Zen 2 might not offer the IPC increase you expect

[HardReset]

You're confusing performance and IPC - no worries, insanely easy to do.

Ryzen already outperforms Coffee Lake, per clock, by about 3% in Cinebench. This is mostly thanks to its superior SMT scaling.

Zen 2 would only need about a 10~12% IPC improvement, in Cinebench, for 4.0~4.1GHz with the Zen 2 sample.

At 4.6Ghz, the 2700X would match the 9900k at 4.7GHz.

My comparison was against i9-9900K, not Zen. That's about only thing that can be directly calculated from scores.

Usually IPC means single core IPC and so SMT is left out of equation. It's currently unknown how good SMT is on Zen2. We do already know that Zen IPC is equal to i9-9900K on Cinebench single thread. And Cinebench single thread is what we are interested in. Assuming Zen2 is 3% slower on single thread vs multi thread, 13% IPC improvement on multi thread is 10% on single thread. That means 10% single thread IPC improvement Zen2 clock speed was around 4.27 GHz, not 4 GHz.

This IPC improvement only applies to Cinebench of course.

 

[Arbie]

About Adored he's a well known bullshiter

Adoredtv is a scammer.

AdoredTV categorizes every piece of info he uses as known, almost certain, leaked, speculation etc. He identifies the source of every piece of data. He analyzes all this and makes sincere best-effort estimates based on the results, with all assumptions noted. He pulls no punches on Intel or AMD ("Vega is garbage"). He scripts all this in a logical, straightforward way where every conclusion is supported. He presents it without drama or even being in the picture - unlike most tech Youtubers.

But - you call him names. Guess that counts for more, to some people.

Perhaps, having disposed of AdoredTV so persuasively, you can tell us of even one other such journo who does anything close to the above. If they're in the same class I may go for Patron there too. So far I haven't found any.

 

[looncraz]

My comparison was against i9-9900K, not Zen. That's about only thing that can be directly calculated from scores.

Usually IPC means single core IPC and so SMT is left out of equation. It's currently unknown how good SMT is on Zen2. We do already know that Zen IPC is equal to i9-9900K on Cinebench single thread. And Cinebench single thread is what we are interested in. Assuming Zen2 is 3% slower on single thread vs multi thread, 13% IPC improvement on multi thread is 10% on single thread. That means 10% single thread IPC improvement Zen2 clock speed was around 4.27 GHz, not 4 GHz.

This IPC improvement only applies to Cinebench of course.

CES demo was 9900k vs Zen 2 8-core of unknown specifications. It was a multi-threaded task, so we CAN NOT derive IPC from it precisely because we don't know SMT scaling may have changed. In all probability, judging by the 12-core leaks on Userbenchmark, scaling isn't terribly different, but may be a little worse, which would mean Zen 2's IPC has increased even more than the benchmark would suggest.

However, we have two main points of reference:

1. We know how fast the 9900k needs to be to get the scored it got - 4.7GHz all core.
2. We know how the Ryzen 8-core CPUs behave and scale in Cinebench with frequency.

Having said reference points, we can derive a relationship between the 9900k and the 2700X on any basis we please - frequency, SMT, etc...

From this, we know that the 9900k is SLOWER per clock than the 2700X in Cinebench R15 multi-threaded. We know why, in this case, it's AMD's SMT outperforming Hyper-threading.

Since we know that a 2700X at 4.7GHz would beat a 9900k at 4.7GHz we know that the Zen 2 core must have been operating below that unless there were major performance regressions or no performance gains what-so-over and everything AMD did was with frequency.

How can we test the frequency hypothesis?

Simple: we know 7nm's power behavior - we have shipping 7nm products, technical docs, and AMD's own slides. At the same frequency, 7nm uses half the power of 14LPP or it can provide 25% higher performance at the same power. I'm not going to bother, but we can draw a graph of this function that shows how power for the CCDs behave. The IO die, however, will use the same-ish power as first gen Ryzen's uncore area - so about 15W. From this, and accounting for nominal VRM and power supply efficiencies, we arrive at a core power value of only 50W. An 1800X, with 100W going to the cores alone, operates at or under 4.0GHz. A 2700X operates closer to 4.15GHz at that core power.

Using that information, we can, from power alone, estimate a range of 4~4.15Ghz. We can give ourselves some more margin for error, but there's no way to get the cores beyond 4.3GHz unless power efficiency is insanely good and IPC gain sadly small.

Given that Userbenchmark and Sandra results show healthy IPC gains and how Cinebench responds to certain architectural changes, we would expect Cinebench, if anything, to respond better than the Userbenchmark mixed results. For a 2700X to match Zen 2's single threaded mixed results it needs to operate roughly 10% faster.

If we use the 10% IPC gain value we end up with the same approximate range of 4.0~4.15GHz, scaling the 2700X results.

This means that the CES sample had to have been operating at or below 4.2GHz, for certain, indicating at LEAST 11.9% higher performance in Cinebench, per clock, than a 9900k. That represents an 8.6% performance increase, the absolute low end of what to expect - in Cinebench R15 multi-threaded - from Zen 2.

The more probable range is higher - I used the worst case all along here.

 

[HardReset]

CES demo was 9900k vs Zen 2 8-core of unknown specifications. It was a multi-threaded task, so we CAN NOT derive IPC from it precisely because we don't know SMT scaling may have changed. In all probability, judging by the 12-core leaks on Userbenchmark, scaling isn't terribly different, but may be a little worse, which would mean Zen 2's IPC has increased even more than the benchmark would suggest.

However, we have two main points of reference:

1. We know how fast the 9900k needs to be to get the scored it got - 4.7GHz all core.
2. We know how the Ryzen 8-core CPUs behave and scale in Cinebench with frequency.

Having said reference points, we can derive a relationship between the 9900k and the 2700X on any basis we please - frequency, SMT, etc...

From this, we know that the 9900k is SLOWER per clock than the 2700X in Cinebench R15 multi-threaded. We know why, in this case, it's AMD's SMT outperforming Hyper-threading.

Since we know that a 2700X at 4.7GHz would beat a 9900k at 4.7GHz we know that the Zen 2 core must have been operating below that unless there were major performance regressions or no performance gains what-so-over and everything AMD did was with frequency.

Problem is still that we cannot say how much single thread IPC is gained when looking only for multi threaded results. Because of SMT, single thread gain is also smaller than multi thread gain if we assume both Zen and Zen2 has "equally effective" SMT.

What I agree is that Zen2 was surely working below 4.7 GHz assuming 9900K had full core turbo clocks and they were not limited.

How can we test the frequency hypothesis?

Simple: we know 7nm's power behavior - we have shipping 7nm products, technical docs, and AMD's own slides. At the same frequency, 7nm uses half the power of 14LPP or it can provide 25% higher performance at the same power. I'm not going to bother, but we can draw a graph of this function that shows how power for the CCDs behave. The IO die, however, will use the same-ish power as first gen Ryzen's uncore area - so about 15W. From this, and accounting for nominal VRM and power supply efficiencies, we arrive at a core power value of only 50W. An 1800X, with 100W going to the cores alone, operates at or under 4.0GHz. A 2700X operates closer to 4.15GHz at that core power.

AMD's 25% higher performance on same power is way underrated. Against 16nm TSMC FF 7 nm offers 38% more performance on same power or 60% less power. Since we know that TSMC's 16nm FF overclocks better than GF/Samsung 14nm LPP, on higher frequencies power consumption difference should be even higher. Sources https://www.tsmc.com/english/dedicatedFoundry/technology/7nm.htm
https://www.tsmc.com/english/dedicatedFoundry/technology/10nm.htm

Using that information, we can, from power alone, estimate a range of 4~4.15Ghz. We can give ourselves some more margin for error, but there's no way to get the cores beyond 4.3GHz unless power efficiency is insanely good and IPC gain sadly small.

Given that Userbenchmark and Sandra results show healthy IPC gains and how Cinebench responds to certain architectural changes, we would expect Cinebench, if anything, to respond better than the Userbenchmark mixed results. For a 2700X to match Zen 2's single threaded mixed results it needs to operate roughly 10% faster.

If we use the 10% IPC gain value we end up with the same approximate range of 4.0~4.15GHz, scaling the 2700X results.

This means that the CES sample had to have been operating at or below 4.2GHz, for certain, indicating at LEAST 11.9% higher performance in Cinebench, per clock, than a 9900k. That represents an 8.6% performance increase, the absolute low end of what to expect - in Cinebench R15 multi-threaded - from Zen 2.

The more probable range is higher - I used the worst case all along here.

Power consumption was system power consumption, not CPU power consumption. Since we didn't have 2700X system as a reference, calculations are very hard to make.

What we also don't know is 7nm tech's effiency curve. Since Zen CPU clocks are very high considering what process is, 7nm tech on same frequency could be ultra energy-efficient. So yes, power effiency may well be insanely good on lower clock speeds.

So based on how energy efficient 7nm tech should be, I predict around 4.3 GHz clock speeds on all cores. I also expect that getting around 5 GHz all core clocks is very possible without exploding energy consumption.

Luckily AMD promised to tell more about Zen2 in just three weeks.

 

[looncraz]

Problem is still that we cannot say how much single thread IPC is gained when looking only for multi threaded results.

No, not with complete accuracy, but we have three other benchmarks for correlation and I have years and years of experience estimating multi-pipeline process performance (from factory throughput to CPUs) to help me along

What I agree is that Zen2 was surely working below 4.7 GHz assuming 9900K had full core turbo clocks and they were not limited.

They were not limited, we know that from the scores.

AMD's 25% higher performance on same power is way underrated.

Yes, because there's a frequency cap We also don't truly know if that >25% is pure frequency or from IPC * frequency.

Power consumption was system power consumption, not CPU power consumption. Since we didn't have 2700X system as a reference, calculations are very hard to make.

Yes, it was system power consumption, which is why you have to remove the baseline power. AMD, I believe intentionally, allowed us to see the idle power without power management enabled before the test began. This gave us a baseline - we can straight-up remove the idle power and look for the core power gain and compare that with 2700X configured similarly. OR, we can take the full system power and remove standard values and percentages to arrive at roughly the same values.

It was a 65W CPU with about 50~55W going to the cores. SR/PR needs double that power for ~4GHz.

What we also don't know is 7nm tech's effiency curve.

We actually do - 7nm products exist in the wild (I own one ). I know I have a spreadsheet somewhere that lets me graph these up... I probably named it something stupid.. (I have, literally, like 300 "results48844.ods" files that are me working things out, LOL! Damn I'm bad at file names...

So based on how energy efficient 7nm tech should be, I predict around 4.3 GHz clock speeds on all cores. I also expect that getting around 5 GHz all core clocks is very possible without exploding energy consumption.

Luckily AMD promised to tell more about Zen2 in just three weeks.

This is reasonable, my estimates are in the same ballpark, though there are some design arguments that could limit frequency to 4.7GHz or even allow 5.2GHz+ with good cooling. It's all about the timing and details regarding the LSU and reworked L1 cache. The FPU was simply widened from all appearances.[/QUOTE]

 

[HardReset]

No, not with complete accuracy, but we have three other benchmarks for correlation and I have years and years of experience estimating multi-pipeline process performance (from factory throughput to CPUs) to help me along

Sounds good.

They were not limited, we know that from the scores.

Good point.

Yes, because there's a frequency cap We also don't truly know if that >25% is pure frequency or from IPC * frequency.

25% is frequency only when talking about process.

Yes, it was system power consumption, which is why you have to remove the baseline power. AMD, I believe intentionally, allowed us to see the idle power without power management enabled before the test began. This gave us a baseline - we can straight-up remove the idle power and look for the core power gain and compare that with 2700X configured similarly. OR, we can take the full system power and remove standard values and percentages to arrive at roughly the same values.

It was a 65W CPU with about 50~55W going to the cores. SR/PR needs double that power for ~4GHz.

AMD might have used less power hungry motherboard. Since AMD motherboards are pretty power hungry, it makes estimating quite hard. They even could have used x500 series chipset, not available right now.

65W CPU is sadly only rumor.

We actually do - 7nm products exist in the wild (I own one ). I know I have a spreadsheet somewhere that lets me graph these up... I probably named it something stupid.. (I have, literally, like 300 "results48844.ods" files that are me working things out, LOL! Damn I'm bad at file names...

I know what you mean but 7nm CPU's are nowhere to be found and Vega is pretty bad example of frequency scaling.

This is reasonable, my estimates are in the same ballpark, though there are some design arguments that could limit frequency to 4.7GHz or even allow 5.2GHz+ with good cooling. It's all about the timing and details regarding the LSU and reworked L1 cache. The FPU was simply widened from all appearances.

Since Zen is only limited by manufacturing tech and AMD had tons of time to tweak Zen2, it's hard to believe it wouldn't clock at least on par with i9-9900K. Very probably it clocks much better. Too bad AMD still hadn't said much about Zen2 changes. Overall they may be so huge that we are not talking about reworked Zen but more like new architecture. It's well known that Zen was rushed into market and so for example AVX units were basically same as on Excavator. Also branch predictor design is mostly from Excavator.

 

[Bullwinkle M]

Now that the fanboys have all chimed in....

We need to accurately compare AMD vs Intel for a single core at the exact same frequency on the same process to see who wins and by how much

When Intel hits 7nm, we will finally have a valid test

Until then, Intel's process looks superior to anything AMD has @ 14nm and above

Comparing 7nm to 14nm is pretty lame, as an actual 7nm process should give the same performance at the same frequency using what? 1/2 the power?

AMD cannot do that now at 7nm, which is why we should only compare apples to apples

 

[gamerk2]

It should be noted out that AMD did have an official slide showing they achieved 1.25 times the performance from the 7nm node alone

https://hothardware.com/ContentImag...r-gains.jpg.ashx?maxwidth=1170&maxheight=1170

Performance isn't IPC of course and that figure likely includes clock gains and IPC gains (from the reduced distance the electricity has to travel).

Performance at the same power envelope, which is key here. So if everything else were kept equal, you'd see a theoretical 25% IPC gain. But throw in power optimizations and so on, and that declines fast.

I eagerly await AMD fans proclaiming how 10% IPC increases per generation is a solid rate of improvement. :/

 

[Evernessince]

Performance at the same power envelope, which is key here. So if everything else were kept equal, you'd see a theoretical 25% IPC gain. But throw in power optimizations and so on, and that declines fast.

I eagerly await AMD fans proclaiming how 10% IPC increases per generation is a solid rate of improvement. :/

Well first off I was referring to performance, not IPC. Second, that's 1.25x at the same power consumption. In addition, AMD is also claiming twice the cores at the same power as well. Mind you that number is nothing amazing when moving to a smaller node. It's not AMD's fault that Intel can't get off 14nm.

FYI a 10% IPC increase each generation would be pretty good, especially when Intel is either only getting 3% or 0% (like 8000 to 9000 series)

 

[Arbie]

https://www.pcgamer.com/hands-on-with-the-amd-radeon-vii/

That article says roughly: "... the base clockspeed for the GPU is 1450MHz... AMD is going with 300W TDP for the Radeon VII, nearly the same as Vega 64. AMD says Vega 7nm is mostly the same architecture as the previous Vega, with small tweaks."

SO --> with power and architecture almost the same, it looks like going from GF 14nm to TSMC 7nm only allows an increase of 12% in base clock speeds (1200MHz Vega 64 to 1450MHz Radeon VII).

I could easily be missing something in this simplistic comparison... hope so...

 

[Evernessince]

https://www.pcgamer.com/hands-on-with-the-amd-radeon-vii/

That article says roughly: "... the base clockspeed for the GPU is 1450MHz... AMD is going with 300W TDP for the Radeon VII, nearly the same as Vega 64. AMD says Vega 7nm is mostly the same architecture as the previous Vega, with small tweaks."

SO --> with power and architecture almost the same, it looks like going from GF 14nm to TSMC 7nm only allows an increase of 12% in base clock speeds (1200MHz Vega 64 to 1450MHz Radeon VII).

I could easily be missing something in this simplistic comparison... hope so...

You are replying to the wrong article. You meant to reply to the rumored Navi article surely.

Your numbers are correct though. The Radeon VII was just a die shrink. Instead of adding more cores to increase performance they decided to just shrink the die to increase yields. Likely needed given that 7nm is still young.

I expect more out of products on a more mature 7nm process that actually improve the architecture.

 

[Arbie]

No, I thought it could apply here, strictly as a data point on process gain in general. Then one can add estimates of IPC / architecture specific to CPUs.

It does look like clocks are not going to move "much", even with this numerically huge node shrink. That's not to imply that any of this is easy, of course... I had just hoped for more.

 

[Evernessince]

No, I thought it could apply here, strictly as a data point on process gain in general. Then one can add estimates of IPC / architecture specific to CPUs.

It does look like clocks are not going to move "much", even with this numerically huge node shrink. That's not to imply that any of this is easy, of course... I had just hoped for more.

Got ya. Well going off the VII numbers of a 12% increase in clocks, 4.35 GHz (2700X max boost clock) x 1.12 (A 12% increase) = 4.872 GHz. Mind you that's at the same power and assumes the node hasn't matured since then (which is has) so it's a conservative estimate. At that frequency, even a small bump in IPC would be huge. For example, let's say AMD only gets a 10% IPC increase. You can do the math to find the effective speed compared to the 2700X. 4.872 GHz x 1.10 = 5.3593 GHz. Mind you that's not the actual speed the processor is running, it's merely a value representing the frequency a 2700X would have to run at in order to have equal performance based off the 10% IPC assumption and 12% frequency increases from node.

Even though that's just a modest estimate, a 2700X at that frequency scores 2502 in Cinebench while the 9900K only scores 2044.

 

[HardReset]

https://www.pcgamer.com/hands-on-with-the-amd-radeon-vii/

That article says roughly: "... the base clockspeed for the GPU is 1450MHz... AMD is going with 300W TDP for the Radeon VII, nearly the same as Vega 64. AMD says Vega 7nm is mostly the same architecture as the previous Vega, with small tweaks."

SO --> with power and architecture almost the same, it looks like going from GF 14nm to TSMC 7nm only allows an increase of 12% in base clock speeds (1200MHz Vega 64 to 1450MHz Radeon VII).

I could easily be missing something in this simplistic comparison... hope so...

You are. TDP is NOT power consumption. It has never been. Practically Radeon VII consumes less power than Vega 64 even it has higher TDP. It depends on what game is used but Radeon VII uses 5-20% less power than Vega 64. So 12% more clocks and 5-20% less power is what should be expected. And: there are architectural differences too so straight comparison is not possible.

Also Radeon VII overclocks around 20-25% higher clocks than Vega64. And overclock results are good indicator how manufacturing process affects. GCN architecture is not known for high clocks so expect Zen2 to clock much better.

 

[Arbie]

Radeon VII uses 5-20% less power than Vega 64. So 12% more clocks and 5-20% less power is what should be expected. And: there are architectural differences too so straight comparison is not possible. Also Radeon VII overclocks around 20-25% higher clocks than Vega64. And overclock results are good indicator how manufacturing process affects. GCN architecture is not known for high clocks so expect Zen2 to clock much better.

All good points, and certainly more encouraging. I had wondered if the boost clock might not be the better comparison. Extrapolating from today's TR 2950X (3.5/4.4GHz, 180W) with the factors you mention indicates that the AdoredTV December predictions for a 16c/32t Ryzen 3850X at 4.3/5.1GHz, 135W are not unreasonable. Pricing is up to AMD but he figured they could go as low as $500. I'll probably buy one if it's anywhere near that.

 

[HardReset]

All good points, and certainly more encouraging. I had wondered if the boost clock might not be the better comparison. Extrapolating from today's TR 2950X (3.5/4.4GHz, 180W) with the factors you mention indicates that the AdoredTV December predictions for a 16c/32t Ryzen 3850X at 4.3/5.1GHz, 135W are not unreasonable. Pricing is up to AMD but he figured they could go as low as $500. I'll probably buy one if it's anywhere near that.

Boost clock is better as it tells where either architecture or manufacturing tech limits are.

AdoredTV "predictions" are just guesses, nothing more. If best AMD can show January is engineering sample, how even AMD can know what they are going to release six months later? Also 16-core Ryzen makes no sense, since Intel is only offering 8 cores right now. There is no reason to release more than 8-core Ryzen 3000, perhaps 12-core if Intel releases 10-core Core. But 16-core is more imagination than realism.

 

[looncraz]

If best AMD can show January is engineering sample, how even AMD can know what they are going to release six months later?

I've had that same problem with Adored's list. If it was a leaked forward-looking lineup, it makes perfect sense.. a "what could we potentially accomplish" type of list.

Thing with that, though, is that those lists usually serve as a basis for advancement. AMD's old Exascale whitepaper is becoming reality - Cray is building it.

I fully believe Adored had a genuine leaker provide him with such a list. I believe some of those details are tailor-made to be leaked... by AMD. It's a lot easier to leak some proposed lineup and see how the public might react rather than release something and upset the people, in any event. This would make the most sense to feel out for pricing tolerance.

People like me have no issues with a $500 12-core Ryzen, but that presents a problem for where to put a 16-core, because I'm not willing to spend substantially more for a 16-core.... I just have too little use for one, at this time. However, if it's $600, I'm on board, but they would need to push the 12-core price down to $430 or some such.

Also 16-core Ryzen makes no sense, since Intel is only offering 8 cores right now. There is no reason to release more than 8-core Ryzen 3000, perhaps 12-core if Intel releases 10-core Core. But 16-core is more imagination than realism.

So AMD had no reason to launch the 8-core Ryzen, then, since Intel's mainstream socket only had four cores.

AMD will do what they feel adds the most value to gain sales. Simple as that. If we only see 4.35GHz or so, then 16-core options are essentially a requirement. If we see 4.8Ghz+, 16-core options are just gravy.

 

[Arbie]

+1 looncraz. Exactly what I was in the process of typing! Thank you.

Next month will be very interesting.

 

[HardReset]

I've had that same problem with Adored's list. If it was a leaked forward-looking lineup, it makes perfect sense.. a "what could we potentially accomplish" type of list.

Thing with that, though, is that those lists usually serve as a basis for advancement. AMD's old Exascale whitepaper is becoming reality - Cray is building it.

I fully believe Adored had a genuine leaker provide him with such a list. I believe some of those details are tailor-made to be leaked... by AMD. It's a lot easier to leak some proposed lineup and see how the public might react rather than release something and upset the people, in any event. This would make the most sense to feel out for pricing tolerance.

People like me have no issues with a $500 12-core Ryzen, but that presents a problem for where to put a 16-core, because I'm not willing to spend substantially more for a 16-core.... I just have too little use for one, at this time. However, if it's $600, I'm on board, but they would need to push the 12-core price down to $430 or some such.

AMD has no need to "create" that kind of list as anyone can make similar "leak". When AMD showed off Ryzen ES, it was clear that in no time rumours about 16-core Zen will appear.

So AMD had no reason to launch the 8-core Ryzen, then, since Intel's mainstream socket only had four cores.

AMD will do what they feel adds the most value to gain sales. Simple as that. If we only see 4.35GHz or so, then 16-core options are essentially a requirement. If we see 4.8Ghz+, 16-core options are just gravy.

AMD had already 8 core CPU's on market so at least 8 core for Ryzen was a must. Another reason is Zeppelin CCX-design that has 8 cores. So unless majority of Zeppelins were partially broken (less than 8 cores work), there was no real reason not to release 8 cores into desktop too. Also AMD promised 8 cores to AM4 more than 2 years before Ryzen actually was available. This time AMD has been Very silent about Zen2.

Not so simple. If Zen2 desktop model has 16 cores, there is no coming back. Also Zen3 must have at least 16 cores. And Zen 4. And... Adding cores is very easy way to get more CPU power. But it can only be done once. Because after adding more cores, it becomes standard and offering less cores is not real option. 5 GHz won't be problem for Zen2. Remember that Zen was made with process that was designed for mobile phone chips. Considering that Zen reached very high clock speeds.

 

[Evernessince]

AMD has no need to "create" that kind of list as anyone can make similar "leak". When AMD showed off Ryzen ES, it was clear that in no time rumours about 16-core Zen will appear.



AMD had already 8 core CPU's on market so at least 8 core for Ryzen was a must. Another reason is Zeppelin CCX-design that has 8 cores. So unless majority of Zeppelins were partially broken (less than 8 cores work), there was no real reason not to release 8 cores into desktop too. Also AMD promised 8 cores to AM4 more than 2 years before Ryzen actually was available. This time AMD has been Very silent about Zen2.

Not so simple. If Zen2 desktop model has 16 cores, there is no coming back. Also Zen3 must have at least 16 cores. And Zen 4. And... Adding cores is very easy way to get more CPU power. But it can only be done once. Because after adding more cores, it becomes standard and offering less cores is not real option. 5 GHz won't be problem for Zen2. Remember that Zen was made with process that was designed for mobile phone chips. Considering that Zen reached very high clock speeds.

We'll we know AMD could do 16 cores on desktop but you are questioning "will they?" and it's a good question. I personally think they will but after a few months. At launch I'm betting 12 cores and a 16 core model to counter whatever Intel may launch. AMD does have a good bit of wiggle room on the price as well, given that Intel's 8 core 9900K is $550. I would not be surprised if Intel drops it's 10 core consumer processor for something like $650 and then AMD releases their 16 core at the same price.

 

[looncraz]

AMD has no need to "create" that kind of list

They HAVE to make such a list. Otherwise they will have no clue as to what to do as a business. It's a projected product lineup or hypothetical scenario based on expected capabilities. I've made about 50 such lists - two of which were purposefully leaked (in part, of course) by the company in order to say things we couldn't officially say... because you don't get in trouble for a leak that says you will have a $300 part with X capabilities when what you finally release is a $400 part with X-y capabilities... but if you had come out and said "We expect to be able to sell this at $300 with X capabilities" and failed to do so your stock price tanks... and you could possibly even be sued.

When AMD showed off Ryzen ES, it was clear that in no time rumours about 16-core Zen will appear.

The rumor for 16-cores first. The list may or may not be real, no idea, but the rumor came before the architectural reveal. This same source said Rome was a 9-chip CPU - an IO die and 8 chiplets... no one really believed that - I certainly didn't. Well... it has nine chips with an IO die.

This time AMD has been Very silent about Zen2.

Officially? Absolutely. We only know its FPU capabilities, front end adjustment, LSU, L1 cache reworking, etc... directly from AMD. And we know it can match a stock 9900k in Cinebench while using a crazy ton less power.

From credible or verified leaks, we know it's at least capable 4.5GHz, we know the 16-core variant is capable of at least 4.2GHz, and we can deduce the likely power required to achieve the base clocks for a given core count - and those numbers, despite how it first appears, actually make sense with Adored's list.... because the IO die is a bit hungry. 15~20W TDP for it, alone. 8-core use ~50W at ~4Ghz, double that to 16 cores and you're at 115~120W actual power usage for a 16-core ~4GHz CPU. Strange, but it just works out.

Not so simple. If Zen2 desktop model has 16 cores, there is no coming back.

That's not an issue. Zen 3 will likely not change core counts - 7nm+ isn't much denser. 5nm will be, IIRC, about 80% denser. It will have to be considered if higher core counts are needed. I suspect 16 cores will be a plateau for a couple generations.

Zen 4 will likely be a DDR5 design with the chiplets mounted on top of the IO die. The CPU chiplet dies (CCDs) will be tiny and will fit nicely atop the IO die. The CCD(s) can be located directly above the IMC, allowing a very large reduction in latency from the IMC to the CCD(s). Memory latency should drop to 50ns, maybe lower, assuming DDR4-like latencies (I haven't had a chance - or cause - to investigate DDR5).

5 GHz won't be problem for Zen2. Remember that Zen was made with process that was designed for mobile phone chips. Considering that Zen reached very high clock speeds.

Purely mathematically, Zen 2 should be able to reach almost 5.3GHz with the best dies. This is based on transistor switching times and distance-based latencies. The reality will surely be lower than that, but it was me doing that math that took me from "Adored's list is just insanely stupid and impossible" to "umm... well... maybe, I suppose."

 

[HardReset]

We'll we know AMD could do 16 cores on desktop but you are questioning "will they?" and it's a good question. I personally think they will but after a few months. At launch I'm betting 12 cores and a 16 core model to counter whatever Intel may launch. AMD does have a good bit of wiggle room on the price as well, given that Intel's 8 core 9900K is $550. I would not be surprised if Intel drops it's 10 core consumer processor for something like $650 and then AMD releases their 16 core at the same price.

AMD could beat Intel 10-core with Ryzen 12-core so why release Ryzen 16-core? Also 12-core on launch would raise bar, then Zen3 must have 12-core model too. And that's not good for profits.

They HAVE to make such a list. Otherwise they will have no clue as to what to do as a business. It's a projected product lineup or hypothetical scenario based on expected capabilities. I've made about 50 such lists - two of which were purposefully leaked (in part, of course) by the company in order to say things we couldn't officially say... because you don't get in trouble for a leak that says you will have a $300 part with X capabilities when what you finally release is a $400 part with X-y capabilities... but if you had come out and said "We expect to be able to sell this at $300 with X capabilities" and failed to do so your stock price tanks... and you could possibly even be sued.

This is internet. AMD has no need for that because there are many people who will make same kind of list for free and without any risk.

The rumor for 16-cores first. The list may or may not be real, no idea, but the rumor came before the architectural reveal. This same source said Rome was a 9-chip CPU - an IO die and 8 chiplets... no one really believed that - I certainly didn't. Well... it has nine chips with an IO die.

For Rome being 9-core CPU was very easy to guess. How? Easy. Let's just assume AMD wants to retain 2*4 core module (2 CCX) because that makes software optimization easier. Possible solutions for 64-core CPU:

- Like Epyc Naples but just 8 modules instead of 4. Epyc Naples had severe problems with core to core communication so this is not an option.

- Putting 2 modules on one chip and four of these to single package like on Epyc Naples. This would make manufacturing much harder. No way.

- Raise CCX core count to 8 so module would have 16 cores. This is only option that makes some sense, but it also makes manufacturing harder and software optimization needs to be aware that module now has 2*8 cores instead 2*4.

- 8 modules 2*4 CCX with some kind of "central chip". This would retain software optimizations and also make sure all modules have same memory latency.

- Some kind of "mesh network" like on Intel CPU's. This is very expensive to manufacture.

So it was very good guess that 64-core Epyc would be with "9 cores".

Another problem is that AMD told about Epyc Rome over six months ago. Ryzen 3000 info is still coming.

Officially? Absolutely. We only know its FPU capabilities, front end adjustment, LSU, L1 cache reworking, etc... directly from AMD. And we know it can match a stock 9900k in Cinebench while using a crazy ton less power.

AMD said they made some adjustments, just didn't say more precisely what. Considering to what Intel told about Sunny Cove, AMD didn't tell anything else than double wide AVX units.

From credible or verified leaks, we know it's at least capable 4.5GHz, we know the 16-core variant is capable of at least 4.2GHz, and we can deduce the likely power required to achieve the base clocks for a given core count - and those numbers, despite how it first appears, actually make sense with Adored's list.... because the IO die is a bit hungry. 15~20W TDP for it, alone. 8-core use ~50W at ~4Ghz, double that to 16 cores and you're at 115~120W actual power usage for a 16-core ~4GHz CPU. Strange, but it just works out.

We don't need any leaks to tell it's capable for at least 5 GHz.

What remains to be seen is balance between power consumption and clock speeds.

That's not an issue. Zen 3 will likely not change core counts - 7nm+ isn't much denser. 5nm will be, IIRC, about 80% denser. It will have to be considered if higher core counts are needed. I suspect 16 cores will be a plateau for a couple generations.

I suspect also that 16-cores is enough for long time but also say that faster AMD get's 16-cores into mainstream, less AMD will make profit long term.

Zen 4 will likely be a DDR5 design with the chiplets mounted on top of the IO die. The CPU chiplet dies (CCDs) will be tiny and will fit nicely atop the IO die. The CCD(s) can be located directly above the IMC, allowing a very large reduction in latency from the IMC to the CCD(s). Memory latency should drop to 50ns, maybe lower, assuming DDR4-like latencies (I haven't had a chance - or cause - to investigate DDR5).

OK, I will see later if this prediction holds true

Purely mathematically, Zen 2 should be able to reach almost 5.3GHz with the best dies. This is based on transistor switching times and distance-based latencies. The reality will surely be lower than that, but it was me doing that math that took me from "Adored's list is just insanely stupid and impossible" to "umm... well... maybe, I suppose."

Reality for Ryzen was way higher it should have been. Based on manufacturing process, getting 3.6 GHz base clock for 8 cores with 95W and that performance should have been impossible.

 

[Arbie]

Regarding the likelihood of a Ryzen 16-core, it's interesting that Threadripper has suddenly been deleted from the AMD roadmap. That would make some room.

 

[Evernessince]

AMD could beat Intel 10-core with Ryzen 12-core so why release Ryzen 16-core? Also 12-core on launch would raise bar, then Zen3 must have 12-core model too. And that's not good for profits.

Raising the core count bar is more of a problem for Intel right now then AMD. AMD's Zen 2 CCX modules are even smaller then first gen, meaning higher yields and lower costs. That's above what they were getting before. AMD's costs increase linearly, Intel's increase exponentially. It benefits AMD to force Intel to start making higher core count consumer CPUs as it will eat into their margins.

In addition, you are assuming that AMD wants to toe the line and just barely beat the competition and risk loosing any chance they have to become the top dog in the market. It would be extremely foolish if AMD held back it's own technology for $$$ and that's not something I'd expect at all from Lisa Sue, she isn't a bean counter like Intels CEO. Mind you this is really only the start for Chiplet based CPUs, you have yet to see the benefit active interposers will have. 16 cores is nothing, active interposers allow AMD to far exceed 128 cores AND get lower latency then a monolithic design.

 

[Mahone]

AdoredTV categorizes every piece of info he uses as known, almost certain, leaked, speculation etc. He identifies the source of every piece of data. He analyzes all this and makes sincere best-effort estimates based on the results, with all assumptions noted. He pulls no punches on Intel or AMD ("Vega is garbage"). He scripts all this in a logical, straightforward way where every conclusion is supported. He presents it without drama or even being in the picture - unlike most tech Youtubers.

But - you call him names. Guess that counts for more, to some people.

Perhaps, having disposed of AdoredTV so persuasively, you can tell us of even one other such journo who does anything close to the above. If they're in the same class I may go for Patron there too. So far I haven't found any.

I call him names because he deserves it. I've been following him since his breakthrough video about gimpworks in 2015.Let's start shall we he's got a lot of flaws one of them is he can't take criticism. He got into an argument with someone and told him to kill himself - who does that? Called people who bought Nvidia products > Nv*****s. Bragged about how he's at home while his pregnant girlfriend has to work. He defended the RX 480 power draw from pci-e - which later amd fixed with a software update (some people had their mobo's fried cause of it, There was such a shitstorm going on at one point that AMD literally cancelled a review shipment package. How he attacked Steve from HUB/Techspot about fake benchmarking the fx 8350 and that 720p benchmarks are wrong -> I literally lolled at that one. Also people saying he isn't an AMD fanboy. He used to have like 20 vids just about AMD and how shady the competition is - always bashing NVIDIA and Intel. People saying he shat on Vega -> watch that video again he was really displeased about it. Just listen to his voice when he talks about AMD and when he does about Intel/Nvidia. He has such pure hatred towards those 2 companies it's unbelliavable. He later even admits in one of his vids that he USED to be an AMD FANBOY-> USED to again had to rofl
The guy literally turned amd fanboyism into religion.
There's way more **** about this guy just google it. Too bad he deleted his original reddit account you could all see for yourself what a sweet guy he is.

edit: here's a few ss
https://I.imgur.com/npf7nw4.png
https://archive.fo/Y0EAa/a777d5ef0016b1b37f33bc58d8b9e119eb51ae08.png