AMD Q4 Earnings Report & Zen 2 Will Have Spectre Fix

[LAwLz]

1 hour ago, Trixanity said:

AMD has no money to count *laughs in debt*

A negative number is still a number you can count

 

Anyway, totally agree with your post.

AMD are doing the right thing by saying as little as possible and letting Intel take the flack.

I am not hating on AMD for doing it (although if they had been more clear with their first statement then they wouldn't have had to update it). I just don't think people should make a bunch of baseless assumptions, praise AMD for being quiet and hating on Intel for stating facts in a real manner.

 

I am not sure if it's sad or hilarious how different the reactions in this thread are compared to the reactions in the Intel thread where they also announced hardware level fixes in upcoming processors.

 

1 hour ago, Trixanity said:

I think it's fair to say that all data available points to that the older the CPU the bigger the performance loss so in no logical scenario would Skylake be slower than Broadwell unless the point is to say that patched Skylake is slower than unpatched Broadwell. Which may be true but it's a pointless comparison. You don't want an unpatched product so the only thing you can lament is the loss in performance which goes for any Intel user and increasingly so the older your product is.

That's what I was thinking too, but the way I interpret Taf's post was that Skylake has regressed more than Broadwell. Here is the quote:

20 hours ago, Taf the Ghost said:

Meltdown is an Intel issue where they found higher IPC but cutting out steps that are attackable. Skylake cores have now functionally gone backwards in IPC from launch and their generational improvements.

It's true that the performance per clock is lower on Skylake with patches added than it was at launch, but the same can be said for any processor (even AMD ones). So I don't get why Skylake was specifically called out unless the intention is to imply that Skylake had a bigger regression than Broadwell.

It might very well be a bigger regression on Skylake, but it's not something I have heard of and would like to read more about.

 

 

Also, I would like to point out that IPC has not gone down. IPC is "instructions per clock" and processors can still do as many instructions as they could before regardless of the patch. The reason why the patches lowers performance is because they increase the number of instructions necessary to do certain things.

IPC has stayed the same, but performance per clock has gone down. Processors still executes instructions at the same speed, it's just that there are more instructions to execute.

 

Might seem like nitpicking but I think that's an important distinction to make in order to make people understand what is happening. If people learn this now then maybe we can avoid all the "Intel is lying about IPC gains with their new processors!" claims. I wouldn't be surprised if some youtuber or blogger will get this wrong and stir up a bunch of unnecessary controversy.

[leadeater]

1 hour ago, LAwLz said:

Also, I would like to point out that IPC has not gone down. IPC is "instructions per clock" and processors can still do as many instructions as they could before regardless of the patch. The reason why the patches lowers performance is because they increase the number of instructions necessary to do certain things.

IPC has stayed the same, but performance per clock has gone down. Processors still executes instructions at the same speed, it's just that there are more instructions to execute.

That and there is another factor which makes application performance slower. There is a much larger process latency hit due to context switching that is necessary when accessing kernel memory/sys calls after that patch because there is no hidden kernel memory area in user space anymore. These are some of the extra instructions required that you mention but it's not just there is more it's the delay in switching as well.

 

You can do just as many instructions per clock as before but it's more that there are comparatively lengthy stalls when the processor/OS has to access system memory. This is why synthetic benchmarks do not show any performance loss as they are performing single tasks at a time never switching memory areas. One of the best examples to show this effect is the File Copy and Zip some publications did, the files per second took massive hits post patches since there is a lot of sys calls involved with file system access. To be clear this does not make the file system slower or the storage device itself slower its the interaction with the application and system that is now slower, single large file copies are essentially unaffected.

[Taf the Ghost]

3 hours ago, leadeater said:

That and there is another factor which makes application performance slower. There is a much larger process latency hit due to context switching that is necessary when accessing kernel memory/sys calls after that patch because there is no hidden kernel memory area in user space anymore. These are some of the extra instructions required that you mention but it's not just there is more it's the delay in switching as well.

 

You can do just as many instructions per clock as before but it's more that there are comparatively lengthy stalls when the processor/OS has to access system memory. This is why synthetic benchmarks do not show any performance loss as they are performing single tasks at a time never switching memory areas. One of the best examples to show this effect is the File Copy and Zip some publications did, the files per second took massive hits post patches since there is a lot of sys calls involved with file system access. To be clear this does not make the file system slower or the storage device itself slower its the interaction with the application and system that is now slower, single large file copies are essentially unaffected.

@LAwLz

 

Yeah, I should have been more precise with my statement. Theoretical IPC is unchanged, as the dies haven't changed. As a functional amount of "work per time", which I offhanded to IPC, Intel has regressed. In some tasks, but an utterly massive amount. The fixes introduce bottlenecks to the way the processor works.

 

And, depending on where the state of the Spectre/Meltdown patches are when Ryzen 2000 series drops in April, that testing could be really, really wild.

[LAwLz]

11 minutes ago, Taf the Ghost said:

Yeah, I should have been more precise with my statement. Theoretical IPC is unchanged, as the dies haven't changed. As a functional amount of "work per time", which I offhanded to IPC, Intel has regressed. In some tasks, but an utterly massive amount. The fixes introduce bottlenecks to the way the processor works.

Yes I get that, but what I don't get was why you brought up Skylake. The way I interpreted your post was that Skylake now performs worse than Broadwell when patches are applied to both.

If that's not what you mean then I don't see why you brought up Skylake specifically, saying that it performs worse than previous generations.

[RagnarokDel]

7 hours ago, mr moose said:

Next year?  it seems like we were debating the value of AMD's "long platform support" only a few months ago. 

 

 

not sure what you're trying to say? that means we still have 23 month on AM4, at least.

[Taf the Ghost]

1 hour ago, LAwLz said:

Yes I get that, but what I don't get was why you brought up Skylake. The way I interpreted your post was that Skylake now performs worse than Broadwell when patches are applied to both.

If that's not what you mean then I don't see why you brought up Skylake specifically, saying that it performs worse than previous generations.

I was referencing the improvement from Broadwell to Skylake, which was quite small (lots of clocks & cache stuff, though). From a "Then to Now", Skylake cores aren't effectively faster because of the patches. With patches on any Intel product, they're going backwards, but Skylake is the current generation. 

 

"Skylake cores have now functionally gone backwards in IPC from launch and their generational improvements."

 

I shorthanded with "IPC", but the rest is correct. Skylake parts are now all functionally slower than when they launched. After all of the patching is done, it'll be kind of interesting to see what Clock for Clock un-patched previous generations look like compared to Skylake (and to itself). Obviously, the issues have to be fixed, especially in the Server space, but it looks like Intel has, in a lot of work flows, seen a negative performance "improvement" over the last 3 years. It's only the Clocks & Efficiency that don't make this catastrophic. 

[Hunter259]

10 hours ago, Shiftstealth said:

"GPZ Variant 3 (Rogue Data Cache Load or Meltdown) is not applicable to AMD processors.

• We believe AMD processors are not susceptible due to our use of privilege level protections within paging architecture and no mitigation is required."

 

https://www.amd.com/en/corporate/speculative-execution

What does that have to do with what I said? I just said it wasn't Intel trying to cheat....

[LAwLz]

4 hours ago, Taf the Ghost said:

From a "Then to Now", Skylake cores aren't effectively faster because of the patches.

Well again, that applies to any processor because all processors are going to take a performance hit.

 

4 hours ago, Taf the Ghost said:

With patches on any Intel product, they're going backwards, but Skylake is the current generation. 

Skylake isn't current gen. Skylake is 2 generations old at this point. It was replaced over a year ago.

Since then we have had both Kaby Lake (previous gen) as well as Coffee Lake (current gen). Coffee Lake was released like 4 months ago.

 

4 hours ago, Taf the Ghost said:

"Skylake cores have now functionally gone backwards in IPC from launch and their generational improvements."

You can't compare patched Skylake to unpatched Broadwell... That's just not a fair comparison, and newer chips are actually handling the patch better than previous generations.

 

4 hours ago, Taf the Ghost said:

Skylake parts are now all functionally slower than when they launched.

All processors are, including Ryzen and Broadwell processors.

 

4 hours ago, Taf the Ghost said:

but it looks like Intel has, in a lot of work flows, seen a negative performance "improvement" over the last 3 years.

That's simply not true.

3 years ago the best "mainstream" processor you could buy from Intel was the 4790K.

Quad core (hyperthreaded) Haswell processor at 4.4GHz.

 

Today the best one is the 8700K.

Six-core (with hyperthreading) Coffee Lake at 4.3GHz

 

Comparison between the two in Cinebench 15:

Single thread:

4790K - 181

8700K - 197

Performance increase - 9%

 

Multi-thread:

4790K - 894

8700K - 1364

Performance increase - 53%

 

Since the patches will a performance impact of ~5% on average (lower on newer processors) I really doubt there will be an overall regression.

 

I don't really see your point though. Are you trying to make it sound like Intel are the bad guys because performance today with the patches are worse than performance 3 years ago? Even if that was true, I don't think that matters. The patches needs to be applied to older processors too. It's not Intel's fault an exploit was discovered in processors which needs to be fixed at the cost of performance. If you compare the old processors against the new ones with patches applied then the new ones will be faster. They have not gotten slower like you are trying to say. In fact, since the patch has a greater performance impact on older processors the gap between old and new ones is larger now than before the patches.

If there used to be a 5% difference between each new generation of Intel processors, that difference might be 6 or 7% now with the patches applied (numbers pulled out of my butt, but you get the point).

 

 

24 minutes ago, Hunter259 said:

What does that have to do with what I said? I just said it wasn't Intel trying to cheat....

I think Shiftstealth misunderstood your post and thought you're saying AMD processors are vulnerable to Meltdown (which you didn't say, and they aren't).

[mr moose]

4 hours ago, RagnarokDel said:

not sure what you're trying to say? that means we still have 23 month on AM4, at least.

 

Just musing because only a few months ago people were making it sound like AM4 had many years left in the platform (significant enough to justify buying AMD without any other reasoning).   When I read your post I realised that "many years" was next year already.

 

Nothing to do with your comments, just some musings I had reading them.

[linustouchtips]

cool i like the current lineup and since they aren't getting into any hotwater (as of right now) it looks to be smooth sailing (before another bug or something) to the next release. i'd give it a day or two before wallstreet starts slamming them again.

[dalekphalm]

On 1/31/2018 at 8:10 AM, NinJake said:

It's nice to see AMD stepping up to the plate and talking about how they are going about to fix spectre. (No word on meltdown yet)

 

But then take a look over at Intel as they kneel in a corner, counting their money and hoping everyone forgets about these issues if they keep their mouth shut.

AMD clearly states they do not believe Zen is affected by Meltdown - multiple people have listed the quote above.

On 1/31/2018 at 8:58 AM, hey_yo_ said:

With that said, it makes me wonder why would Intel need to ditch the existing LGA 1151 just to fix the Spectre v2 in their 9th gen processors?

I don't think changing sockets have anything to do with fixing Spectre - do you have a source for that? I assume they changed the socket for one of two reasons:

1. Just because, to force people to buy new hardware (less likely)

2. To implement additional improvements of the new architecture (more likely)

3 hours ago, mr moose said:

 

Just musing because only a few months ago people were making it sound like AM4 had many years left in the platform (significant enough to justify buying AMD without any other reasoning).   When I read your post I realised that "many years" was next year already.

 

Nothing to do with your comments, just some musings I had reading them.

"Next year" can be misleading. We're at the very beginning of the 2nd month of 2018. If they guarantee it through to 2019, that leaves just shy of 2 years of support, still.

 

2 years of additional support for AM4 is pretty good. That'll be Ryzen 2000 series for sure, and possibly even Zen 2 (unconfirmed).

[cj09beira]

6 minutes ago, dalekphalm said:

AMD clearly states they do not believe Zen is affected by Meltdown - multiple people have listed the quote above.

I don't think changing sockets have anything to do with fixing Spectre - do you have a source for that? I assume they changed the socket for one of two reasons:

1. Just because, to force people to buy new hardware (less likely)

2. To implement additional improvements of the new architecture (more likely)

"Next year" can be misleading. We're at the very beginning of the 2nd month of 2018. If they guarantee it through to 2019, that leaves just shy of 2 years of support, still.

 

2 years of additional support for AM4 is pretty good. That'll be Ryzen 2000 series for sure, and possibly even Zen 2 (unconfirmed).

there was one interview where an amd employee said that they are aiming for a bios update for zen 2 to work on am4

[Energycore]

On 1/31/2018 at 7:48 AM, The Benjamins said:

Sorry I am dyslexic.

You mean dylsexyc?

[ItsMitch]

5 minutes ago, Energycore said:

You mean dylsexyc?

nice meme

[Trixanity]

24 minutes ago, dalekphalm said:

Next year" can be misleading. We're at the very beginning of the 2nd month of 2018. If they guarantee it through to 2019, that leaves just shy of 2 years of support, still.

 

2 years of additional support for AM4 is pretty good. That'll be Ryzen 2000 series for sure, and possibly even Zen 2 (unconfirmed).

AMD has stated support through 2020, so Zen2 is pretty much guaranteed to run on the same socket given that it's due to launch in 1H 2019. Zen3 might also but it's a bit up in the air at this point. They'll be required to move to a new socket with DDR5 and PCI-E 5.0 which would fit around 2020 anyway.

We're potentially looking at 5 years of support for the AM4 platform. That's pretty damn good in the current climate.

[TheSilverNiko]

Good for them. I hope 2018 continues this trend.

[mr moose]

50 minutes ago, dalekphalm said:

"Next year" can be misleading. We're at the very beginning of the 2nd month of 2018. If they guarantee it through to 2019, that leaves just shy of 2 years of support, still.

 

2 years of additional support for AM4 is pretty good. That'll be Ryzen 2000 series for sure, and possibly even Zen 2 (unconfirmed).

I know,  but 2.5 years is still only half of the average upgrade cycle.  And this is being generous to those arguments I referred to,  because it was only a few months ago people were arguing about it and there are no guarantees that AMD will release the last AM4 stuff right at the end of 2019. They may introduce AM5 (or whatever is next) early in 2020 instead.  It's still an awful lot of speculation to be basing whole hardware/platform upgrades on when we have more relevant benchmarks and current prices to go of.

 

And all that aside, as I said earlier, it was just me musing over those previous arguments because the way I read ragnorak's post that made it seem a lot worse than the 3-4 years support people where talking (only a few months ago).  

[mr moose]

42 minutes ago, Trixanity said:

AMD has stated support through 2020, so Zen2 is pretty much guaranteed to run on the same socket given that it's due to launch in 1H 2019. Zen3 might also but it's a bit up in the air at this point. They'll be required to move to a new socket with DDR5 and PCI-E 5.0 which would fit around 2020 anyway.

We're potentially looking at 5 years of support for the AM4 platform. That's pretty damn good in the current climate.

It's not beyond the realms that AM4 could support CPU's ranging over 6 years.  But it's important to differentiate between being able to run a CPU and actually being able to use the features and performance of it.

 

EDIT: forgot to add, Zen3 is road mapped for AM4 in 2019.

[leadeater]

5 minutes ago, mr moose said:

It's not beyond the realms that AM4 could support CPU's ranging over 6 years.  But it's important to differentiate between being able to run a CPU and actually being able to use the features and performance of it.

 

EDIT: forgot to add, Zen3 is road mapped for AM4 in 2019.

Just like the difference between AM2+ and AM3. AM3 processors could work in both but only AM3 had DDR3 support.

[Taf the Ghost]

3 hours ago, Trixanity said:

AMD has stated support through 2020, so Zen2 is pretty much guaranteed to run on the same socket given that it's due to launch in 1H 2019. Zen3 might also but it's a bit up in the air at this point. They'll be required to move to a new socket with DDR5 and PCI-E 5.0 which would fit around 2020 anyway.

We're potentially looking at 5 years of support for the AM4 platform. That's pretty damn good in the current climate.

AMD will support AM4 through 2020. They've never committed to a new mainstream Desktop CPU then. (I.e. it'll be the Zen2 APUs in 2020 while Desktop goes to AM5, unless DDR5 ends up late. Then, we could see a Zen2+ on AM4.) 

[mr moose]

46 minutes ago, Taf the Ghost said:

AMD will support AM4 through 2020. They've never committed to a new mainstream Desktop CPU then. (I.e. it'll be the Zen2 APUs in 2020 while Desktop goes to AM5, unless DDR5 ends up late. Then, we could see a Zen2+ on AM4.) 

They have mapped Zen3 for AM4 socket in 2019.     It would be very unfair of them to not allow Zen 3 to run on current AM4 motherboards.  Although if ddr5 (or whatever comes next) becomes a thing on desktop, just don't expect to have your cake and eat it too.

 

 

[Jito463]

5 hours ago, Energycore said:

You mean dylsexyc?

You mean lysdexic?

4 hours ago, leadeater said:

Just like the difference between AM2+ and AM3. AM3 processors could work in both but only AM3 had DDR3 support.

Yep, but at least the option was there.  There were also some hybrid boards that took either DDR 2 or DDR 3.

[mr moose]

55 minutes ago, Jito463 said:

You mean lysdexic?

Yep, but at least the option was there.  There were also some hybrid boards that took either DDR 2 or DDR 3.

Problem was if you bought early in the platform like I did, You wouldn't have had a hybrid board, I was stuck with HT v2 and even though I could run the newer CPU's, it was marginal IPC improvements only (FSB being limited by the HT).    So yes I had the option, but all things considered it wasn't a very good option. If all you could afford for the next 4 years was a new CPU it was better than nothing.  But then having a $200 CPU that could run at 1.8Ghz FSB being limited to 1Ghz is painful.

 

Maybe with AM4 there will be no performance hit staying with current ram and MOBO features so might be very different.   I can't imagine a huge leap between ddr4 and whatever is next (not in gaming anyway). 

[Taf the Ghost]

2 hours ago, mr moose said:

They have mapped Zen3 for AM4 socket in 2019.     It would be very unfair of them to not allow Zen 3 to run on current AM4 motherboards.  Although if ddr5 (or whatever comes next) becomes a thing on desktop, just don't expect to have your cake and eat it too.

 

 

"Zen3", from any roadmaps we've seen, has pointed to 2020. And did you mean to place a different video? There isn't a roadmap in that one.

 

Ryzen 3000 series, or "Zen2" or Matisse, will work on AM4, along with the APUs off that process.

 

"Zen3", the 2nd major uArch upgrade to Zen cores, should have a DDR5 controller, however we simply don't know if DDR5 will be on time. Both AMD & Intel are going to run into that issue, so we'll know somewhere in late 2018/early 2019 whether the new memory will be available on-time.

[mr moose]

Just now, Taf the Ghost said:

"Zen3", from any roadmaps we've seen, has pointed to 2020. And did you mean to place a different video? There isn't a roadmap in that one.

 

Ryzen 3000 series, or "Zen2" or Matisse, will work on AM4, along with the APUs off that process.

 

"Zen3", the 2nd major uArch upgrade to Zen cores, should have a DDR5 controller, however we simply don't know if DDR5 will be on time. Both AMD & Intel are going to run into that issue, so we'll know somewhere in late 2018/early 2019 whether the new memory will be available on-time.

No, I posted that video (direct from AMD) because it clearly states Zen3 on "future ready AM4", which is why I said it would very unfair of AMD not to have Zen3 work on AM4. 

 

With regard to the roadmap:

 

https://hothardware.com/news/amd-confirms-am4-socket-support-future-ryzen-processors-2020

https://www.eteknix.com/amd-roadmap-leak-shows-2nd-gen-ryzen-coming-q1-2018/

 

Even if these are leaks or speculation by media, AMD still advertised ZEN3 specifically as "AM4 future ready" .