Speculation: Ryzen 4000 series/Zen 3

[bsp2020]

There is no fixing anymore, Zen3 will have a massive IPC advantage (~20-25%) over Skylake(-X) in everything but AVX512 while clocking to almost 5Ghz for 1T workloads. Intel will have a choice of lowering the prices or doing some insane marketing shenanigans.

Or AMD will charge higher price to make more profit. No need to price their CPUs so low when Intel can't match it and there are high demand for all products from both Intel and AMD.

 

[HurleyBird]

There is no fixing anymore, Zen3 will have a massive IPC advantage (~20-25%) over Skylake(-X) in everything but AVX512 while clocking to almost 5Ghz for 1T workloads. Intel will have a choice of lowering the prices or doing some insane marketing shenanigans.

What he means is that Zen 1 benched disproportionately well in legacy CPU-Z, so they changed the benchmark after Zen 1 launch to better reflect the average performance across most other applications.

 

[inf64]

What he means is that Zen 1 benched disproportionately well in legacy CPU-Z, so they changed the benchmark after Zen 1 launch to better reflect the average performance across most other applications.

I know what he means . I'm just saying that Zen3 is now so much better IPC wise that there will be no point to modify the benchmarks.

 

[moinmoin]

Exactly. Different workloads can produce wildly different results. On my 3900x:

CBR20 - 4175 MHz
Prime95 Small FFTs- 3990 MHz (?)

It's been like that since Matisse. Summit Ridge was for more predictable.

A big change in that regard for Matisse was the much higher clock selection time granularity of UEFI CPPC2 which allows to change the frequency every 1-2ms. This means heavy instructions won't pull down light instructions as much as in the past. CBR20 apparently has more light instructions with the higher time granularity allowing it to boost higher than Prime95 Small FFTs. I expect this behavior to be improved further in future Zen gens.

 

[jamescox]

It's going to be like that no matter whether AMD stockpiles the chips or not. They can only order so many wafers over such-and-such a period of time. Best to launch with at least a limited supply to gauge interest, and immediately start taking more wafers as they become available. Which is more-or-less what they did with Matisse.

(plus dice that bin well enough for EPYC will probably wind up in Milan instead)



See @Hitman928 below



Exactly. Different workloads can produce wildly different results. On my 3900x:

CBR20 - 4175 MHz
Prime95 Small FFTs- 3990 MHz (?)

It's been like that since Matisse. Summit Ridge was for more predictable.



You are not alone. I don't think we'll see them in a 'monopoly' position so much as a 'we'll keep selling you the same crap for longer than our technology mandates we need to' position. Expect to see some variation of Zen3 on the market as AMD's main offering until 2022. Booooooring.



You have a thread for that.

From the few articles that I have read on the subject, it sounds like the cycle time for processing a 7 nm wafer is on the order of 80 days. So AMD had to be manufacturing Zen 3 die more than 3 months ago to launch in October. It was probably more than 3 months ago since the packaging and testing process takes time and they need to stockpile supply. They can’t try to fill current demand by increasing wafer starts now. If they increased wafer starts, the extra parts probably wouldn’t be available for about 4 months. AMD’s MCM strategy makes them quite a bit more flexible. I don’t know when Epyc Milan is launching, but that takes huge numbers of cpu die, so they should have been making huge numbers of them over the last few months to be able to bin and validate enough of them for an Epyc launch. They are just going to make as many cpu die as they can. All parts use the same die, so they aren’t going to have an oversupply of one die and an under-supply of another. They also have the advantage that the cpu die are quite small, so they can fit a lot of them on each wafer. They can’t launch with low supply to gauge demand and then increase production due to long lead times.

Also, AMD could stand to be in a dominate position for a much longer time, especially in the server market. The replacement cycles are much longer and people are less likely to take what they perceive as a risk. It will take a lot longer for Intel’s monopoly to dissolve. Although, at this point, I think both companies are going to have trouble against ARM unless Nvidia really screws it up. A lot of servers do not actually need much cpu performance. Storage servers are almost appliances is, and in some cases, are already being replaced with appliances (NAS devicesMans such). Also, we are starting to see some high performance ARM server chips anyway.

 

[dr1337]

Or AMD will charge higher price to make more profit.

Eh I don't think they're going to do that, not with this generation at least. I do think they might release the 5800x and 5900x first for max volume and profit. I guess they could tack on another $50 to the msrp. Regardless, it is going to be interesting how they position their current lineup against zen 3.

 

[simas]

Or AMD will charge higher price to make more profit. No need to price their CPUs so low when Intel can't match it and there are high demand for all products from both Intel and AMD.

that is my concern as well. as person who is buying for himself, there are price points (not sure of the proper term) where manufacturer starts to lose my interest and then loses my attention altogether regardless of the performance. I.e. most recent NVIDIA releases of 'flagship' product, i simply don't care at that price point even if it runs circles around around any competition while single happy birthday..

All comes to perceived value, if I spend X what I am getting for that X in areas I do care about (i.e. specific programs or specific games). for example, I will welcome more CPU power as one of my favorite games is Total War series that has huge CPU bottlenecks when large armies meet.

 

[StinkyPinky]

If those IPC gains are true it makes me want to upgrade from my 3900x. Seems excessive....but hey, I'm a tech nerd.

 

[scannall]

There is no fixing anymore, Zen3 will have a massive IPC advantage (~20-25%) over Skylake(-X) in everything but AVX512 while clocking to almost 5Ghz for 1T workloads. Intel will have a choice of lowering the prices or doing some insane marketing shenanigans.

Nah, Intel still has branding and mindshare. AMD and Ryzen are certainly doing well at rebuilding their own, but Intel still has 3-5 years breathing room. While I am fairly sure they know it, they aren't really doing much aside from patching holes in the boat when they really need to build an all new one. Will they? Won't they? Who knows, certainly none of us on this board. But it will be fun to watch in any event.

 

[amrnuke]

If those IPC gains are true it makes me want to upgrade from my 3900x. Seems excessive....but hey, I'm a tech nerd.

Eh, if you have the disposable income and your work is important, there's no reason not to.

For me, a 3600 rarely hangs on things, and only then briefly, but with the disposable income I'll probably move to an 8 or 12 core Zen3 chip. Probably get a next gen graphics card to compel me to finally upgrade to FS2020 too.

 

[Carfax83]

Not sure whether this legit, but CPU Monkey has uploaded some Zen 3 results in CBR20 for the 5950x, 5900x, 5800x and 5600x.

Cinebench R20 (Single-Core) CPU benchmark list (cpu-monkey.com) *edit* fixed link!

Assuming this is accurate, looks like 15% IPC increase for certain after you consider clockspeed enhancements. But doesn't CB use floating point? So I guess the rumors of the 50% IPC increase for floating point must be incorrect.

15% IPC is good, but I was hoping for around 20% or so.

 

[Makaveli]

Not sure whether this legit, but CPU Monkey has uploaded some Zen 3 results in CBR20 for the 5950x, 5900x, 5800x and 5600x.

Cinebench R20 (Single-Core) CPU benchmark list (cpu-monkey.com) *edit* fixed link!

Assuming this is accurate, looks like 15% IPC increase for certain after you consider clockspeed enhancements. But doesn't CB use floating point? So I guess the rumors of the 50% IPC increase for floating point must be incorrect.

15% IPC is good, but I was hoping for around 20% or so.

Since I will be going from 3800X to 5800X i'll happily take 19.29%.

 

[Carfax83]

Since I will be going from 3800X to 5800X i'll happily take 19.29%.

I suppose when you consider that Zen 3 is using an enhanced version of the same process node as Zen 2, then 15% IPC is quite good. But if rumors are true, Intel looks like they are targeting much larger IPC increases with Golden Cove and its successor. Zen 3 is inevitably going to face off against Golden Cove next year, unless AMD can get Zen 4 out by late next year as well.

 

[Hitman928]

Not sure whether this legit, but CPU Monkey has uploaded some Zen 3 results in CBR20 for the 5950x, 5900x, 5800x and 5600x.

Cinebench R20 (Single-Core) CPU benchmark list (cpu-monkey.com) *edit* fixed link!

Assuming this is accurate, looks like 15% IPC increase for certain after you consider clockspeed enhancements. But doesn't CB use floating point? So I guess the rumors of the 50% IPC increase for floating point must be incorrect.

15% IPC is good, but I was hoping for around 20% or so.

I would not trust any pre-release numbers from cpu-monkey. I am highly doubtful they got the scoop here. I think it's much more likely they are just guessing here.

 

[A///]

Nice score. Now wait for the patch to "correct" the Zen3 performance like they did with the first zen.

Why would you patch an ES?

 

[scannall]

I suppose when you consider that Zen 3 is using an enhanced version of the same process node as Zen 2, then 15% IPC is quite good. But if rumors are true, Intel looks like they are targeting much larger IPC increases with Golden Cove and its successor. Zen 3 is inevitably going to face off against Golden Cove next year, unless AMD can get Zen 4 out by late next year as well.

And that's fine. Once parts are actually in the wild I will look at them. A cold beer is worth far more than a rumor. ;-)

 

[TheGiant]

Well guys anyone Remember adored?
Let's wait for the reviews or reputable leaks

 

[Markfw]

Well guys anyone Remember adored?
Let's wait for the reviews or reputable leaks

I won't bet the farm on it, but after what AMD said and all the leaks, I really do believe we will get about 15% IPC uplift. Not sure of turbo or all core turbo, don't care. But I am pretty sure the gaming crown is going to AMD, they already had everything else.

 

[maddie]

I suppose when you consider that Zen 3 is using an enhanced version of the same process node as Zen 2, then 15% IPC is quite good. But if rumors are true, Intel looks like they are targeting much larger IPC increases with Golden Cove and its successor. Zen 3 is inevitably going to face off against Golden Cove next year, unless AMD can get Zen 4 out by late next year as well.

There buries all of the old excuses about why IPC gains were so minimal in the past decade. I remember asking if there were theoretical limits. What was it again? IPC improvement is soooo hard to achieve , so cut Intel some slack. Well here we are, back in the race.

 

[leoneazzurro]

AMD is trying to put emphasis on gaming with the Zen3 presentation:

https://www.amd.com/en/events/gaming-2020#.X3Z8AiHjtLg.twitter

They seems very sure of their product, let's hope this will be confirmed in actual tests.

 

[Nereus77]

I think people are going to have a good time testing Zen 3 to its limits. Exciting times ahead.

 

[inf64]

I would not trust any pre-release numbers from cpu-monkey. I am highly doubtful they got the scoop here. I think it's much more likely they are just guessing here.

Alderlake for desktop is very unlikey to show up in 2021 . Zen4 is Q1 2022 as how things look right now, just in time to make AMD great again with 20+ cores , DDR5 and IPC jump - all this versus hybrid AL 8+8 . Fun times!

 

[moinmoin]

Intel still has 3-5 years breathing room.

So that'd be 8-10 years from the point where its node progress showed serious cracks, and 6-8 years from the point where its competition started becoming competitive again. The time scale honestly is eye-watering mind boggling at this point, especially for happening within the otherwise fast moving tech industry. Intel is incredible lucky to be as old-established within the markets as it is.

 

[TheGiant]

There buries all of the old excuses about why IPC gains were so minimal in the past decade. I remember asking if there were theoretical limits. What was it again? IPC improvement is soooo hard to achieve , so cut Intel some slack. Well here we are, back in the race.

The excuses have a name
10nm f. Up

 

[Kedas]

I know what he means . I'm just saying that Zen3 is now so much better IPC wise that there will be no point to modify the benchmarks.

Maybe OR Zen3 doesn't register as needing a correction and we see the original score like zen1, meaning we don't see zen3 vs zen2 but with correction factor vs without correction factor....(i hope that is not true)