Zen 2 Review Summary

[juanrga]

Gamers Nexus saw 2% or less gains with their update. It's just that a couple reviewers like Anandtech were really pushed off the mark.

Did they retest power consumption and temps with the new BIOS?

 

[Dan_D]

Did they retest power consumption and temps with the new BIOS?

Temps were unchanged. I didn't check power consumption, but I have no reason to think it changed. I've never seen drastically different power consumption over a BIOS upgrade.

 

[DuronBurgerMan]

Hardware Unboxed tested 36 games at 1080p with both CPUs (9900k and 3900X) stock, and both overclocked. Used 2080 Ti for GPU.

His results were a 6% average gap across all games when both CPUs were stock, and a 5% average gap when both CPUs were overclocked (9900k @ 5GHz, and 3900X with PBO + Auto OC).

 

[Dan_D]

Hardware Unboxed tested 36 games at 1080p with both CPUs (9900k and 3900X) stock, and both overclocked. Used 2080 Ti for GPU.

His results were a 6% average gap across all games when both CPUs were stock, and a 5% average gap when both CPUs were overclocked (9900k @ 5GHz, and 3900X with PBO + Auto OC).

PBO is an automatic overclocking solution.

 

[DuronBurgerMan]

PBO is an automatic overclocking solution.

GN went into detail about the differences between "PBO" and "Auto OC" settings in this video. PBO doesn't explicitly enable the +200 MHz max boost clock that "Auto OC" does. Supposedly. So in Ryzen 3k they are two different settings that both enable some level of overclocking. The maximum available is if both are turned on.

Again, supposedly. I don't have a 3900X in my hands (yet) to test myself. The results are not exactly earth-shattering though.

 

[juanrga]

Temps were unchanged. I didn't check power consumption, but I have no reason to think it changed. I've never seen drastically different power consumption over a BIOS upgrade.

The Stilt measured 27W and 7°C deltas with a change of BIOS on his Asus board because a version was flawed.

 

[Randall Stephens]

GN went into detail about the differences between "PBO" and "Auto OC" settings in this video. PBO doesn't explicitly enable the +200 MHz max boost clock that "Auto OC" does. Supposedly. So in Ryzen 3k they are two different settings that both enable some level of overclocking. The maximum available is if both are turned on.

Again, supposedly. I don't have a 3900X in my hands (yet) to test myself. The results are not exactly earth-shattering though.

it's almost as if the marketing people and the engineers never actually sat down and discussed what was actually possible with the early silicon. perhaps the later silicon will have more potential, but I'm not holding out hope.

 

[legcramp]

Hardware Unboxed tested 36 games at 1080p with both CPUs (9900k and 3900X) stock, and both overclocked. Used 2080 Ti for GPU.

His results were a 6% average gap across all games when both CPUs were stock, and a 5% average gap when both CPUs were overclocked (9900k @ 5GHz, and 3900X with PBO + Auto OC).

So close so close! Can't wait for Intel's response and AMD's next update.

I love that this launch there is not much of "Wait for apps/games updates to increase Ryzen performance" it's just so good out of the gate.

There's literally nothing bad I can say about this launch except Intel has an edge in high fps gaming and Ryzen could use a little more headroom to overclock/tinker with.

 

[Dan_D]

The Stilt measured 27W and 7°C deltas with a change of BIOS on his Asus board because a version was flawed.

Fair enough. That can certainly happen.

 

[NWRMidnight]

So close so close! Can't wait for Intel's response and AMD's next update.

I love that this launch there is not much of "Wait for apps/games updates to increase Ryzen performance" it's just so good out of the gate.

There's literally nothing bad I can say about this launch except Intel has an edge in high fps gaming and Ryzen could use a little more headroom to overclock/tinker with.

I think we are seeing the end of over clocking as we know it now. Even from Intel, and they know it. That's why they are milking as much as they can from 14mm.

 

[DuronBurgerMan]

Did they retest power consumption and temps with the new BIOS?

He doesn't say in the video, except that he notes his 3600 test is identical (0% outside margin of error for any test), and his 3900X results increase by a maximum of 2%, but for the most part are the same also - whether he's including temps/power usage in his 'it's basically the same' statement, I don't know.

He shows a chart of boosting behavior, and an increase is noted only in the single core boosting behavior, and only by approximately +100 MHz at the most (usually less than that). I doubt this would result in +20 watts and +7C, but I don't know this. The_Stilt may have seen different results because, like Anand, his initial test configuration BIOS was more off the mark than GN's.

 

[IdiotInCharge]

I think we are seeing the end of over clocking as we know it now. Even from Intel, and they know it. That's why they are milking as much as they can from 14mm.

Ryzen has overclocked like ass over three revisions and two nodes. For their desktop parts, AMD is running them to the edge of their design.

Intel has gotten significant use out of 32nm, 22nm, and now 14nm- and while they've quite publicly stumbled with 10nm, the jury is still out on the potential of dedicated 10nm desktop parts.

But it looks like AMD is going to have to revise Zen to actually hit higher clockspeeds if that's a game that they want to compete in going forward.

 

[DuronBurgerMan]

Ryzen has overclocked like ass over three revisions and two nodes. For their desktop parts, AMD is running them to the edge of their design.

Yeah, AMD continues a trend of leaving little-to-nothing on the table. I had hoped they might change that practice this time around, but it seems to be basic company philosophy at this point, so I doubt it will change next time around.

Intel has gotten significant use out of 32nm, 22nm, and now 14nm- and while they've quite publicly stumbled with 10nm, the jury is still out on the potential of dedicated 10nm desktop parts.

The days 'shrink the node and get automagical clockspeed increases' are over. I suspect - don't know - that 10nm doesn't scale well with clockspeed (yet?), and that's why we continue to see Skylake refreshes.

But it looks like AMD is going to have to revise Zen to actually hit higher clockspeeds if that's a game that they want to compete in going forward.

Perhaps. The TSMC 7nm process is new. AMD has had a pretty consistent 'add a couple hundred MHz' trend each generation. I bet AMD and TSMC wring a little more out of 7nm over time. Not a lot. But a little. AMD may yet eek out 5GHz boost next go around. Or not. Who knows?

But ~5GHz appears to be the wall for just about everything. I remember golden sample Sandy Bridge overclocks hitting 5GHz (and 4.5 being pretty achievable for most samples, tweaked right). Things haven't changed all that much in terms of max clockspeeds since then.

 

[IdiotInCharge]

I had hoped they might change that practice this time around, but it seems to be basic company philosophy at this point, so I doubt it will change next time around.

Servers. Among other metrics, task energy matters for servers, and you'll get more work per watt if the architecture is tuned for lower clockspeeds and you shove more cores in a socket. Ryzen has used one and is now one or two Zen dies, while Epyc goes up to four, etc, but they're the same dies all around.

The days 'shrink the node and get automagical clockspeed increases' are over. I suspect - don't know - that 10nm doesn't scale well with clockspeed (yet?), and that's why we continue to see Skylake refreshes.

Intel's 10nm process has been broken and is about four years late, per their own planning. It's perhaps the most epic stumble we've seen from them. Ever.

Perhaps. The TSMC 7nm process is new. AMD has had a pretty consistent 'add a couple hundred MHz' trend each generation. I bet AMD and TSMC wring a little more out of 7nm over time. Not a lot. But a little. AMD may yet eek out 5GHz boost next go around. Or not. Who knows?

I don't see this happening soon- the only market for such CPUs, the only real market, is enthusiasts. AMD has tuned Zen for the server market as best they can, and they'd have to decide on different tuning for servers perhaps with input from TSMC for a higher clockspeed target for future Zen revisions. The point is that even if it were technically possible AMD doesn't have a financial motivation to do so.

It would take Intel bringing the heat back to the market across the board for AMD to even consider such a costly change. And that might happen, just not soon.

But ~5GHz appears to be the wall for just about everything. I remember golden sample Sandy Bridge overclocks hitting 5GHz (and 4.5 being pretty achievable for most samples, tweaked right). Things haven't changed all that much in terms of max clockspeeds since then.

I think that it's certainly possible for CPUs to be designed to run faster, but again, there's little motivation to do so as there is almost no market demand for such a product. In order to target say 6GHz as an example and do so in volume production to the point that there's a real return on investment- let alone the margins that Intel is accustomed to presenting to their stockholders- every consumer and entry-level consumer part would have to be architected to that end, and then parts could be binned for different purposes, including mobile, basic desktop, enthusiast desktop, and the consumer-socketed enterprise parts, at a minimum.

 

[NWRMidnight]

Ryzen has overclocked like ass over three revisions and two nodes. For their desktop parts, AMD is running them to the edge of their design.

Intel has gotten significant use out of 32nm, 22nm, and now 14nm- and while they've quite publicly stumbled with 10nm, the jury is still out on the potential of dedicated 10nm desktop parts.

But it looks like AMD is going to have to revise Zen to actually hit higher clockspeeds if that's a game that they want to compete in going forward.

I don't think you quite get it:. Intel is facing clock reductions with die shrink. Which means they will also have to run their new processors not based off 14mm near the edge as well to compete. As I said, overclocking as we know it is coming to an end.

 

[IdiotInCharge]

I don't think you quite get it:. Intel is facing clock reductions with die shrink. Which means they will also have to run their new processors not based off 14mm near the edge as well to compete. As I said, overclocking as we know it is coming to an end.

I don't think you get it: Intel has only produced mobile parts on 10nm. We have no idea what the desktop parts that would compete with Zen 2 on the desktop will do.

And those mobile Intel parts wipe the floor with Zen.

 

[Randall Stephens]

I don't think you get it: Intel has only produced mobile parts on 10nm. We have no idea what the desktop parts that would compete with Zen 2 on the desktop will do.

And those mobile Intel parts wipe the floor with Zen.

Looks like they lost 100Mhz of boost from the 8650U 14nm to the leaked 1065G7. For being 12nm Zen+ cores, the mobile AMD parts look fairly decent for what they are.

 

[IdiotInCharge]

Looks like they lost 100Mhz of boost from the 8650U 14nm to the leaked 1065G7. For being 12nm Zen+ cores, the mobile AMD parts look fairly decent for what they are.

Yet gained performance at those speeds- which means that they increased battery life, which for ultrabooks, matters more than outright performance. Should also decrease thermal load and thus noise, also important.

Where AMD stands is, quite literally, not available. No competitive part whatsoever.

And I say that as someone that wishes that they did.

 

[Randall Stephens]

Yet gained performance at those speeds- which means that they increased battery life, which for ultrabooks, matters more than outright performance. Should also decrease thermal load and thus noise, also important.

Where AMD stands is, quite literally, not available. No competitive part whatsoever.

And I say that as someone that wishes that they did.

Competitiveness isn't solely based on performance, cost also plays a factor, as well as availability. When will the latest intel parts actually be available for purchase? Wife needs a new ultrathin laptop

 

[IdiotInCharge]

Competitiveness isn't solely based on performance, cost also plays a factor, as well as availability. When will the latest intel parts actually be available for purchase? Wife needs a new ultrathin laptop

This month, I believe. Dell has been sampling the new XPS' about.

 

[somebrains]

AMD launch - Intel response- amd price cut is always good.

Hopefully the pattern holds.

 

[Randall Stephens]

AMD launch - Intel response- amd price cut is always good.

Hopefully the pattern holds.

There's no price cut for the new stuff. The new amd stuff will have no direct competition coming from Intel for quite a while. The Intel stuff games a bit better, but most pure gamers in this world have a cell phone and/or a console

 

[N4CR]

Yet gained performance at those speeds- which means that they increased battery life, which for ultrabooks, matters more than outright performance. Should also decrease thermal load and thus noise, also important.

Where AMD stands is, quite literally, not available. No competitive part whatsoever.

And I say that as someone that wishes that they did.

That's irrelevant to consumers because you can't market that. End of story Intel regressed on 10nm in overall performance hence it's nowhere yet.

 

[NWRMidnight]

I don't think you get it: Intel has only produced mobile parts on 10nm. We have no idea what the desktop parts that would compete with Zen 2 on the desktop will do.

And those mobile Intel parts wipe the floor with Zen.

HAHA, why is your response about mobile processors, when we are talking about desktop processors? Their 10mm parts are so good, that we will be seeing desktop processors when??? Oh wait.. not going to happen, they are skipping it and going straight to 7nm. As I said, you don't get it.

When Intel finally releases a desktop processor on 7nm, it will be lower clocks with more cores, compared to what they offer now, with little head room for overclocking, just as Ryzen. That's my call.. we will just have to wait till 2021 to find out.

 

[juanrga]

The days 'shrink the node and get automagical clockspeed increases' are over. I suspect - don't know - that 10nm doesn't scale well with clockspeed (yet?), and that's why we continue to see Skylake refreshes.

Kabylake used 14+ and Coffelake uses 14++

Cannonlake used 10 and Icelake uses 10+. Higher clockspeed would be achieved by 10++

Perhaps. The TSMC 7nm process is new. AMD has had a pretty consistent 'add a couple hundred MHz' trend each generation. I bet AMD and TSMC wring a little more out of 7nm over time. Not a lot. But a little. AMD may yet eek out 5GHz boost next go around. Or not. Who knows?

AMD has been loosing MHz pretty consistently before partially recovering.

Fastest 4C

Piledriver........... 4.2/4.3GHz
Steamroller........ 4.0/4.2GHz
Excavator.......... 3.8/4.0GHz
Zen................... 3.5/3.7GHz
Zen+................. 3.6/4.0GHz
Zen2................. 3.8/4.2GHz?

Zen3................. 4.0/4.4GHz?

But ~5GHz appears to be the wall for just about everything. I remember golden sample Sandy Bridge overclocks hitting 5GHz (and 4.5 being pretty achievable for most samples, tweaked right). Things haven't changed all that much in terms of max clockspeeds since then.

It is the frequency wall. We reach the limits of silicon, time ago.

 

[IdiotInCharge]

That's irrelevant to consumers because you can't market that.

You can't market longer battery life in a mobile device...?

End of story Intel regressed on 10nm in overall performance hence it's nowhere yet.

Quite the opposite- the story hasn't yet started.

HAHA, why is your response about mobile processors, when we are talking about desktop processors? Their 10mm parts are so good, that we will be seeing desktop processors when???

Because 10nm was brought it relative to overclocking potential or the lack thereof.

Oh wait.. not going to happen, they are skipping it and going straight to 7nm. As I said, you don't get it.

Source? I understand if you don't have one.

When Intel finally releases a desktop processor on 7nm, it will be lower clocks with more cores, compared to what they offer now, with little head room for overclocking, just as Ryzen. That's my call.. we will just have to wait till 2021 to find out.

I won't bet on Intel performing as poorly as AMD.

 

[DuronBurgerMan]

Cannonlake used 10 and Icelake uses 10+. Higher clockspeed would be achieved by 10++

Probably true. How much higher is anybody's guess. My suspicion is still probably lower than 14++.

AMD has been loosing MHz pretty consistently before partially recovering.

Provided values don't really work for this argument. Bulldozer and derived designs were designed for higher initial clockspeeds than Zen. Steamroller and Excavator did not have 'enthusiast' FX branded parts, and were thus not directly comparable to Piledriver in this respect.

Being a clean sheet design (more or less), Zen shares little with Bulldozer, and so not comparable in the same way contemporary Intel parts were. Zen has consistently gained clocks with each iteration, I don't expect that this will stop until we hit the ~5GHz silicon wall, or until Zen derivatives are replaced with another clean sheet design.

Fastest 4C

This is disingenuous, juanrga. Up until now in this thread the metric being used for max clocks has been max boost clock (1C-2C). You change the metric being used to improve the perception of your position. Bad form, sir. You know better!

Max Boost, 1C-2C. Zen.

Zen: 4.1 (XFR), 4.2 (TR)
Zen+: 4.35, 4.4 (TR)
Zen 2: 4.6 (July), 4.7 (Sept). TR: ???
Zen 3: ??? 4.9-5GHz maybe?

It is the frequency wall. We reach the limits of silicon, time ago.

True.

 

[DuronBurgerMan]

You can't market longer battery life in a mobile device...?

Indeed. That was a silly argument on his part (not yours). However, I will note that large-scale business laptop purchasing decisions are as much price as battery life. Both matter. I suspect on this front, Intel has a superior battery life argument and will retain this even if Zen 2 mobile products show up, but Zen 2 is quite efficient and it may be that AMD can make a "good enough battery life for the money" argument. Right now, Zen/Zen+ products aren't really doing that, IMHO.

Quite the opposite- the story hasn't yet started.

As I mentioned to juanrga, I think it will be quite some time before Intel 10nm starts showing similar clockspeeds to 14nm. However, with the massive IPC increases, it doesn't necessarily have to. There's a point where clockspeeds on 10nm will be good enough that superior IPC can make up the difference and deliver a faster overall product. Intel surely knows better than we do when that moment will come. Probably not in 2020 (or at least not first half of 2020) because another Skylake refresh is planned for that. But after that? Probably.

Source? I understand if you don't have one.

Yeah I don't buy his argument either. I think we will see 10nm desktop parts. Just not this year, and not until at least 2H 2020 (but more likely 2021). Until then, 10nm mobile products will be excellent performers in that space.

I won't bet on Intel performing as poorly as AMD.

Intel has a history of leaving a decent enough amount of overclocking room on their CPUs. I don't see any reason for this to change. AMD has a history of taking every last MHz they can, although I give them credit for being very clever about how they do this. We can make decent arguments both ways as to which approach is better.

 

[NWRMidnight]

Source? I understand if you don't have one.

I understand you are oblivious to reality and need people to source information that has been top news for a few months now, that you chose to ignore:

https://www.theinquirer.net/inquirer/news/3074879/intel-roadmap-leak-desktop-core-14nm-two-years

https://www.extremetech.com/computi...unch-7nm-chips-in-2021-ice-lake-ships-in-june

https://www.google.com/amp/s/www.te...tel-roadmap-shows-10nm-desktop-cpus-wont.html

https://www.google.com/amp/s/wccfte...14nm-comet-lake-10nm-ice-lake-tiger-lake/amp/

The last link talks about 2022 for 10nm LGA parts... Which implies its even worse...

There are many more articles out there... But Intel's own road map tells all.

 

[DuronBurgerMan]

I understand you are oblivious to reality and need people to source information that has been top news for a few months now, that you chose to ignore:

https://www.theinquirer.net/inquirer/news/3074879/intel-roadmap-leak-desktop-core-14nm-two-years

https://www.extremetech.com/computi...unch-7nm-chips-in-2021-ice-lake-ships-in-june

https://www.google.com/amp/s/www.te...tel-roadmap-shows-10nm-desktop-cpus-wont.html

https://www.google.com/amp/s/wccfte...14nm-comet-lake-10nm-ice-lake-tiger-lake/amp/

The last link talks about 2022 for 10nm LGA parts... Which implies its even worse...

There are many more articles out there... But Intel's own road map tells all.

Not sure how much I trust all the 'leaked' stuff. But in any event doesn't tell us much we don't already know. We'll get at least one more refresh of Skylake on 14nm (number of pluses TBD). After that, we'll probably get 10nm desktop parts. At what clockspeed... who knows?

 

[NWRMidnight]

Not sure how much I trust all the 'leaked' stuff. But in any event doesn't tell us much we don't already know. We'll get at least one more refresh of Skylake on 14nm (number of pluses TBD). After that, we'll probably get 10nm desktop parts. At what clockspeed... who knows?

What we already know is nothing from intel till late 2020 (big maybe) or 2021. Nearly all tech sites have said that intel is skipping 10nm for desktop and going to 7nm due to their issues with 10nm. But time will tell.

 

[juanrga]

Probably true. How much higher is anybody's guess. My suspicion is still probably lower than 14++.

The image posted gives some hint.

Provided values don't really work for this argument. Bulldozer and derived designs were designed for higher initial clockspeeds than Zen. Steamroller and Excavator did not have 'enthusiast' FX branded parts, and were thus not directly comparable to Piledriver in this respect.

Steamroller and Excavator FX branded parts were cancelled because the nodes used couldn't hit same clocks than the 32SOI node used by Piledriver. That is why clocks reduced, as I showed.

This is disingenuous, juanrga. Up until now in this thread the metric being used for max clocks has been max boost clock (1C-2C). You change the metric being used to improve the perception of your position. Bad form, sir. You know better!

Max Boost, 1C-2C. Zen.

Zen: 4.1 (XFR), 4.2 (TR)
Zen+: 4.35, 4.4 (TR)
Zen 2: 4.6 (July), 4.7 (Sept). TR: ???
Zen 3: ??? 4.9-5GHz maybe?

I have given base clocks and max 1C clocks. E.g. the "3.6/4.0GHz" I quoted for Zen+ is the base clock and max boost of the R5-2500X.

I am quoting the max clock achieved for equivalent SKUs. You are quoting the max clocks of each gen independently of the SKU. You are mixing node improvements with high binding for some premium products. If I was to include special binding, then I had quoted the clocks of the special binned 8C FX-9590, instead the "4.2/4.3GHz" I quoted for mainstream 4C Piledriver.

 

[DuronBurgerMan]

Steamroller and Excavator FX branded parts were cancelled because the nodes used couldn't hit same clocks than the 32SOI node used by Piledriver. That is why clocks reduced, as I showed.

They were also cancelled because the entire uarch was a big steaming pile of shit.

I have given base clocks and max 1C clocks. E.g. the "3.6/4.0GHz" I quoted for Zen+ is the base clock and max boost of the R5-2500X.

I am quoting the max clock achieved for equivalent SKUs. You are quoting the max clocks of each gen independently of the SKU. You are mixing node improvements with high binding for some premium products. If I was to include special binding, then I had quoted the clocks of the special binned 8C FX-9590, instead the "4.2/4.3GHz" I quoted for mainstream 4C Piledriver.

Equivalent SKUs? Equivalent on what? Core count? Market segmentation? Nobody here cares what a 2500X does. Even fewer care what Steamroller did. Plenty care about what a 3900X does.

I take the max values associated with a particular generation, in a particular socket (this is why I give TR values separately) - I use the same benchmark for Intel products, albeit I give a Intel credit for a higher OC ceiling as well (I'd do the same for AMD, but their max boost clock is typically at or above the max OC ceiling anyway, so it doesn't matter).

Edit: and just to provide a little evidence behind the 'nobody cares about a 2500X' statement, look at the Userbenchmark data:

https://cpu.userbenchmark.com/Compare/AMD-Ryzen-5-2500X-vs-AMD-Ryzen-9-3900X/m567224vs4044

267 benchmarks for the 2500X (out for almost a year). 1,538 for the 3900X (which has been out for only a week and a half). The Conclusion Statement at the bottom is literally "This Exists?"

2500X isn't an 'equivalent SKU' to much of anything.

 

[IdiotInCharge]

2500X isn't an 'equivalent SKU' to much of anything.

I was typing something else, but I had to look up what these things are. They look like single-CCX Zen+ dies (I thought they were just graphics-less APUs at first), which at least looks somewhat attractive compared to the lowest-end Intel parts.

Still, lack of retail availability does make any detailed comparison pretty moot.

 

[DuronBurgerMan]

I was typing something else, but I had to look up what these things are. They look like single-CCX Zen+ dies (I thought they were just graphics-less APUs at first), which at least looks somewhat attractive compared to the lowest-end Intel parts.

Still, lack of retail availability does make any detailed comparison pretty moot.

I had to look them up too. Non-APU Zen+ products with less than 6 cores... yes, they theoretically exist. But no, I don't actually know anybody who's ever had one.

 

[Ruiner]

3600 review at TPU. Negligible loss to 3600x. Almost no overclock. Roughly 15 percent faster than 1600, 10% over 2600 at 1080p.