do you think there is a possibility to be real?

 

[NostaSeronx]

Do you think there is a possibility to be real?

Yes.
"Most recent tapeout of Large APU Chipset involving 600+ engineering resources across 5 time zones, 15B+ active transistors in 7nm tech node."

Apple A12X => 10B at 122 mm squared on 7nm
Large APU on 7nm => 15B+ at 122 mm squared < x mm squared < 398 mm squared
Centriq 2400 => 18B at 398 mm squared on 10nm

It should be around ~200 mm squared hopefully.