Speculation: Ryzen 4000 series/Zen 3

[amrnuke]

From what has been told/alluded to/hinted at, AMD intends to release Ryzen 4000 CPUs in Q3/Q4 2020 however we haven't had a single leak so far. It's kinda unusual — looks like motherboard vendors haven't received any samples yet.

Any ideas why? Are they postponing the release or just not rushing the things in order to sell the stock? Even now we still have tons of unsold 1st/2nd gen Ryzen CPUs.

I think architecturally the chips are going to be more similar to Zen 2, than Zen 2 was to Zen+. As a result, it may be the case that AMD hasn't released to motherboard vendors because they don't expect major issues; or alternatively, what I think is the case, is that AMD have told mobo vendors, et al, to lock things down and stop benchmarking ES and pre-release CPUs on benchmarks with public databases.

 

[itsmydamnation]

I think architecturally the chips are going to be more similar to Zen 2, than Zen 2 was to Zen+.

AMD verbiage particularly Forst Norrods has been the exact opposite. Zen2 was a more "minor"update and Zen3 is a bigger update. I dont trust any of the rumors but what we have so far is:

8 Core CCX ( amd presentation)
big (50%) L1 bandwidth increase (rumor)
4x 256 FMA (rumor)
big FPU performance incerease (rumor)
3d stacking of SRAM ( few rumors , new patients)

We haven't heard any rumors on the front-end/INT core but even if you just assume only evolutionary changes, that still better /bigger branch/fetch/predict/retire/PRF etc .

assuming all are true holistically that a very big uplift in the Core to memory sub system/a cache design. The requirements of the IF/cache coherency will change massively if AMD stack stacking 100's/1000's of MB of SRAM on each CCD.

 

[Olikan]

AMD verbiage particularly Forst Norrods has been the exact opposite. Zen2 was a more "minor"update and Zen3 is a bigger update. I dont trust any of the rumors but what we have so far is:

8 Core CCX ( amd presentation)
big (50%) L1 bandwidth increase (rumor)
4x 256 FMA (rumor)
big FPU performance incerease (rumor)
3d stacking of SRAM ( few rumors , new patients)

We haven't heard any rumors on the front-end/INT core but even if you just assume only evolutionary changes, that still better /bigger branch/fetch/predict/retire/PRF etc .

assuming all are true holistically that a very big uplift in the Core to memory sub system/a cache design. The requirements of the IF/cache coherency will change massively if AMD stack stacking 100's/1000's of MB of SRAM on each CCD.

Also this, on linux zen3 patch: "Future SMCA systems may see a new version of the Load Store unit bank type as indicated by its McaType value."

[PATCH 0/5] MCA and EDAC updates for AMD Family 19h

lore.kernel.org

 

[uzzi38]

I think architecturally the chips are going to be more similar to Zen 2, than Zen 2 was to Zen+. As a result, it may be the case that AMD hasn't released to motherboard vendors because they don't expect major issues; or alternatively, what I think is the case, is that AMD have told mobo vendors, et al, to lock things down and stop benchmarking ES and pre-release CPUs on benchmarks with public databases.

If you mean in terms of stuff like chipets etc, you're correct.

But in pretty much every other way, Zen 3 is a much larger departure than Zen 2 was from Zen. On a floormap level, Zen 2 and Zen have a lot of similarities.

And you really doubt how secretive AMD are being. If anything, and I mean ANYTHING that described Zen 3 properly were to go public right now, heads would roll.

 

[Ajay]

AMD has already said that Zen3 will be a significant departure from Zen2 and hinted that the performance increase will be 20% or more.

I don't recall seeing anything from AMD on future Zen3 performance. It is highly doubtful that they would specify a performance number anyway.

 

[uzzi38]

I don't recall seeing anything from AMD on future Zen3 performance. It is highly doubtful that they would specify a performance number anyway.

They didn't give a performance figure, no. But nothing's wrong with hoping for about a 20% uplift overall.

But they did definitely say that Zen 3 is a major architectural upgrade.

 

[amrnuke]

If you mean in terms of stuff like chipets etc, you're correct.

But in pretty much every other way, Zen 3 is a much larger departure than Zen 2 was from Zen. On a floormap level, Zen 2 and Zen have a lot of similarities.

And you really doubt how secretive AMD are being. If anything, and I mean ANYTHING that described Zen 3 properly were to go public right now, heads would roll.

I meant chiplets and IOD. On each chiplet things sound to be changing massively, but the question for me is - if the IOD remains essentially the same, what does it matter what's going on "behind" the IOD, at least as far as the motherboard is concerned? I think Windows scheduling might need an update if the CCX changes from 4 to 8, but that's no harder than what has already been done from a scheduler standpoint with the update from Zen+ to Zen2.

As for secretive, I agree. I think they have just locked stuff down entirely and anyone receiving samples is being told that if anything leaks, Lisa Su herself will personally sterilize the person responsible.

Which is good, because I personally think Zen3 is going to be HUGE. AMD know the faults in the chiplet + IOD configuration and have already worked to correct some of that with huge L3$. And they know where the next work needs to be done to reduce latency. And combining that with 7nm+ EUV they are going to have a great time.

 

[DisEnchantment]

Also this, on linux zen3 patch: "Future SMCA systems may see a new version of the Load Store unit bank type as indicated by its McaType value."

[PATCH 0/5] MCA and EDAC updates for AMD Family 19h

lore.kernel.org

And you really doubt how secretive AMD are being. If anything, and I mean ANYTHING that described Zen 3 properly were to go public right now, heads would roll.

Funny thing about the patch, it has no Model numbers . They are really tightening the lids on this one.

Or perhaps this time around they plan it this way?

 

[uzzi38]

I meant chiplets and IOD. On each chiplet things sound to be changing massively, but the question for me is - if the IOD remains essentially the same, what does it matter what's going on "behind" the IOD, at least as far as the motherboard is concerned? I think Windows scheduling might need an update if the CCX changes from 4 to 8, but that's no harder than what has already been done from a scheduler standpoint with the update from Zen+ to Zen2.

It already is.

https://www.reddit.com/r/Amd/comments/ex9nip

 

[amrnuke]

It already is.

https://www.reddit.com/r/Amd/comments/ex9nip

While that would be good news, the title says "might be" and the post says "mostly my speculation", "this all is just speculation", "if my hypothesis is true", "but still, just speculations for now."

 

[uzzi38]

While that would be good news, the title says "might be" and the post says "mostly my speculation", "this all is just speculation", "if my hypothesis is true", "but still, just speculations for now."

Yes, but you can be certain it'll happen purely because of the possibility of Alder Lake-S and Sapphire Rapids, the both of which are likely to be similar to Zen 3.

 

[inf64]

Let's recap what we know so far about Zen3:

1)Zen3 is a new uarchitecture , just like Zen1 was. Zen2 was not a new uarchitecture jump as per Forrest Norrod's statements.

AMD stated that they had a 40% IPC goal when it came to next gen core(Zen1) and it's reasonable to assume they had a similar goal set for the next new core (Zen3). They did overshoot the 40% target with Zen1, the IPC jump was >50% vs the last gen of Bulldozer (Carrizo).
If Zen1 was baseline 1 in IPC, this goal would put Zen3 at 1.4. That would mean that Zen3 would need to have 1.4/1.15/1.035~=1.176 or 17.6% higher IPC ( on average) Vs Zen2 core which is perfectly in line with Norrod's comments. My expectation is between 17 and 20% higher IPC vs Zen2.

2) Zen3 will have new CCX topology and *possibly* HBM memory. Also it is going to likely have redesigned pipeline, lower instruction latencies, reworked cache structure(faster, bigger L1s and bigger, faster and fully shared L3 cache) and *rumored* reworked FP units.Zen3 should have lower memory latency due to mentioned cache/CCX changes which will drastically improve performance in memory sensitive workloads.

3) Zen3 will be built on 7nm+ node that will have higher density, better power/mm^2 and higher fpeak. AMD could cram more cores (if needed) and remain within similar power brackets.

AMD knew what they NEED to hit IPC wise with Zen2,Zen3 and Zen4 if they wanted to be competitive Vs Willow Cove and Golden Cove cores. I think they will hit the targets easily, whether intel will hit theirs remains to be seen.

 

[Ajay]

There is no chance Zen3 will use HBM memory.

 

[lobz]

There is no chance Zen3 will use HBM memory.

also no chance for 40% IPC increase or anything in that ballpark 18-20% would be pretty nice!

 

[uzzi38]

also no chance for 40% IPC increase or anything in that ballpark 18-20% would be pretty nice!

You should read what he wrote again. He said 40% over Zen, not 40% over Zen 2.

 

[DisEnchantment]

I can't seem to recollect who it was that made a good description of how stalling can get very pronounced with high clock designs. Be it waiting on the memory subsystem, the cache subsystem and so on.
Zen Core with a much lower latency DDR5 kits and additional control channels would be a big step already. AMD mention even though memory latency can be reduced there is always a back down period before the memory is able to handle new requests on the same channel. DDR5 attemps to work around and other problems with more channels . Removed Zen4 stuffs. A reduction of inter core and cache latency etc would bring substantial benefits before touching other more radical architecture upgrades.
Add some ~200 megahertz to that and we would be seeing some decent gains.

 

[Ajay]

I can't seem to recollect who it was that made a good description of how stalling can get very pronounced with high clock designs. Be it waiting on the memory subsystem, the cache subsystem and so on.

Higher clocked designs usually have longer pipelines which really take a hit on mispredictions, not on stalls (save current state and run another thread). If the data or instructions for the other thread aren't in cache, then the second thread will stall also (SMT-2).

 

[inf64]

also no chance for 40% IPC increase or anything in that ballpark 18-20% would be pretty nice!

I clearly stated that 40% IPC goal is Vs Zen1 core. Please re-read my comment one more time.

 

[Richie Rich]

I clearly stated that 40% IPC goal is Vs Zen1 core. Please re-read my comment one more time.

This means Zen3 needs to bring +22% IPC over Zen2. IMHO this is still achievable target with "completely new uarch" as Norrod said.

 

[inf64]

This means Zen3 needs to bring +22% IPC over Zen2. IMHO this is still achievable target with "completely new uarch" as Norrod said.

I already did the back of the envelope math in my post, Zen3 needs 17.6% IPC jump over Zen2 to achieve that.

 

[amrnuke]

This means Zen3 needs to bring +22% IPC over Zen2. IMHO this is still achievable target with "completely new uarch" as Norrod said.

How did you get 22%?

Also, FYI, your signature is misleading. "fastest clocked Ryzen beaten by iPhone CPU" is at best an incomplete statement, because the fastest clock Intel is also beaten by iPhone CPU. And also, they are beaten by the iPhone CPU in simple, single-core tests, and so your statement is misleading because it implies it is a global truth.

You of course know that the Ryzen and Intel chips beat the A13 in CB20, wPrime, Blender, Corona, KeyShot, light baking, software compilation, Tensorflow, Euler3D, DigiCortex, Excel, Photoshop, 3dF Zephyr, OCR, VMWare virtualization, MySQL, Java, WinRAR, 7zip, Veracrypt, HEVC, AVC encoding, ACO, Battlefield V, Civ VI, Far Cry 5, Metro Exodus, Rage 2, Sekiro, SotTR, The Witcher 3, Wolfenstein II, WCG, Rosetta@Home, and dozens of others.

Again, just FYI. It might be a better and more truthful look, if you consider adjusting it to include other benchmarks where Intel and Ryzen chips destroy the iPhone CPU, or at least put "fastest clocked Ryzen and Intel CPUs that are designed for multitasking and parallel processing are beaten by an iPhone CPU that is geared toward single-thread performance".

 

[lobz]

You should read what he wrote again. He said 40% over Zen, not 40% over Zen 2.

sorry about that then

 

[lobz]

I clearly stated that 40% IPC goal is Vs Zen1 core. Please re-read my comment one more time.

sorry man, I misunderstood you

 

[DisEnchantment]

Also, FYI, your signature is misleading. "fastest clocked Ryzen beaten by iPhone CPU" is at best an incomplete statement, because the fastest clock Intel is also beaten by iPhone CPU. And also, they are beaten by the iPhone CPU in simple, single-core tests, and so your statement is misleading because it implies it is a global truth.

You of course know that the Ryzen and Intel chips beat the A13 in CB20, wPrime, Blender, Corona, KeyShot, light baking, software compilation, Tensorflow, Euler3D, DigiCortex, Excel, Photoshop, 3dF Zephyr, OCR, VMWare virtualization, MySQL, Java, WinRAR, 7zip, Veracrypt, HEVC, AVC encoding, ACO, Battlefield V, Civ VI, Far Cry 5, Metro Exodus, Rage 2, Sekiro, SotTR, The Witcher 3, Wolfenstein II, WCG, Rosetta@Home, and dozens of others.

Lets avoid this on this thread.

 

[Richie Rich]

How did you get 22%?

Zen2 jump was +15% so:
1.15 x 1.22 = 1.40 (+40%)
1.15 x 1.176 = 1.35 (+35%)... I guess you calculate also with Zen1+ jump.

You of course know that the Ryzen and Intel chips beat the A13 in CB20, wPrime, Blender, Corona, KeyShot, light baking, software compilation, Tensorflow, Euler3D, DigiCortex, Excel, Photoshop, 3dF Zephyr, OCR, VMWare virtualization, MySQL, Java, WinRAR, 7zip, Veracrypt, HEVC, AVC encoding, ....

It's off topic here, but you are welcome to discuss A13 performance in its thread here. At page 17 and 18 there are SPEC bench and also Geek Bench 5.1 comparison. For example A13 in Povray benchmark is +25% faster which means 2.2x higher IPC (+121%) so Blender might be very similar to it. It's kind of unfair because A13 core has double transistors, 4x more L1$ and 16x more L2$ than Zen2 core. Different core for different market.