Zen Theoritical performance.

[Vogner16]

Per discussion on AMD cpu thread I decided to do some testing to find what the real performance of amd's next gen Zen cpu processors will be capable of in real comparable scenarios.

http://imgur.com/a/Qe3uo
http://imgur.com/gallery/Qe3uo/new
Those may be the same link

summary of this testing is

Cinebench R15

ST: 83
MT:538

3dmark 11 physics and combo (with 7970 overclocked 1060)

Phy:6277
Combo:5982

Test machine was as in the pictures amd FX 8350 @ 3.52Ghz with 16gb 1866mhz ram

First round I simply changed multiplier to 17.5 from my overclock.

second round I lowered voltages to auto and enabled turbo core in bios.
It is worth noting that I didn't see the core clock change during single thread testing and is the reason I kept the score of 83 even when enabling the core to clock to 4.2 with turbo mode.

AMD claims Zen will reach 40% IPC improvement over excavator. We believe zen's SMT design is possible to reach 3.5Ghz as a base clock. giving this comparision ground.

as excavator is claimed to be 10% IPC improvement over Piledriver we assume 50% improvement in IPC over the 8350 in my test machine.

To begin with Cinebench R15

ST performance 1.5 * 83 = 124
MT performance 1.5 * 538 = 807

This assumes the comparable zen part will be a quad core 8 thread like its predecessor.

now for 3dmark 11

phy: 1.5* 6277 = 9415.5
Combo: 1.5 * 5982 = 8973 ( I know GPU has a large part to do with this, but why not)

From the pictures you can see haswell ST can reach 165 @ 4.4Ghz so our score of 124 isn't really that far off.

the 3770K clocked at 3.4 (with turbo) only gets a score of 138.

I am confident after this testing that Zen will produce single core performance with the help of turbo core enabled close to that of haswell cpu's when placed clock per clock.

When you factor in the 16 threads zen will have this leaves us with one might beast (in theory)

Will zen live up the hype? you decide!

 

[synphul]

I don't put much faith in speculation. The ipc improvements are just that for now, theory. Devil's canyon i5's were theorized to hit 5ghz on air cooling and reality said otherwise. There's also not much faith to be had in synthetic benchmarks, often times the fx 8350 scores higher in things like passmark yet the i5's outperform them in just about every category in the real world. Contradictory to synthetic tests.

Not to say zen will for sure be a flop but far too early to call it a success either. Speculation is much like daydreaming and the only thing that really matters is when a chip from either camp is in hand, in a rig and putting out measurable performance in programs people actually use.

Some of their proposed tech sounds good but as far as getting all excited over 8 cores and 16 threads I think it's also important to realize how many times a simple 4c/4t cpu like an i5 bests the current fx 8c/8t design. If twice the cores/threads aren't getting the job done the majority of the time adding more of the same isn't going to have much impact. So many programs are still failing to make efficient use of 8 threads and it's only in a few special circumstances that all 8 threads get to really shine for amd. Doubling those threads only goes further down the path of diminishing returns.

It's not just amd, it's just the way things are. For instance in games, most of the time with the exception of maybe 1 or 2, the i5 produces the same fps as a quad core i7. The i7 having twice the thread count. If a hyperthreaded quad core isn't producing even close to 80-90% more performance, imagine how fast the diminishing returns would be on an i7 5960x with 8 cores and 16 threads. With 8c/16t on a cpu with proven high ipc filling such a small niche it's easy to say the same for amd with 8c/16t.

Unfortunately it sounds like more sales hype and gimmicks waiting for us. Even though the tech may be legit it's simply not needed. Adding 3 or 4 more stereos to a car that already has one doesn't automatically make it better, it just means it contains useless amenities.

 

[bicycle_repair_man]

I'm with synphul on this; everything is just speculation and should be taken with a truck-load of salt. I remember when the FX-8xxx series were hyped to the hills, but when they were released, they were a disappointment. This was shortly after Intel had struck it big with Sandy Bridge, so that didn't do AMD any favours either.

I hope that Zen turns out to be a success for AMD; competition is good for the consumer and nobody wants Intel to dominate the market for desktop microprocessors. Focusing on cores and threads, as synphul said, is largely irrelevant. What AMD need to do is focus on IPC and TDP. If they can manufacture a CPU with strong IPC, a sub-100W TDP, 8 threads and a reasonable price tag, then I reckon that a lot of people will start to listen. I hope this happens, but going on past experience, my gut feeling is that Zen will disappoint.

 

[synphul]

Well they already have 8c/8t models which is more than enough for heavily threaded situations. It's not a matter of pre judgement on amd, it's just that companies have burned folks in the past with all the hype. People don't want to see a bunch of multicolored charts and graphs with all these wild prediction arrows of performance that don't scale to the graph. (Notice how the 40% ipc improvement on the amd adverts show an arrow rising over 100% for dramatic effect?)

Percentage improvements sound all well and good but gets hopes up and then in reality they end up saying oh well, I guess it only improves game xyz by 2fps or I guess it only shaves 3 seconds off a video encoding project. People feel let down. I'd much rather a company, whether amd, intel, nvidia or whoever just quietly say 'hey, our new product is out' and then see benchmarks with 15-20fps improvements. They wouldn't have to make a huge deal about it, people would be shouting it all over the web.

Going back to the heavy threading, I think zen would do well to concentrate on ipc improvements and core efficiency/productivity. If they're going with an smt model similar to hyper threading then 4c/8t with stronger ipc performance would compete well against i7's. Already talking about 8c/16t models makes me think they might be going back to their old ways of 'if you can't make the cores strong enough just add more of them' and it's a design that hasn't panned out for them. I don't have any proof of it, it could turn out to be a killer chip. More that history has a tendency to repeat itself. Seeing zen topping at even 6c/12t would lead me to believe they're actually confident about their ipc improvements.

 

[gamerk316]

I worked out the math in the AMD sticky thread. If you assume clocks remain at @4GHz and 40% IPC is the "typical" case, and removing the 20% scaling hit you take with CMT, I came out to a 50.75% theoretical overall performance gain, disregarding the SMT implementation. If you assume clocks are lowered to the 3.5GHz range, and 20% IPC is the typical case, I came out to 23% performance gain. So depending on the exact details, between 20% and 50% gain over PD is likely.

 

[RobCrezz]

I hope its better than your expected results to be honest. IF it can only match Haswell in IPC, then unless its very cheap its already going to be behind in the performance stakes, by the time its released intel will be on Kaby Lake if not Cannonlake...

 

[LookItsRain]

I would really like to see amd pull this off, as sandy/ivy bridge level ipc is still good enough for games today, add on the extra core/thread count at a great price and you have an excellent cpu....but i wont be making plans based on speculation alone. Intels recent tick tock cycles have been underwhelming for the desktop performance market.

 

[Vogner16]

Glad to see everyone is interested in this topic.

I agree 8350 scales poorly from real world to synthetic benchmarks but I think it has a lot to do with its base design.

As the 8350 is a CMT design amd will always be behind with single thread performance compared to intel's SMT design. If you read the Amd sticky we have gone over this in great detail looking at the block diagram of zen to what we know of amd's claims of performance improvements.

While you are all correct in the fact of amd's consistant overhyping of new cpu's ever since the first Athlon being a wild success, I have faith that zen will pick up the massive gap in performance between current CMT style cores and intels 6th generation of SMT cores.

As lookitsrain said Sandy and Ivy both have amazing IPC for todays games despite being 4 years old. If amd is to match the IPC of even the super old Sandy and maintains the clock speed of sandy, with twice the cores allows for both great gaming experience for single thread games that piledriver suffered at and massively multithreaded tasks such as handbrake and photoshop photo compression.

Time will tell, but For once I believe that amd will deliver on its promices. they don't claim to be faster than skylake or kaby lake which will be released at the same time, but for the price of a zen cpu, amd should deliver a excellent cpu for 90% of tasks.

I gain this from what amd claims zen will improve and how zen is physically structured.

As an amd fan I hope that this cpu will return them to the high end computing market and bring intel's prices down to reality. If you guys haven't noticed the 6700K is the most expensive launch price for a mainstream I7 in the history of core I series. Nobody wants to see what their performance monolopy will do to prices. I don't want to pay $3000 for a highish performance computer in 3 years time just like I had to in 1995. If zen is what it looks like it will be competition can keep that price down to $1500.

 

[salgado18]

Per discussion on AMD cpu thread I decided to do some testing to find what the real performance of amd's next gen Zen cpu processors will be capable of in real comparable scenarios.

http://imgur.com/a/Qe3uo
http://imgur.com/gallery/Qe3uo/new
Those may be the same link

summary of this testing is

Cinebench R15

ST: 83
MT:538

3dmark 11 physics and combo (with 7970 overclocked 1060)

Phy:6277
Combo:5982

Test machine was as in the pictures amd FX 8350 @ 3.52Ghz with 16gb 1866mhz ram

First round I simply changed multiplier to 17.5 from my overclock.

second round I lowered voltages to auto and enabled turbo core in bios.
It is worth noting that I didn't see the core clock change during single thread testing and is the reason I kept the score of 83 even when enabling the core to clock to 4.2 with turbo mode.

AMD claims Zen will reach 40% IPC improvement over excavator. We believe zen's SMT design is possible to reach 3.5Ghz as a base clock. giving this comparision ground.

as excavator is claimed to be 10% IPC improvement over Piledriver we assume 50% improvement in IPC over the 8350 in my test machine.

To begin with Cinebench R15

ST performance 1.5 * 83 = 124
MT performance 1.5 * 538 = 807

This assumes the comparable zen part will be a quad core 8 thread like its predecessor.

now for 3dmark 11

phy: 1.5* 6277 = 9415.5
Combo: 1.5 * 5982 = 8973 ( I know GPU has a large part to do with this, but why not)

From the pictures you can see haswell ST can reach 165 @ 4.4Ghz so our score of 124 isn't really that far off.

the 3770K clocked at 3.4 (with turbo) only gets a score of 138.

I am confident after this testing that Zen will produce single core performance with the help of turbo core enabled close to that of haswell cpu's when placed clock per clock.

When you factor in the 16 threads zen will have this leaves us with one might beast (in theory)

Will zen live up the hype? you decide!

Nice tests, but I believe it missed one very important point: what is the performance of Ivy Bridge and Haswell at that same clocks? There has to be some articles around that compare them, and then you could match their clocks, test and add 50% for a good Zen (speculative) comparison.

EDIT: this article has comparison of the past four six-core cpus from intel, all on single-thread Cinebench R15, all on the same 3.2GHz clock. Maybe you could do the test at 3.2 to compare with them?

Cinebench R15, single thread, all CPUs at 3.2 GHz
990X 98
3960X 111
4960X 117
5960X 127

http://www.anandtech.com/show/8426/the-intel-haswell-e-cpu-review-core-i7-5960x-i7-5930k-i7-5820k-tested/2

 

[Vogner16]

Nice tests, but I believe it missed one very important point: what is the performance of Ivy Bridge and Haswell at that same clocks? There has to be some articles around that compare them, and then you could match their clocks, test and add 50% for a good Zen (speculative) comparison.

EDIT: this article has comparison of the past four six-core cpus from intel, all on single-thread Cinebench R15, all on the same 3.2GHz clock. Maybe you could do the test at 3.2 to compare with them?

Cinebench R15, single thread, all CPUs at 3.2 GHz
990X 98
3960X 111
4960X 117
5960X 127

http://www.anandtech.com/show/8426/the-intel-haswell-e-cpu-review-core-i7-5960x-i7-5930k-i7-5820k-tested/2

I had the cinebench comparision photo in that imigr post. showed 4.4ghz haswell and 3.4 ghz ivy.

I don't have either cpu so I cant test them at 3.5 to get a full compare

that Ivy however is pretty close

Claims to get 138 points on ivy in ST test. that's pretty close to the calculated 124 I found. at neer identical clocks.
Remember I didn't account for the CMT to SMT improvement but as im unsure what the % improvement would be I don't know where to put it. gamerk claims smt will improve scores over cmt by 20%.

if we normalize both zen theoretical and ivy to be at 3.5 we calculate zen to be 124 and ivy to be (140-143)

assuming 0% improvement from the cmt to smt switch zen will have the single core power close to sandy bridge.

assuming 5% from CMT to SMT we get 130.2
assuming 10% 136.4
assuming 15% 142
assuming what gamerk said smt will gain us of 20% we have ST score of 148.8 at 3.5 ghz.

148 places zen slighty above ivy bridge in IPC but slightly below haswell, as haswell at 3.5 is said to get roughly 151.

account for turbo core to hit 3.9??? we really have some good potential.

 

[gamerk316]

assuming what gamerk said smt will gain us of 20% we have ST score of 148.8 at 3.5 ghz.

Assuming no other inherent bottlenecks in the SMT design, yes. The PD scaling factor is about 6.6x the performance when all 8 cores are used, so there's about a 17.5% performance loss due to CMT. If we assume perfect scaling, you gain most of that back. Note scaling will be less then perfect, but I could see a 15% performance boost in multithreaded workloads simply due to better scaling across cores.

 

[gamerk316]

I'm going to do the math slightly differently:

Performance = (Clock * IPC) * NumberOfCores

Taking your baseline:

Cinebench R15

ST: 83
MT:538

I can solve for IPC for this specific benchmark. Let's do the single-core case first:

83 = IPC * 3.52
IPC = 83/3.52
IPC = ~23.58

And multithreaded:

538 = 3.52 * IPC * 8
538 = 28.16 * IPC
IPC = ~19.11

The lower IPC is due to CMTs poor scaling. You would expect performance of 83 * 8 = 664, but only get 538, a loss of 19%, which is about what you'd expect for CMT. Zen will likely reduce this somewhat, but we can't know by how much until we get real benchmark results.

Let's assume clocks for Zen remain at about 3.52GHz at the top for now, and let's disregard SMT for now since it makes the math harder. We'll assume a 30% IPC boost, taking the halfway between what AMD states, and what pessimists believe. We'll use this to solve for Zen's theoretical performance.

Performance = 3.52 * (23.58 + (23.58 * .3))
Performance = 3.52 * 30.636
Performance = ~107.84 rounds to 108

Again, we don't know for what degree CMT's limitations are removed, so these results will likely be on the low side. In addition, I'm not factoring in SMT since I don't know how well it will scale, so I can't account for it. So these results are also understated due to not factoring in SMT.

Performance = 3.52 * (19.11 + (19.11 * .3)) * 8
Performance = 3.52 * 24.843 * 8
Performance = 699.57 rounds to 700

Comparing to PD:

Single thread: 83 vs 108 (26.178% improvement)
Multi-thread: 538 vs 700 (26.1712% improvement)

I again note the MT numbers are on the low end, due to not factoring in the expected improvement to core scaling, and not factoring in SMT.

Now, someone get numbers for IB/Haswell/Skylake at a 3,5GHz clock, and compare, and you get your theoretical head to head comparison with Zen.

 

[salgado18]

I'm going to do the math slightly differently:

Performance = (Clock * IPC) * NumberOfCores

Taking your baseline:

Cinebench R15

ST: 83
MT:538

I can solve for IPC for this specific benchmark. Let's do the single-core case first:

83 = IPC * 3.52
IPC = 83/3.52
IPC = ~23.58

And multithreaded:

538 = 3.52 * IPC * 8
538 = 28.16 * IPC
IPC = ~19.11

The lower IPC is due to CMTs poor scaling. You would expect performance of 83 * 8 = 664, but only get 538, a loss of 19%, which is about what you'd expect for CMT. Zen will likely reduce this somewhat, but we can't know by how much until we get real benchmark results.

Let's assume clocks for Zen remain at about 3.52GHz at the top for now, and let's disregard SMT for now since it makes the math harder. We'll assume a 30% IPC boost, taking the halfway between what AMD states, and what pessimists believe. We'll use this to solve for Zen's theoretical performance.

Performance = 3.52 * (23.58 + (23.58 * .3))
Performance = 3.52 * 30.636
Performance = ~107.84 rounds to 108

Again, we don't know for what degree CMT's limitations are removed, so these results will likely be on the low side. In addition, I'm not factoring in SMT since I don't know how well it will scale, so I can't account for it. So these results are also understated due to not factoring in SMT.

Performance = 3.52 * (19.11 + (19.11 * .3)) * 8
Performance = 3.52 * 24.843 * 8
Performance = 699.57 rounds to 700

Comparing to PD:

Single thread: 83 vs 108 (26.178% improvement)
Multi-thread: 538 vs 700 (26.1712% improvement)

I again note the MT numbers are on the low end, due to not factoring in the expected improvement to core scaling, and not factoring in SMT.

Now, someone get numbers for IB/Haswell/Skylake at a 3,5GHz clock, and compare, and you get your theoretical head to head comparison with Zen.

I just got you every Intel since Nehalem at 3.2GHz

Don't know if it can be easily compared, but Bulldozer (FX 8120) did 71 points in single-thread at 3.2GHz. I'm going to extrapolate, and assume a 60% improvement over Bulldozer (I believe it's more, but let's stick to it).

So, Zen at 3.2 GHz could do around 114 points. Comparing with the others, it is right in the middle between Ivy Bridge and Haswell.

I won't compare multi-threaded works, because the switch from CMT to SMT will affect the scaling in an unpredictable way. But single threaded performance doesn't look that bad.

Of course, I should do three passes and average, but I ran out of time. Will do it again later, and post here the results.

 

[Vogner16]

So, Zen at 3.2 GHz could do around 114 points. Comparing with the others, it is right in the middle between Ivy Bridge and Haswell.

While using different cpu I arrived at the same conclusion. gamerk316 you keep using IPC gains lower than what amd claims by a significant margin. 30% ipc improvement over PD is simply too low. I have never seen a company claim a 20% higher performance than what they get real world. At the least you should use 40% improvement over PD.

EDIT: for example Intel claimed 5% improvement in IPC over broadwell and they got 2.7%. that's a difference of 2.3% IPC, much less than the 20% less you are using if amd claims 40% over excavator (aka 50% over PD), then they will at least get 45% over PD and worst case scenario 40% improvement over PD.

 

[Vogner16]

Performance = 3.52 * (23.58 + (23.58 * .4))
Performance = 3.52 * 33.012
Performance = ~116.2 rounds to 116

Again, we don't know for what degree CMT's limitations are removed, so these results will likely be on the low side. In addition, I'm not factoring in SMT since I don't know how well it will scale, so I can't account for it. So these results are also understated due to not factoring in SMT.

Performance = 3.52 * (19.11 + (19.11 * .4)) * 8
Performance = 3.52 * 26.754 * 8
Performance = 753.39 rounds to 754

Comparing to PD:

Single thread: 83 vs 116 (39.759% improvement)
Multi-thread: 538 vs 754 (40.148% improvement)

Equaling amd's claims of 40% IPC improvement albeit not over excavator. the extra 10% it needs may come from SMT

 

[Vogner16]

the full 50% improvement would give us score of

this is where I include 10% for SMT

Performance= 116 * 1.1 = 127.6
Performance= 754 * 1.1 = 829.4

 

[Suzuki@1473780084@1473780163]

We don't know yet (I think?). But regardless, Intel likely can just up their performance as they release slow upgrades to make as much money as possible. I'm curious to see how Zen performs but I don't think it's going to be something spectacular.

 

[jdwii]

Not to be that guy but IPC isn't exactly the end all of everything, things like SSE tend to do better on intel CPU's compared to Amd. When i hear IPC i now say what instruction set? In some cases IPC might only be 15% faster in others 60%.

 

[Cazalan]

One of the more interesting things I heard lately is AMD will continue the Carrizo trend and have lower base clocks but with much higher turbos (1-2 cores). This would fit nicely with the DX12 profiles I've seen where only 2 cores really max out. Considering what the PS4 and XBone can do with 8 wimpy Jaguar cores, the prospect of 8 Zen cores at 2.8Ghz(3.6-4Ghz Turbo) would be a tremendous upgrade.

I don't think AMD will be competing with the Intel E series or even the higher clocked i7s. What they could do however is offer a healthy competition in the lower range of the i5/i7 series where the bulk of sales are made. As enthusiasts most are only thinking the creme of the crop. The i7-6700K 4Ghz(4.2Ghz Turbo) but if you look out there the bulk of the systems are made with the lower speed T or TE version. i7-6700T 2.8Ghz(3.6Ghz Turbo) or i7-6700TE 2.4Ghz(3.4Ghz Turbo). The same goes for the i5 field.

I don't expect miracles from Zen but they could certainly regain some market share with what has been proposed thus far.

 

[salgado18]

One of the more interesting things I heard lately is AMD will continue the Carrizo trend and have lower base clocks but with much higher turbos (1-2 cores). This would fit nicely with the DX12 profiles I've seen where only 2 cores really max out. Considering what the PS4 and XBone can do with 8 wimpy Jaguar cores, the prospect of 8 Zen cores at 2.8Ghz(3.6-4Ghz Turbo) would be a tremendous upgrade.

I don't think AMD will be competing with the Intel E series or even the higher clocked i7s. What they could do however is offer a healthy competition in the lower range of the i5/i7 series where the bulk of sales are made. As enthusiasts most are only thinking the creme of the crop. The i7-6700K 4Ghz(4.2Ghz Turbo) but if you look out there the bulk of the systems are made with the lower speed T or TE version. i7-6700T 2.8Ghz(3.6Ghz Turbo) or i7-6700TE 2.4Ghz(3.4Ghz Turbo). The same goes for the i5 field.

I don't expect miracles from Zen but they could certainly regain some market share with what has been proposed thus far.

Remember that Zen 8 cores is not going against 4 core i7s. IPC will be close enough that a 6-core Zen will be faster than a 4-core i7.

Also, if they have an 8-core Zen with 2.8-3.0 GHz base, and 3.5-4.0 GHz turbo, they will be about 15-20% slower than the top-of-the-line i7-5960x, which has 3.0 base and 3.5 turbo.

Everyone is talking about them reaching lower clocks, but forget that Intel's high core count CPUs have lower clocks as well. We can definitely see a 4-core Zen at the 3.6-4.0 base clock, which would almost compete with the i7-6700k, and probably cheaper due to the lower IPC.

Zen will be a great product if they reach the 40% IPC over Excavator, and manage to make the 8-core an 125-140W part.

 

[RobCrezz]

Is this all assuming Intel will not be advancing before then though? if they match haswell then they will still be 2-3 generations behind intel. That said if they are cheap, have 8 cores and 16 threads and match Haswell IPC, I will be buying one for sure!

 

[LookItsRain]

Is this all assuming Intel will not be advancing before then though? if they match haswell then they will still be 2-3 generations behind intel. That said if they are cheap, have 8 cores and 16 threads and match Haswell IPC, I will be buying one for sure!

Exactly, ivy/haswell level ipc is still more than enough for games, but add extra cores for a great price onto that, ill be buying one.

 

[gamerk316]

So, Zen at 3.2 GHz could do around 114 points. Comparing with the others, it is right in the middle between Ivy Bridge and Haswell.

While using different cpu I arrived at the same conclusion. gamerk316 you keep using IPC gains lower than what amd claims by a significant margin. 30% ipc improvement over PD is simply too low. I have never seen a company claim a 20% higher performance than what they get real world. At the least you should use 40% improvement over PD.

EDIT: for example Intel claimed 5% improvement in IPC over broadwell and they got 2.7%. that's a difference of 2.3% IPC, much less than the 20% less you are using if amd claims 40% over excavator (aka 50% over PD), then they will at least get 45% over PD and worst case scenario 40% improvement over PD.

AMD has had a recent history of using "best case" numbers when talking about performance gains, going back to the Phenom days. Why do you think we still get charts that says "Up to 150% performance improvements!" every time we get a new GPU on the market? As a general rule, if a company claims improvements of X, the "typical" improvement will end up being X/2. I'm giving AMD 10% more improvement then I think they're actually going to achieve.

I fully expect Zen to be a benchmark monster, just like BD was. I also expect, in typical usage, Zen will lag current Intel CPUs, just like BD did.

 

[RobCrezz]

BD benchmark monster? Only in the sense that it was bad.. Was only "ok" in some multithreadded benches.

 

[salgado18]

BD benchmark monster? Only in the sense that it was bad.. Was only "ok" in some multithreadded benches.

I think he means that, in threaded benchmarks, BD was able to rival Sandy Bridge i7, but in everyday use, like games and productivity software, it lagged behind badly.

I bought an FX 8120, overclocked it to 4.0 GHz, and tested a test scene in Blender. Then I took the scene to a friend with an i5-2500k, overclocked to 4.5 GHz, and tested again. The Bulldozer was around 20% faster in that test.

Then we got to Skyrim, and I couldn't go past 50 fps, while he had over 70. :/