Speculation: Ryzen 4000 series/Zen 3

Page 208 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
Ahhh you mean that Zen 3 would have a higher IPC?
Like this:
With that said, AMD is all we got at the moment for desktop. But team x86 is going to have to shape up soon, or else there will be competitive derivates of ARM incoming. I'm quite sure.
Well if you had just inserted the same image not just 2 times but a 3rd time as well, maybe you would have understood the difference between IPC and ST performance, which he wrote.
 
  • Like
Reactions: Tlh97

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
As far as speculation goes for Zen 3, with the move to an 8 core CCX with shared L3$, this seems to obviate the idea of a 4 core 5300X or 5500. I do, however, think that a 3xxxXT might be a good solution for those chips to bring them up to date.

Given the 4xxx APU layouts, I do wonder if Zen 3 APUs are really going to see as much of a benefit, or if they'll essentially be 4xxxXT chips at that point?

They likely will still have 4 core parts. Whether or not they will have the full cache? That is another story. Maybe the 5100 will have half the cache and the 5300 will have the full cache.
 

.vodka

Golden Member
Dec 5, 2014
1,203
1,537
136
Eh? I thought AMD said there were zero changes to the memory controller. Works be nice if there was improvement here.

The current I/O die can already do >5GHz memory clocks if needed. It's not the bottleneck.

Fabric clock tied 1:1 to mclk and uclk is, that's what got a ~100MHz uplift through refinements. Faster fclk is always welcome!

So, if you get a good sample, it'll do 2000MHz fclk just like a good 3rd gen sample can do 1900MHz fclk.

Mmm there's some chats on twitter about Zen3 having a much redesigned pipeline.
Very interesting. Wanna know the details :)

That was a very interesting thread.

Can't wait to read the architecture deep dive with a nice cup of coffee.


-------------------------------------




5950x GB5. DDR4-2400

~2022-2024 single core
~17400-17800 multi core

That is an absolutely ridiculous single core result. This thing is built on the same f*** node as the 3rd gen XT parts?? F*** hell.

A good 3900x-3950x will do 1350-1400 points.

+49.7%???

Either AMD is sandbagging with that +19% IPC increase, or there's something fishy going on.
 
Last edited:

Hitman928

Diamond Member
Apr 15, 2012
5,182
7,633
136
Those gb5 results seem rather fishy. It being identified as a iMacPro device and running at 6 GHz doesn't seem legit but maybe gb5 is just not reading the system accurately at all. I'm inclined to believe it's just not a legitimate score though. We'll find out soon enough.
 
  • Like
Reactions: Tlh97 and coercitiv

.vodka

Golden Member
Dec 5, 2014
1,203
1,537
136
Maybe the 5950x holds the max boost frequency (4.9ghz) much better?

Maybe. But then, both the 3900x and 3950x are specified for ~4.65GHz and ~4.75GHz max boost which they can't really hold, more like ~4.5GHz steady

+200-250MHz holding steady near 4.9GHz can't really explain the remaining uplift from the provided +19% IPC figure

~1350 * 1.19 = 1607 points

~2022/1607 = +25%

5950X processor_frequency .gb5 :
~6022 MHz

3950X processor_frequency . gb5:
~4600 MHz

:fearscream:

Oh, you're right. The 5950x eng sample gb5 does report ~6GHz

Hmm, I can't imagine a mac on LN2, can these boards even overclock? lol, this is definitely fishy, but then maybe GB5 can't quite read Zen3 parts correctly yet

Could this be a hackintosh? There are no AM4 macs out there IIRC...

----------------------------------------------

That sample's part number states 4.6GHz as the boost frequency...4.6GHz + 25% = 5.75GHz

Well, that's closer.

3800x @ 5.85GHz does 1725 single core, assuming linear scaling (which it isn't) @ 5.75GHz that would be 1695 points, +19% for IPC = 2017.5 points which is dead on to 2022 points, but then the sample is running 300MHz faster... need more coffee

napkin math says it mostly adds up... It does seem to be an overclocked 5950x sample.

Someone's having fun!
 
Last edited:

HurleyBird

Platinum Member
Apr 22, 2003
2,670
1,250
136
napkin math says it mostly adds up... It does seem to be an overclocked 5950x sample.

Might also explain the low memory clocks, since reducing how hard the memory controller has to work should improve stability.

Of course, it's possible things aren't being reported correctly, but I think you're right. Linear scaling would put a 4.9 GHz Zen 3 core at 1651 points then, but scaling isn't going to linear and you'll gain points from faster memory. I'm guessing in the range of 1700-1750 for a stock 5950X.
 
  • Like
Reactions: Tlh97 and Carfax83

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
What Mitigation you are talking about? 20% IPC increase is enough "mitigation"

Actually that's what I was referring to, though I was expecting more like +15% IPC and MT clockspeed increases courtesy of N7+. Now I'm not sure where AMD's custom N7 process lies in TSMC's hierarchy. Maybe N7+ would actually perform worse?
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
The current I/O die can already do >5GHz memory clocks if needed. It's not the bottleneck.

Fabric clock tied 1:1 to mclk and uclk is, that's what got a ~100MHz uplift through refinements. Faster fclk is always welcome!

So, if you get a good sample, it'll do 2000MHz fclk just like a good 3rd gen sample can do 1900MHz fclk.



That was a very interesting thread.

Can't wait to read the architecture deep dive with a nice cup of coffee.


-------------------------------------




5950x GB5. DDR4-2400

~2022-2024 single core
~17400-17800 multi core

That is an absolutely ridiculous single core result. This thing is built on the same f*** node as the 3rd gen XT parts?? F*** hell.

A good 3900x-3950x will do 1350-1400 points.

+49.7%???

Either AMD is sandbagging with that +19% IPC increase, or there's something fishy going on.
Ryzen was never particularly good in GeekBench, so could be that its perf increase is like CS:GO vs Battlefield (~50% vs ~5%). OK maybe not 50% but it seems to be affected by the improvements a lot heavier than the 19% average.
 

Det0x

Golden Member
Sep 11, 2014
1,027
2,953
136

Something strange with these scores, why do i match the "5950X @ 6ghz" multicore score with my PBO 3950x on ambient cooling ?

1439
Single-Core Score

17206
Multi-Core Score


The 5950x only have one core @ 6 ghz ?
But it should still beat my 3950x multicore score by around 19% thanks to the IPC gain, or is there really zero IPC gain to be had in the geekbench 5 benchmarks (?)
 
Last edited:

.vodka

Golden Member
Dec 5, 2014
1,203
1,537
136
Something strange with these scores, why do i match the multicore score with my PBO 3950x on ambient cooling ?

1439
Single-Core Score

17206
Multi-Core Score


You're running DDR4-3666, probably with decent timings. That 5950x sample is running DDR4-2400 with probably vomit inducing timings... On top of that, it's clocked @ 6GHz.

In other words, it'd take you less time to go shopping for groceries than for that CPU to get what it needs from memory to work with... apart from the fabric clock being ~1.2GHz, which caps performance for the entire chip. Despite having an 8 core CCX now with a 32MB L3 cache mitigating these kind of low performance memory situations, low fabric clock can't really be dealt with.

It'd probably be a very different picture if it wasn't hamstrung by memory+fabric performance and all those 6GHz cores could better stretch their legs.

Whoever is playing with that LN2 pot was likely just aiming for single core testing
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
Maybe. But then, both the 3900x and 3950x are specified for ~4.65GHz and ~4.75GHz max boost which they can't really hold, more like ~4.5GHz steady

+200-250MHz holding steady near 4.9GHz can't really explain the remaining uplift from the provided +19% IPC figure

~1350 * 1.19 = 1607 points

~2022/1607 = +25%



Oh, you're right. The 5950x eng sample gb5 does report ~6GHz

Hmm, I can't imagine a mac on LN2, can these boards even overclock? lol, this is definitely fishy, but then maybe GB5 can't quite read Zen3 parts correctly yet

Could this be a hackintosh? There are no AM4 macs out there IIRC...

----------------------------------------------

That sample's part number states 4.6GHz as the boost frequency...4.6GHz + 25% = 5.75GHz

Well, that's closer.

3800x @ 5.85GHz does 1725 single core, assuming linear scaling (which it isn't) @ 5.75GHz that would be 1695 points, +19% for IPC = 2017.5 points which is dead on to 2022 points, but then the sample is running 300MHz faster... need more coffee

napkin math says it mostly adds up... It does seem to be an overclocked 5950x sample.

Someone's having fun!

I usually get around 1379 SC on my 3900X.

The chip was overclocked to 6 Ghz. Pretty amazing if you ask me.
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
You're running DDR4-3666, probably with decent timings. That 5950x sample is running DDR4-2400 with probably vomit inducing timings... On top of that, it's clocked @ 6GHz.

In other words, it'd take you less time to go shopping for groceries than for that CPU to get what it needs from memory to work with... apart from the fabric clock being ~1.2GHz, which caps performance for the entire chip. Despite having an 8 core CCX now with a 32MB L3 cache mitigating these kind of low performance memory situations, low fabric clock can't really be dealt with.

It'd probably be a very different picture if it wasn't hamstrung by memory+fabric performance and all those 6GHz cores could better stretch their legs.

Whoever is playing with that LN2 pot was likely just aiming for single core testing

Yep, they likely overclocked a single core. It may not even be on LN2.

The 5xxx series presents some very unique opportunities for overclocking.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
Something strange with these scores, why do i match the "5950X @ 6ghz" multicore score with my PBO 3950x on ambient cooling ?

1439
Single-Core Score

17206
Multi-Core Score


The 5950x only have one core @ 6 ghz ?
But it should still beat my 3950x multicore score by around 19% thanks to the IPC gain, or is there really zero IPC gain to be had in the geekbench 5 benchmarks (?)
Surely it's tuned for st score here, these chips operating at close to space temps don't necessarily work / function the way they normally do
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
That's some real dangerous talk here. If this was Path of Exile, I'd say your comment has a 13% chance to attract monsters from beyond.

In hindsight this estimate seems to have been a "lowball" to put it in a manner that someone else could similarly describe as just a bit of an understatement.

Just create a thread and title it: "ARM vs x86." Or, "ARM vs The World." Or, something along those lines. That should be all encompassing, I think?

Get some old DEC engineers to StrongARM him into it and I think we'll have established a decent containment zone.
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
The 5600X is an interesting chip now that I think about it. They increased the boost clocks, very slightly reduced the base clocks, and dropped the TDP from 95W to 65W vs. the XT.
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
That would be a decent clock jump if we were to compare with 4600U which has 2.1 base and 4.0 turbo. Roughly 25% ST performance jump over Renoir.

At LEAST 25%. The unified cache will help Cezanne quite a bit, and AMD could very well increase the amount of cache.

I hope we get a bit of surprise with Zen 3 multicore performance. AMD has done a ton of work on power efficiency, and I am curious to see where the multicore boost lands.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
That should manage to keep it on par and ahead of Tiger Lake for their 4 and 6 core SKUs. Just the CCX reorganization should make a big difference for performance for effectively doubling the L3 visible to each core, which is far more importantat 4/8 MB than the 16/32 of desktop.