AMD Discloses Initial Zen 2 Details

Megalith

24-bit/48kHz
Staff member
Joined
Aug 20, 2006
Messages
13,000
WikiChip has published their take on AMD’s next iteration of Zen, which, being built on TSMC’s 7nm process node, will “deliver 2x the density and offer 0.5x the power at the same performance, or >1.25x the performance at the same power.” Aside from IPC improvements, the chip will see an increase in both dispatch and retire bandwidth, along with in-silicon Spectre mitigations previously provided via firmware and software.

AMD has made a large set of enhancements to Zen 2. To feed the widened execution units which were improved in throughput, the front-end had to be adjusted. For that reason, the branch prediction unit has been reworked. This includes improvements to the prefetcher and various undisclosed optimizations to the instruction cache.
 
I like the sound of that. In the past I've usually gone for 45/65w Intel CPUs for routers and servers that didn't need a beefy CPU.
I'm really looking forward to using AMD CPUs again.
 
Are we expecting clock speed increases? That seems to be the only area AMD is behind
 
Are we expecting clock speed increases? That seems to be the only area AMD is behind


An arch shrink should bring a small increase in speed. Probably not enough to overtake Intel single thread even with the enhancements.
 
An arch shrink should bring a small increase in speed. Probably not enough to overtake Intel single thread even with the enhancements.

Zen 2 was built to take on 10nm Ice Lake. It will be faster than any of Intel's desktop parts in single/multithread.
 
ETA? Anyone have any idea.

sometime in 2019. Amd is likely binning the highest end parts for eypc cpu's. Once supply can meet demand in those markets, will we see the desktop am4 refresh. I'd guess March-June.
 
Zen 2 was built to take on 10nm Ice Lake. It will be faster than any of Intel's desktop parts in single/multithread.

AMD 2700X base/boost is 3.7/4.3 GHz. the 9900K is 3.6/4.7/5.0 (all core/single core boost) GHz, and that doesn't even take into account the IPC difference... so it's highly unlikely that Zen 2 would be able to compete directly single core compared to the highest end Intel, even on 7mn... The differences in single core performance are even more pronounced in the server side of CPUs. Obviously, there's a price differential, but you said "any," not bang for buck.
 
AMD 2700X base/boost is 3.7/4.3 GHz. the 9900K is 3.6/4.7/5.0 (all core/single core boost) GHz, and that doesn't even take into account the IPC difference... so it's highly unlikely that Zen 2 would be able to compete directly single core compared to the highest end Intel, even on 7mn... The differences in single core performance are even more pronounced in the server side of CPUs. Obviously, there's a price differential, but you said "any," not bang for buck.

All AMD needs is a 15% bump to clock speed to match the 9900K. That's not an unreasonable amount to get out of the shrink to 7nm based off of the clock boosts that we already see from the 7nm shrink for the MI60 The actual IPC difference is rather minimal, we're talking single digit percentage differences ( https://www.techspot.com/article/1616-4ghz-ryzen-2nd-gen-vs-core-8th-gen/ ).

So if AMD delivers the 1.25x perf bump they are claiming then they will best the 9900K at single threaded workloads. This doesn't exactly strike me as a "highly unlikely" scenario. A sure thing? No, of course not. "highly unlikely" though? No, no not really.

On the server side of things the difference in single-core performance is actually far *less* pronounced as Xeons don't reach nearly the same clock speeds that the i7/i9's do. Top end Xeon (28-core) is 2.8ghz base and the top-end Epyc (32-core) is 2.7ghz base. In the server world it's already sub-10% difference in single-thread performance *and* AMD is doing it while consistently delivering more cores and far more PCI-E lanes. Rome is going to extend this lead even further
 
All AMD needs is a 15% bump to clock speed to match the 9900K. That's not an unreasonable amount to get out of the shrink to 7nm based off of the clock boosts that we already see from the 7nm shrink for the MI60 The actual IPC difference is rather minimal, we're talking single digit percentage differences ( https://www.techspot.com/article/1616-4ghz-ryzen-2nd-gen-vs-core-8th-gen/ ).

So if AMD delivers the 1.25x perf bump they are claiming then they will best the 9900K at single threaded workloads. This doesn't exactly strike me as a "highly unlikely" scenario. A sure thing? No, of course not. "highly unlikely" though? No, no not really.

On the server side of things the difference in single-core performance is actually far *less* pronounced as Xeons don't reach nearly the same clock speeds that the i7/i9's do. Top end Xeon (28-core) is 2.8ghz base and the top-end Epyc (32-core) is 2.7ghz base. In the server world it's already sub-10% difference in single-thread performance *and* AMD is doing it while consistently delivering more cores and far more PCI-E lanes. Rome is going to extend this lead even further

I'm not sure I agree with you on a 15% performance difference between single threaded tasks - looks more like 20%-25 to me. (See the first set of images here: https://arstechnica.com/gadgets/201...e-i9-9900k-review-top-performance-but-pricey/)

Also, no one was talking about only IPC, nor overclocking. It's the combination of higher clocks and higher IPC that has kept Intel ahead in the high end single threaded race. Considering AMD has failed to do just that for... I don't know how many years, now... It seems highly unlikely they will do that now. Those small percentages of faster clocks and IPC multiply together rather nicely. Multithreaded, however, AMD will make up for it on number of cores.
 
AMD 2700X base/boost is 3.7/4.3 GHz. the 9900K is 3.6/4.7/5.0 (all core/single core boost) GHz, and that doesn't even take into account the IPC difference... so it's highly unlikely that Zen 2 would be able to compete directly single core compared to the highest end Intel, even on 7mn... The differences in single core performance are even more pronounced in the server side of CPUs. Obviously, there's a price differential, but you said "any," not bang for buck.

And I'm sticking to what I said. Zen 2 will beat Intel's current offerings in single/mutlthreaded workloads.
 
Zen 2 will be my next desktop CPU. Its gonna be a good one (I hope) haven't had an AMD CPU since 939 socket.

Same here. I just hope some companies come out with some nice matx motherboards by then. I haven't been impressed with any of them so far.
 
I'm not sure I agree with you on a 15% performance difference between single threaded tasks - looks more like 20%-25 to me. (See the first set of images here: https://arstechnica.com/gadgets/201...e-i9-9900k-review-top-performance-but-pricey/)

Also, no one was talking about only IPC, nor overclocking. It's the combination of higher clocks and higher IPC that has kept Intel ahead in the high end single threaded race. Considering AMD has failed to do just that for... I don't know how many years, now... It seems highly unlikely they will do that now. Those small percentages of faster clocks and IPC multiply together rather nicely. Multithreaded, however, AMD will make up for it on number of cores.

I didn't say a 15% performance difference, I said a 15% *clock speed* difference. Which you can't really agree or disagree with, it's a simple fact. Then on top of that you have the sub-10% IPC difference. I talked about both clock & IPC, so I'm not sure why you seem to be implying I talked about only one of those aspects?

Which your link just agreed with, since 15% plus <10% ends up being around 20-25% overall. AMD is claiming 1.25x in perf. What would Zen * 1.25 then result in? Matching/beating the 9900k.

AMD has failed to match Intel for years, sure, but Ryzen closed that gap massively already and Intel has been completely stagnant for how many years now? It's not like AMD is hitting a moving target. Intel is completely frozen in place.
 
Is Zen2 gonna keep the AM4 socket?

AMD previously committed to AM4 being good for a couple generations and they publicly announced that Rome (Zen2 in Epyc form) will be socket compatible with current Epyc.

So very probably yes, it will be AM4. And very probably you won't need a motherboard upgrade, either, unless you want the PCI-E 4.0 support.
 
Zen 2 was built to take on 10nm Ice Lake. It will be faster than any of Intel's desktop parts in single/multithread.
Thats a bold statement, while I really hope it comes through, as I will soon have a third Ryzen system, all ready to drop in Zen 2. Not sure it will work out that way, but I guess AMD is pretty close right now. We need some sweet clock boosts. 4.5-4.7 would be great!
 
Same here. I just hope some companies come out with some nice matx motherboards by then. I haven't been impressed with any of them so far.

I was about to say the exact same thing. The situation with mitx isn't much better I presume. Gonna need both
 
I was about to say the exact same thing. The situation with mitx isn't much better I presume. Gonna need both
I have been ok with the ITX offerings, well if you want robust vrms than we haven't seen those. But yeah whats up with Micro atx, seems to be relegated to budget boards.
 
Same here. I just hope some companies come out with some nice matx motherboards by then. I haven't been impressed with any of them so far.
Yeah this is what I'm holding out for. My most recent rig really sold me on the matx form factor. Unfortunately between atx and mitx it seems to be getting lost in the shuffle as far as good boards go.

Don't want to go back to the days of gigantic "mid" tower cases again.
 
Thats a bold statement, while I really hope it comes through, as I will soon have a third Ryzen system, all ready to drop in Zen 2. Not sure it will work out that way, but I guess AMD is pretty close right now. We need some sweet clock boosts. 4.5-4.7 would be great!

4.7 will be the PBO clocks on 7nm while 5Ghz will come on Zen 3 on 7nm+.
 
I'd caution everyone here to be wary of putting much faith in the proposed clock speeds until they ship parts.

Remember, Intel once thought NetBurst would scale to 10GHz.
 
Things only look better and better on Zen2. Ryzen/Threadripper Zen+ series beat Intel on price across the board and tie or better raw performance in certain workloads. Alas, they seem to lag a little behind in raw performance for single/few threaded loads thanks to IPC and clock speed, however from most of what I've read this is minimal if even noticeable at all in real world performance as opposed to benchmarks. Still, it seem that with Zen2 they may be able to really close the gap even further and perhaps even top the market for a time. Ultimately through, even if they don't crush Intel in every single benchmark etc... if they can offer near to or even a much better real world experience across the board yet do so at a fraction of Intel's prices, they will have have won the generation so to speak.

I can only hope that the continued success of AMD in CPUs (and even GPUs at every sector save for the ultra high end) means that they will be taken seriously meaning that various developers will stop compiling/optimizing specifically for Intel so that certain software doesn't seem to work or perform quite as well on AMD CPUs instead. AMD has made lots of decisions I feel are more ethical than competition and more open as well, so I consider one glaring omission a major issue - the maintaining of the PSP as an impermeable, proprietary black box backdoor with full access to your system . More info from Coreboot/Libreboot - https://libreboot.org/amd-libre.html and https://libreboot.org/faq.html#amd-platform-security-processor-psp - Even if you put ethics aside entirely, AMD has rightfully marketed themselves around openness in many situations as a foil to the proprietary nature of Nvidia and/or Intel. Offering an option to veritably disable the PSP without harming the function of the computer would well be in line with this, but so far they seem to refuse; some consider this proof of meddling by state level actors or other powerful interests especially in the wake of leaks from Snowden and other mentioning the "feature". This is not to say that the PSP or ME couldn't have some viable uses, but it should be well within the user's control to disable them easily if they wish to do so. Its ironic that many are so concerned about Chinese company or manufacturered hardware regarding backdoors, but those same individuals don't seem so concerned about Intel and AMD's hardware. In any event, as they move onto a new process like Zen2 it would be nice for them to allow disabling the PSP and users should all request that such a feature be implemented; even if only a few of us would likely go through the trouble, it would be a benefit to everyone to have it available and secure AMD considerable reputation in the FLOSS/libre, enthusiast, and security communities.
 
Things only look better and better on Zen2. Ryzen/Threadripper Zen+ series beat Intel on price across the board and tie or better raw performance in certain workloads. Alas, they seem to lag a little behind in raw performance for single/few threaded loads thanks to IPC and clock speed, however from most of what I've read this is minimal if even noticeable at all in real world performance as opposed to benchmarks. Still, it seem that with Zen2 they may be able to really close the gap even further and perhaps even top the market for a time. Ultimately through, even if they don't crush Intel in every single benchmark etc... if they can offer near to or even a much better real world experience across the board yet do so at a fraction of Intel's prices, they will have have won the generation so to speak.

I can only hope that the continued success of AMD in CPUs (and even GPUs at every sector save for the ultra high end) means that they will be taken seriously meaning that various developers will stop compiling/optimizing specifically for Intel so that certain software doesn't seem to work or perform quite as well on AMD CPUs instead. AMD has made lots of decisions I feel are more ethical than competition and more open as well, so I consider one glaring omission a major issue - the maintaining of the PSP as an impermeable, proprietary black box backdoor with full access to your system . More info from Coreboot/Libreboot - https://libreboot.org/amd-libre.html and https://libreboot.org/faq.html#amd-platform-security-processor-psp - Even if you put ethics aside entirely, AMD has rightfully marketed themselves around openness in many situations as a foil to the proprietary nature of Nvidia and/or Intel. Offering an option to veritably disable the PSP without harming the function of the computer would well be in line with this, but so far they seem to refuse; some consider this proof of meddling by state level actors or other powerful interests especially in the wake of leaks from Snowden and other mentioning the "feature". This is not to say that the PSP or ME couldn't have some viable uses, but it should be well within the user's control to disable them easily if they wish to do so. Its ironic that many are so concerned about Chinese company or manufacturered hardware regarding backdoors, but those same individuals don't seem so concerned about Intel and AMD's hardware. In any event, as they move onto a new process like Zen2 it would be nice for them to allow disabling the PSP and users should all request that such a feature be implemented; even if only a few of us would likely go through the trouble, it would be a benefit to everyone to have it available and secure AMD considerable reputation in the FLOSS/libre, enthusiast, and security communities.

Won't happen due to N.S.A. BTW...I'm not a conspiracy nut.
 
Won't happen due to N.S.A. BTW...I'm not a conspiracy nut.

Many believe that, but it seems like a poor plan really. It would be easier and with more plausible deniability to simply use zerodays or any of the major operations (ie intercept and modify haardware, write another Stuxnet, ask for a bloody NSAKEY in Windows again though the last one is a bit foolish to repeat etc) available to a state level actor like the NSA, rather than have a "THIS IS A BACKDOOR" thing in just about every major CPU over the past few years. Furthermore it means that if there are ever leaks or even if someone figures out a vulnerability it is basically available on every piece of hardware out there which isn't a good way to secure a nation or (more importantly to some, sadly) to encourage the sales of American developed hardware worldwide if other nations have reasons to believe that all the hardware is compromised by intent by a gov't agency. Even if nobody cares about the cvil liberty violations, it will harm the stock price and orders the first time someone "important" gets owned by the use of these tools or another bombshell drops in some other way. This is entirely the wrong thing to do when ARM is catching up in performance for general use/tablet /non-enthusiast stuff, and China (already skeptical of our hardware just as we are of theirs) will roll out plenty of alternatives. Its just a losing proposition in every way I can think of.

Both Intel and AMD claim that these are totally innocuous (with AMD's interpretation being somewhat better from what I've read, but worse in that there's no way to disable it the way that can be done on some earlier Intel procs) so there's no reason they shouldn't be able to be verifiable switched off. Maybe its the NSA or corporate greed from some interest that wants control or surveillance, or perhaps AMD just doesn't want to waste the time because they figure too few people will care to put the effort into the change/testing it etc... but if we can press on the latter at least, it will tell us something about the circumstances.
 
Many believe that, but it seems like a poor plan really. It would be easier and with more plausible deniability to simply use zerodays or any of the major operations (ie intercept and modify haardware, write another Stuxnet, ask for a bloody NSAKEY in Windows again though the last one is a bit foolish to repeat etc) available to a state level actor like the NSA, rather than have a "THIS IS A BACKDOOR" thing in just about every major CPU over the past few years. Furthermore it means that if there are ever leaks or even if someone figures out a vulnerability it is basically available on every piece of hardware out there which isn't a good way to secure a nation or (more importantly to some, sadly) to encourage the sales of American developed hardware worldwide if other nations have reasons to believe that all the hardware is compromised by intent by a gov't agency. Even if nobody cares about the cvil liberty violations, it will harm the stock price and orders the first time someone "important" gets owned by the use of these tools or another bombshell drops in some other way. This is entirely the wrong thing to do when ARM is catching up in performance for general use/tablet /non-enthusiast stuff, and China (already skeptical of our hardware just as we are of theirs) will roll out plenty of alternatives. Its just a losing proposition in every way I can think of.

Both Intel and AMD claim that these are totally innocuous (with AMD's interpretation being somewhat better from what I've read, but worse in that there's no way to disable it the way that can be done on some earlier Intel procs) so there's no reason they shouldn't be able to be verifiable switched off. Maybe its the NSA or corporate greed from some interest that wants control or surveillance, or perhaps AMD just doesn't want to waste the time because they figure too few people will care to put the effort into the change/testing it etc... but if we can press on the latter at least, it will tell us something about the circumstances.


Like x 1,000,000
 
Back
Top