Over the holiday weekend a leaked Geekbench benchmark result on an engineering sample AMD Zen processor got tech nerds talking. Other than the showcase that AMD presented a couple weeks back using the Blender render engine, the only information we have on performance claims come from AMD touting a "40% IPC increase" over the latest Bulldozer derivative.
The results from Geekbench show performance from a two physical processor system and a total of 64 cores running at 1.44 GHz. Obviously that clock speed is exceptionally low; AMD demoed Summit Ridge running at 3.0 GHz in the showcase mentioned above. But this does give us an interesting data point with which to do some performance extrapolation. If we assume perfect clock speed scaling, we can guess at performance levels that AMD Zen might see at various clocks.
I needed a quick comparison point and found this Geekbench result from a Xeon E7-8857 v2 running at 3.6 GHz. That is an Ivy Bridge based architecture and though the system has 48 cores, we are only going to a look at single threaded results to focus on the IPC story.
Obviously there are a ton of caveats with looking at data like this. It's possible that AMD Zen platform was running in a very sub-optimal condition. It's possible that the BIOS and motherboard weren't fully cache aware (though I would hope that wouldn't be the case this late in the game). It's possible that the Linux OS was somehow holding back performance of the Zen architecture and needs update. There are many reasons why you shouldn't consider this data a final decision yet; but that doesn't make it any less interesting to see.
In the two graphs below I divide the collection of single threaded results from Geekbench into two halves and there are three data points for each benchmark. The blue line represents the Xeon Ivy Bridge processor running at 3.6 GHz. The light green line shows the results from the AMD Zen processor running at 1.44 GHz as reported by Geekbench. The dark green line shows an extrapolated AMD Zen performance result with perfect scaling by frequency.
Continue reading our preview of AMD Zen single threaded performance!!
There are three tests in this benchmark, the AES encryption, Canny and Gaussian Blur, where the AMD Zen extrapolated results have an edge over Intel's Ivy Bridge. The AES results are probably the most compelling though we have definitely seen AES performance increases in Haswell, Broadwell and Skylake since this Xeon processor was released in early 2014.
Basically all of the other compute task based tests show the advantage going to Intel's Xeon processor running at 3.6 GHz. On average, if these results are at all indicative of what AMD Zen will be in the next 4 months, then it will be ~70-80% of the IPC of Ivy Bridge.
This relative performance graph removes the last three memory tests as it is very unlikely that memory bandwidth or latency will shift by simple linear scaling with clock speed.
What does this mean for AMD and the Zen architecture? It could mean absolutely nothing if the benchmark results we have here from a Zen CPU running at 1.44 GHz are simply not accurate, or not optimized. It might tell us that AMD and its partners have some work to do to optimize the platform in general for better overall performance, but with only months until stated product availability and partners already receiving samples this month, it looks less likely that we'll see dramatic changes from the numbers above.
This is Zen going up against Ivy Bridge and a Xeon processor that was released in Q1 of 2014. AMD needs to make its next CPU competitive not just against this but against Skylake-K and Broadwell-E systems if they want to impress the consumer and the press like they did back at IDF in August. Is it possible that AMD will be able to get Zen to 4.0 GHz and beyond in 8-core systems and thus will utilize clock speed to make up some of this IPC performance gap we show here? It's possible – though AMD needs to do it while maintaining processor thermals within a reasonable level. Doing some quick napkin math, in order for the AMD Zen results to swap wins and loses with the 3.6 GHz Xeon running on Ivy Bridge, this pre-production setup would have to run at around 4.8 GHz.
Again, I must reiterate, the results and graphs we have here are nowhere near finalized and could be an AMD red herring or the result of very early prototype hardware.
If the summer of 2016 was the season of GPUs, it looks like winter 2016/2017 is about do the same for CPUs.
You mean the AMD hype-machine
You mean the AMD hype-machine potentially over-promised and under-delivered? It’s almost like there’s a pattern developing. 😀
LOLERSKATES
LOLERSKATES
Interesting theory and
Interesting theory and opinions you have, but until real products are reviewed and benchmarked, your article will remain just that: theory and opinion.
Yes, exactly. In fact, I
Yes, exactly. In fact, I think that is stated precisely in the article.
Yup, had to make sure that
Yup, had to make sure that you understood that it is indeed an opinion.
I read the same article and I
I read the same article and I think your comment and response comes across as condescending toward Ryan. I don’t see how anyone could read the piece and suggest that he wasn’t aware that it was theoretical.
Let’s not get this all
Let’s not get this all skewed. I made a simple statement that he has interesting opinions, but that is all they are.
And he follows it up with a condenscending remark, and somehow you don’t have a problem with that, but you do with me giving it back to him? lol
“Yup, had to make sure that
“Yup, had to make sure that you understood that it is indeed an opinion”.
You didn’t have to do anything. But you just wanted to sound like smartass – that ended up looking like a moron.
Congratulations.
Oh look a disgruntled AMD
Oh look a disgruntled AMD fanboy.
Actually it’s hypothetical,
Actually it’s hypothetical, not theoretical.
What else would Ryan
What else would Ryan understand, given he wrote the article?
You can make sure by not
You can make sure by not being a condescending drone and actually read the article.
You couldn’t have been more
You couldn’t have been more clear. The reading comprehension problem on the internet is an epidemic.
I thought your piece was an excellent analysis based on known information and you didn’t utter one misleading or unsupported statement. Even the headline isn’t click bait.
Thanks!
Thanks!
Yeah man there wasn’t an
Yeah man there wasn’t an issue with the way you wrote this piece. Perfectly reasoned and unbiased.
+1 to what this dude said.
+1 to what this dude said.
That multi-core performance
That multi-core performance scaling looks more like quad and/or octa socket AMD server configuration where scaling per core drops below 50% or worse. I suspect that a single AMD “32-core” Naples chip is very likely 4x 8-cores dies using MCM (Multi-Chip Module), just like AMD’s previous high core count Opterons (linked together using HyperTransport). Thus with 2x AMD 32-core Naples that is equivalent to 8x 8-cores (for a total of 64 cores). Here is an example of quad socket server with 4x Opteron 6180 SE https://browser.primatelabs.com/v4/cpu/83804 48 cores, and again note the bad multi-core performance scaling. A single Opteron 6180 SE is actually 2x 6-cores die using MCM. Thus with 4x Opteron 6180 SE that is equivalent to 8x 6-cores (for a total of 48 cores).
Passive aggressive AMD
Passive aggressive AMD fanboy? Most likely.
Another way to look at Zen is
Another way to look at Zen is that it’s going to have at least 40% higher IPC than Excavator; but likely 10% less clock speed (maybe even lower). A better way might be to extrapolate off of that to see how competitive it is.. 1.3 * the AMD A12-9800.
Geekbench is a horrible benchmark, but adding this to the extrapolation of Ashes of The Singularity, it’s starting to feel like it’s not going to match 4790K, and maybe not even match 4770K in many scenarios.. Let alone Sky or Kabylake.
Other than being synthetic in
Other than being synthetic in nature, why do you consider Geekbench to be a "horrible benchmark"?
The way I wrote that – I
The way I wrote that – I deserve a call out.
According to Linus Torvalds, many sub-tests of Geekbench fit within the L1 cache, meaning that it’s not really any more useful for predicting real world app performance than dhrystone. (A larger L1 cache will always win.. almost regardless of the rest of the CPU).
Linus Torvald’s opinion on Geekbench:
http://www.realworldtech.com/forum/?threadid=136526&curpostid=136666
My main point was that using the recently released A12-9800 AMD chip should allow for significantly more accurate extrapolation on performance than Geekbench clock scaling.
That part is still 28nm and
That part is still 28nm and still based on excavator its just updated for the new platform.
True, but AMD has publicly
True, but AMD has publicly said Zen’s IPC will be 40% higher than Excavator, so that’s something concrete we can work with.
The only benchmark I’m
The only benchmark I’m interested in is one using clang compile times.
This is completely new
This is completely new Geekbench 4 with new tests etc. Should not be compared to Geekbench 3.
I don’t think this Geekbench
I don’t think this Geekbench test says a lot. I don’t even think the clock speed was reported correctly. Look at the sample identification, it starts with “2”. Which means it’s a prototype sample. Remember the AotS scores? Sample “2” scores were way lower than sample “1” scores.
I really hope for
I really hope for consumer/competition sake that this does not accurately reflect the finalized performance of Zen, otherwise Zen is dead of arrival.
From the demo’s that AMD was showing, ZEN 8 Core vs a Broadwell-E 8 Core both running at 3.0 Ghz, it seemed like performance was nearly identical. Granted that was only running in one test, but I would think AMD knows they have a lot riding on Zen and that having another Bulldozer is almost a death sentence.
IMHO, I believe the end real world performance of Zen will be between Broadwell and Skylake, with trade offs in different benchmarks. If AMD can pull that level of performance and offer good prices, they will have a real winner.
That’s great let the
That’s great let the sycophants worry about the IPC metric only! If Zen is between the Broadwell and Skylake levels of IPC then it will be good enough for more than just gaming! Add to that Zen APUs with Polaris graphics and there will be more value added in the graphics part of the APU equation from AMD. Price/Performance will be what most want from AMD.
Fun read but with so many
Fun read but with so many assumptions and extrapolations I can’t really believe this will be representative of final product.
Probably should have stated
Probably should have stated that at the start of the article.. i can already see bunch of fanboys posting these everywhere and crap like this hurts sales.
Over the years pcper have
Over the years pcper have been fairly accurate on a lot of things.I don’t expect any surprise from amd .end result will be within 15% of what pcper napkin prediction brought here. A die shrink help whats already awsome.this look to be a server chip tho. So this naptin result as zero % chance of accuracy.
Server CPUs have more in
Server CPUs have more in common with client than you might think.
Yes for an engineering
Yes for an engineering sample, and mostly there for early testing and software development it’s not so bad, let’s go and get some of Intel’s early engineering samples runs from the same benchmarking databases and see what the Intel engineering samples do at lower clocks. The real benchmarks are going to have to wait until there are really the final steppings and UEFI/BIOS and optimized OS/Software for Zen after the product is on the market and available to be randomly selected/purchased and then tested properly.
I’m only willing to take as proper testing some Zen 8 core systems meant for the consumer desktop market, and not any meant for other markets. Of course there will be the Price/performance metric for the consumer Zen SKUs against any of Intel’s newest CPU/SOC SKUs for actual gaming. AMD’s Polaris graphics against Intel’s IRIS Pro, at the same SKU price point will be important also, as Intel never appears to be offering its highest tiered graphics in its affordable range of products. I do not expect that Intel will be able to compete with AMD’s or Nvidia’s graphics and integrated Polaris graphics at 14nm will have more ACE units in use, so even Intel’s top end SOC graphics will have to compete with AMD’s Polaris on the APU/SOC graphics front.
AMD would only have to field a Zen/Polaris APU with a single HBM2 stack to really put Intel’s integrated graphics in the rear view mirror. So AMD could create a consumer high end Zen/Polaris APU on an interposer with one stack of HBM2(4GB) configured to act as a L4 cache with the remainder of the memory supplied by DDR4 DRAMs in a single or dual channel configuration while leveraging the HBM2 to provide the Polaris integrated graphics with plenty of bandwidth.
One thing is certain even if a Zen/Polaris/HBM2(Singe HBM2 stack/L4) APU was only supplied with one DDR4 DIMM memory channel by the OEMs, the APU’s HBM2 would do very well at hiding and single channel limited bandwidth/latency, as the HBM2/L4 cache would still be providing its bandwidth for textures and game/gaming engine code to run from. The HBM2 die would have its own 1024 bit dedicated channel to the Zen APU’s CPU cores with the regular DDR4 DIMM based DRAM acting as a second tier lower bandwidth RAM store, so single or dual DIMM DRAM channel would not have any detrimental effect on the APU’s integrated graphics bandwidth needs.
The OEM’s would have to figure out some other way to gimp a Zen/Polaris/HBM2 APU’s graphics performance relative to what they offer for Intel’s pricy kit under Intel’s nefarious incentive packages(Brown envelops with green inside).
By the time you get to the
By the time you get to the engineering sample stage, IPC isn't going to change much. The further tweaks after ES are just that, tweaks. Stepping optimizations to help increase yields and clocks. Changing the IPC of a CPU requires *big* changes to the layout. You don't just flip one transistor somewhere and suddenly some type of operation takes fewer clocks to complete.
Yes but there are also the
Yes but there are also the errata that may be there in an engineering sample that can not be fixed with a miro-code/other-code UEFI/BIOS patch that may require some amended tape out design modification to resolve. So I’ll not place my complete trust in any engineering sample benchmarks. I’ll be looking for the Price/performance metric more than the overall IPC winner, and non gaming Graphics performance will play a bigger part in my final choice for any APU/SOC. I look at non gaming rendering performance more than gaming rendering performance and FPS. So maybe I’ll look at the total numbers of Ray interaction calculations that can be accelerated on the GPU’s FPUs per unit time rather than any FPS metrics that is more a function of ROPs.
I’ll also look to see the total numbers of shader cores, as that will let me have better 3d graphics editor usability/performance in programs like Blender 3D as opposed to simply FPS for games. Intel’s graphics does not play well with high polygon count mesh models in 3D edit mode in many graphics packages, simply because Intel lacks the total numbers of shader cores compared to AMD’s or Nvidia’s graphics. So for non gaming rendering a Zen/Polaris APU may be able to handle large polygon count mesh models in Blender’s 3D editing mode than Intel’s tuned for gaming only(Low polygon count gaming meshes) integrated graphics. Those total shader core counts from AMD and Nvidia offer more FP units for Ray Tracing accelerated on the GPU’s cores also, compared to Intel’s integrated graphics. AMD’s Polaris GCN GPUs have more FP resources/$ than Nvidia’s or Intel(Very few shader cores relative AMD’s total numbers) so on Polaris there are more FP units to accelerate Ray Tracing workloads on the GPU.
People expectarion are high
People expectarion are high in gaming. Ms , Sony and Nintendo have been all tweaking , so hopefully this will help them there.if twitch streamer love it , the rest will follow. Yep twitch streaer and YouTube game streamer will be the real story. People with brand new pc are complaining left and right with intel and NVidia inside .example how come 144hertz slowdown to 60hertz in window full screen world of warcraft etc)) faster rarelly translate to succesfull system on twitch. And now a day twitch gamer stream and YouTube vamer stream is the important metric and you vet to show it off too))
I thought I’d put my 2 cents
I thought I’d put my 2 cents in as someone that has built servers for corporate applications. Low clock speeds and high core counts are completely normal in this kind of appkication
Depending on the chip the single threaded performance may be sub par depending on what the purpose of the chip is. For example xeon e5 chips have poor single thread performance compared to the xeon e3. It’s simply because they have different target markets.
This is not the same AMD that
This is not the same AMD that made such bad choices 5-7 years ago. Different people, different philosophy.
For example the Windows NT6 kernel was great, yet the shell team screwed up with Vista royally. Then another team with different people made the Windows 7 shell and everyone loved it.
Give it some time, and see what comes out. It only really matters to the people buying/building systems and AMD stockholders anyway.
I thought I read somewhere
I thought I read somewhere that 1.4Ghz was just the base clock? And how come it doesn’t list any L3 cache? I think there are way to many assumptions being taken to draw anything meaningful from this.
It might be interesting to
It might be interesting to lock a Broadwell-E to 1.4GHz to provide another datapoint.
That Zen processor had 0 KB
That Zen processor had 0 KB L3 cache.
Thus, the overall performance is not indicative of final product.
A 12 core CPU has an unfair
A 12 core CPU has an unfair advantage against a 32 core CPU on single-core performance. Low core count CPUs always have better single-thread performance than their high core count counterparts. I think you should have factored this in that extrapolation.
I just want to know, how do
I just want to know, how do you get off extrapolating Zen at 3.6GHz? You’re literally just making shit up.
I think you’ve gravely
I think you’ve gravely misunderstood the point of this article.
ANY extrapolation they did would be ‘making shit up’, they ‘get off’ on it because the data they have for the Xeon they are making the comparison with is at that clock speed giving them a verified basis on which to do the ‘napkin math’.
How else would you make an IPC/single-core performance comparison otherwise in lieu of all the details?
You know you are right.
You know you are right. Technically speaking performance should scale linearly, that is for example, double clock speed and, all things being equal, double the amount of work done over the same time period. Unfortunately, not all things are equal, different cores clock and turbo on different clock domains. Memory clock and data rates are on different domains than the core clock, so in fact, his extrapolation is an over estimate, it will likely be slower that what is estimated here…. You’re right he ‘just made this shit up’. It will be worse.
You gotta love the AMD fanboy. They are so entertaining.
It’s dirty math, and assumes
It’s dirty math, and assumes perfect frequency scaling, so naturally it’s not going to be an exact calculation. But for ballpark guesstimates, it’s not all that difficult.
Here’s basically how it works. We know that these are single-threaded benchmarks, so we can automatically throw the core count out the window; we’re looking for per-core IPC and the benchmark only works one core/thread, so we’re already halfway there.
Take the first result on the first picture, “AES”. At 1.44 GHz, one core/thread of this alleged Zen CPU scores 1550. We’d like to find out how it might score at 3.6 GHz.
(1550 / 1.44) = (N / 3.60) – now solve for N.
(1550 / 1.44) = 1076.388888~
Thus,
(1076.3888888~) = (N / 3.60)
(1076.3888888~) x 3.60 = (N / 3.60) x 3.60
3875 = N
Therefore, assuming perfect frequency scaling, one would expect a single core/thread of this alleged Zen CPU at 3.6 GHz should score somewhere around 3875.
No, it’s not a perfect calculation. But in the absence of any other data, it’s a decent way to get a ballpark guess as to what COULD be. Which is why Ryan made that very clear at the beginning of the article.
(note – I am a self-admitted AMD fanboy.)
They already have the samples
They already have the samples running at 3ghz, lets hope they can deliver something around 3.8ghz with 6 cores 12 threads for less than $300, they will have no issues selling such a product.
Low level API’s and higher resolutions will render most cpus mute going forwards anyway, they will become more like memory for most of us, just a component that is needed to make the rest of the rig run.
Vega is where its at, for better or worst.
ivy bridge is xeon e5 v1
ivy bridge is xeon e5 v1
as long as it matches the ipc
as long as it matches the ipc of a i3 6100 with 4 real cores or more at 20-30$ less dollars
it will kick ass
ZEN will have ~twice the IPC
ZEN will have ~twice the IPC of Jaguar and ~twice the clock.
With the same number of real cores ZEN will allow to game at 120fps (four times 30fps = console rate, HT to handle OS overhead).
What matters really is price and efficiency.