AMD Unveils Zen Microarchitecture, Demos Summit Ridge Performance

Page 5 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.
Ok, you might wanna specify what you mean by "better at 3D applications" because that's a broad category, and saying the AMD was better at it then Intel for the past few years is generally false. And no, a $200 AMD CPU didn't get you the same performance as a $1000 Intel model. Look back at those benchmarks I linked and also this one. A $1000 i7-3970X @ 3.5 GHz blows away the $180 FX-8350 @ 4.0 GHz. The FX also loses to the $340 i7-3770K @ 3.5 GHz. The only Intel chip the FX consistently beats is the $240 i5-3570K @ 3.5 GHz. But if you compare it to another $240 Intel chip at the time, the 4C/8T Xeon E3-1230v3 @ 3.3 GHz, the FX would lose to that as well.

So the FX was losing in 3D applications even though it had a clock speed advantage. If they were running at the same clock speed, the AMD chips would be anywhere from 15% - 25% slower than the i7. The FX chips weren't magically 3D killer chips back then. The advantage only came if you needed as many cores as possibly on the smallest budget. However, since the 8350's higher power draw necessitated a more expensive mboard to safely run it, much less overclock it, a Xeon E3 on a basic B75 or H77 mboard was usually not much more expensive.

Once again, this is why Zen is interesting. It's not barely keeping up due to a much higher clock rate. It is tying and beating Intel at the same clock rates. That's a huge improvement over the FX line.
 
I'm worried about Zen and whether or not AMD can come closer to the clock speed of the CPU's it wants to compete with while only consuming 95W.

One example is the 1266 MHz 14nm RX 480 vs the 1708 MHz 16nm 1060. There is a decent clock speed advantage for NVidia. Why wasn't AMD able to get higher clocks and will the sample we saw at 2.8GHz/3.2GHz be the best AMD can do on its 8 Core/16 Thread Zen that is suppose to compete against the 3.2GHz/4.0GHz i7-6900k?

Intel has a TDP advantage allowing for higher clocks. 140W vs AMD's Zen at 95W.

 


well it is too early to judge ...I personally wish that AMD wins this round ...intel is pricing the 10 cores i7 insanely high .. Only AMD can force them to lower that CPU to $999 ...
 

The clock speed advantage is of little importance: if you look at DX12/Vulkan benchmarks, the RX480 still beats the GTX1060.

Why does the RX480 not go as high in frequency? Because AMD's architecture favors being able to do more work per clock for the workloads it has been optimized for (DX12/Vulkan) instead of clock frequency, which means more stuff happening per clock and longer critical paths. Longer critical paths is what limits clock scaling. This is very similar to how Intel's architectures have favored deep ILP optimization over clock speed after Intel ran in a clock brick wall with Netburst.

AMD has been busy repeating Intel's Netburst mistake by making their pipeline longer in the hope of achieving higher clock speeds regardless of power over the last few years, which lead to the 45W 3.5GHz i3 beating the 200-250W FX8350 overclocked to 5GHz in many games.
 


Oops, you forgot the i5 600 series. ;D Has Intel split the core count like this with any of its later i5 lineups?

I have an i5 670, but I'd like to get hold of a 680 just to see if I can get it to 5GHz+ (I'm confident it could, as 4.7 with an i3 550 was easy).

 


I'm saying the architecture has always been better at CPU based 3D rendering in CAD programs like 3D Studio Max and Blender. In that benchmark you are comparing a 32nm chip to a 22nm chip. As for the Xeon E3 1230 v3. It would be difficult to tell if it would be a better choice than an FX-8350. They have not been benchmarked against each other in CPU-based rendering and the Xeon would be $100-$200 more depending on board selection. The main advantage the FX has that allows it to be above the Core i5's is its 8 cores. It loses to most Core i7 due to hyper threading. So the Xeon would be equivalent in that sense. However, when it comes to cache and memory bandwidth it has less. This may negatively affect the score. The E3 1230 typically is 8~10% slower than a 3770, but they were also built for 2 different workloads.
 
Yes, you keep saying that, but you're not explaining why you believe it when the benchmarks are quite contrary to it. If the FX line had the better architecture for these tasks, then they should be able to finish them faster than an Intel chip running at the same clock and with the same number of cores. Instead they consistently are slower in those tasks despite having a clock advantage.

Irrelevant. Smaller lithography simply allows for a smaller chip and generally lower power usage, but doesn't change the design of the microarchitecture, what instruction sets it allows, how efficient the pipeline is, etc. But if you still don't believe that, compare an 8150 @ 3.6 GHz against a 2600K @ 3.4 GHz. Both 32nm chips. In Blender, 3ds, and SolidWorks, the AMD chip still loses. The newer 8350 ties the 2600K, but it's also running 600 MHz faster.

An educated guess isn't difficult if you use a 4770K as a 1230v3 substitute. They share the same architecture, same cache, and HT. Except for ECC RAM support and a locked multiplier, the two are practically identical. A 1230v3 runs 5.5% slower than the 4770K ( 3.3 GHz vs 3.5 GHz ). It's reasonable to estimate the 1230 would score about 5.5% worse than the 4770K. The 8350 consistently comes in ~10% or more behind the 4770K, even with the 500 MHz advantage.

Well, if you want to go for a ridiculous mboard like an Asus Z97 ROG for the Xeon, sure. But why would you do that? There was typically a $60 difference between the two CPUs at the time. But you could drop the Xeon into a basic $50 H81 board with no problem. If you wanted extra features, nice H97 boards were commonly in the $70 - $80 range. The 8350, with its much higher power draw, typically needed a more expensive mboard to function properly, often $90 or more. And if you were going to OC it you needed to spend more on the cooler. As a locked chip, the Xeon's stock cooler was just fine and fairly quiet as well. So the price difference was usually closer to $40 - $50.

Well, I said the i5 was usually right behind the 8350, so I'm not sure what you're getting at.

You mentioned a lot of chips in the previous sentence, so I'm not sure what "it" is referring to. If the Xeon, no, it has the same cache and memory controller as the i7. If the 8350, then yes I agree. That's one of the problems with it I listed before.
 
(ex_bubblehead, only just noticed your comment after a post preview and typo check, hope my reply is ok!)


C.A.'s own summary for the launch of the 8350 summed it up well IMO:

"Would FX-8350 be my first choice in a new build, though? Probably not. Although I’m impressed by the work AMD’s architects have done in the last year, performance remains too workload-dependent. And, inexpensive energy aside, I’m going to go with the more efficient implementation when all else is close to equal."

That's what surprised me looking at the benchmark results at the time. One moment the 8350 matches a 2600K, another it loses to a non-HT 2500K because its IPC is behind, or its modular design means the shared resources are not well exploited in situations such as AE where one would normally expect decent performance from a multi-core CPU. Zen appears to move completely away from this, which is very good news for AMD. I certainly look forward to buying one, seeing how it fares, especially if it's priced even remotely competitively with whatever Intel has at the time. AMD has had certain mbd/chipset advantages in the past aswell (eg. many years ago I bought an AM2 board with a 6000+ instead of an E8400 config because it was the only way I could obtain proper PCIX); I really hope they can see the opportunities here, eg. providing a lot more PCIe lanes so users can exploit all sorts of storage options without sacrificing lanes for multiple GPUs (and I don't necessarily mean for CF/SLI, rather for GPU acceleration).

For content creation, time is money, and time to insight is important. Using an 8350 back then instead of a 3930K would be weird IMO (assuming initial affordibility). The 3930K did cost quite a lot more, but for this kind of task it was a sensible investment. For those who couldn't afford SB-E, the better efficiency, higher IPC (helps interactive performance, etc.) and lower power consumption made SB/IB a no-brainer. Next year though, Zen could easily look like the winner in these other aspects even if it's not actually faster overall; combined with any price advantage, it should do well.

One area where Intel goofed way back was not fitting boards with a decent number of native SATA3 ports; allowing X79 to fester for so long with just two was crazy. I talked to a guy who did actually buy an AMD setup purely because the mbd in question had lots of native SATA3 ports (obviously from the AMD chipset, at least eight I think it was), instead of via poor 3rd-party controllers as was so common on Intel boards (Marvell/ASMedia chips are terrible). Intel has finally fixed this by now of course, but perhaps AMD can do an equivalently sensible thing for AM4, like allow the inclusion of multiple M.2 ports, etc.





Indeed, the only tests in toms' 2012 review where the 8350 is competitive are Photoshop and 7zip, but elsewhere it's way behind, eg. WinRAR and AE (for the latter it's beaten by a 2500K). But then even where the 8350 isn't much slower, once one takes into account power consumption and efficiency, it's a general loss. I kinda read C. Angelini's review as being like what the 8150 should have offered a year earlier.

As a lesson in design decisions, BD and then PD are interesting. An ex-AMD employee said the company had reverted to using a lot of automated design tools, which made it difficult to do late stage optimisations and resulted in maybe a third more transistors being employed.

The information about Zen so far seems to suggest AMD has pulled completely away from this quagmire of poor IPC, low efficiency, high power consumption and a workload-dependent performance profile. I shall infer they're just designing better than they were before. Good news for CPU competition next year I hope. It will be interesting to see how Intel responds, either by big price cuts, or at last perhaps via a proper performance boost (all they really have to do if they wanted to is release an unlocked XEON, but I doubt that'll happen; more likely they'll try tweaking and meddling as they've been doing for the last 5 years, which if so would help AMD a lot), even if that's by using a higher-TDP socket to unleash something more interesting.





That's why I like the 2700K so much, 5GHz every time, so easy to oc, simple TRUE and one fan, it's almost embarassing. I can see why the value of P55 boards has gone up a lot in recent years, ie. those who do oc'ing purely for the technical challenge probably regard most recent Intel CPUs as just too easy and boring, but many of them probably skipped P55 at the time, so now they're going back to explore (just my speculation, but otherwise the bidding on eBay is inexplicable). None have been as easy as SB though, 5 mins with an M4E and I have a 2700K at 5GHz stable every time, built seven of them so far, giving the same performance as a stock 6700K for CB R15 (scores 880), with no complicated cooling issues.

It's ironic though that used SB-Es are even cheaper. Leaving aside mbd availability, we have 3930Ks going for as little as 72 UKP on eBay these days, hard to ignore if one's focus is rendering on a tight budget, especially when a 32GB 2400MHz RAM kit is so cheap now. Alas like I say, finding X79 mbds atm is much harder than it was a year ago.

I suppose this is what people mean by Intel competing with itself, and why they made IB such a heat spewer, to try and roll back the oc scene (delidding vids on YT show load temp drops of 30C+), but all that's done along with boring updates since then is disuade oodles of SB owners from upgrading, ditto those who bought Gulftown systems.

Ironically this very same issue will help AMD, ie. if Zen holds true to this early promise, not only will a heck of a lot of AMD owners upgrade, but I've no doubt plenty of Intel users will dip their toes in an AM4 build aswell. By being behind for so long in the midrange and top-end, AMD producing something good should mean strong demand. I just hope they manage the supply side sensibly to avoid price hiking, something which would spoil what should otherwise be a successful comeback.





That's the funny thing about XEONs, locked as they are, stock coolers result in scarily low noise much of the time. When I use my Dell T7500 (which has two X5570s), I often forget it's turned on because it's so quiet.





Though an easy label, it's not really accurate to say the 8350 has 8 cores. Rather, it has 4 modules, with shared fp resources, lower IPC and a design that means some workloads can't easily exploit the modular design (anyone know if that's a coding issue? Or purely hw?), which is why it so often fails to beat Intel's 4-core products of the day. The design made the 8350 quite strong for certain integer tasks though, eg. Handbrake. But this workload-dependent performance probably put a lot of people off.

Given it's attractive price at the time, I might have been tempted by an 8350, but I didn't like the power consumption, etc. Zen though is looking much better in all respects. Sure, Intel will have moved on by next year somewhat, but this is obviously early silicon so it's likely AMD can and will do more before launch.

For the first time in a long while, I'm hopeful AMD can get back in the game, which is the most +ve I've felt about the company in a long while. Some have pointed out that previous products were hyped and then proved disappointing, which is true, but I'd be astonished if AMD did that again, it'd be market suicide. This time things actually appear reasonably sunny on the horizon for them.

Ian.

 
I think a lot of people share that impression, that PD was what BD should have been. It definitely would have made 2011 - 2013 more interesting.

I think it was more software at the time. I remember Windows patches were released back then that were supposed to help with AMD core scheduling because a lot of resources weren't being utilized. If you look at the CPU benchmarks throughout the last few years, you'll notice the 8350 gaining a little ground as the test suites get updated.
 


Did AMD ever release a patch/update which specifically mentioned the 8350? Along similar lines, what does the DCO (Dual Core Optimizer) patch for Athlon64 X2 and other models actually do?

Ian.

 

AFAIK that would be Microsoft and the Linux community who would be responsible for that, not AMD (apart from AMD contributing to Microsoft and the Linux community to help them, of course.) AFAIK (again) the primary piece of software that would need to be patched would be the thread scheduler or the bit in the operating system that assigns threads to cores and tells them when to execute.

I guess another area of software that might have changed to help AMD (general software, not an operating system) is the C/C++ compiler, like Clang, GCC, Visual C++ and Intel's C++ compiler (yeah, right, thanks Intel! :pt1cable: :rofl: )

 
". If you look at the CPU benchmarks throughout the last few years, you'll notice the 8350 gaining a little ground as the test suites get updated. "

IIRC, that windows 7 scheduler patch did very, very little to help the Bulldozer/PD cpus. I think there was even some cases where it ended up being slightly slower than the unpatched kernel. I had an 8150/Sabertooth setup way back when it first came out and was really excited for that patch, and it ended up not doing much of anything at all.
 
You're right, the scheduling patch wasn't huge. I didn't mean to make it sound like it was. But as software has gotten updated over the years, and/or compilers as Andrew mentioned, all the small improvements have added up.
 
I'm exited that in 2017 the whole 4-core as the norm is going to go to the way side if AMD is successful. AMD has been pushing more cores for a while but they had weak cores so it never forced Intel's hand. So raise a glass that Zen is as good as AMD is hyping as we may finally move beyond 4-cores in the higher end mainstream i.e. non-enthusiast CPU's. This should be lovely for Vulkan / DX12 and maybe help lead the way for a few more truly multi-threaded applications(Yes I understand the complexities of multi-threading and some tasks do not work well on multiple threads).
 
Will AMD follow Intel's design and place the PCI-E controller on the die? If so, are they going to use the same marketing scheme that forces you to pay more for more lanes?

Sorry if this has been asked already.
 

Since AMD is converging APUs and desktop HEDT on a single socket and AMD's APUs have integrated PCIe with 20 lanes total, it would be logical to assume that the primary PCIe lane complement will come from the CPU. If AMD is still aiming for heterogeneous computing, the reduced latency between the CPU and GPU is mandatory.
 


I didn't realize they used that technology already. Thank you for clarifying that for me.

 
We're not really at the point yet where quad-cores are the norm ( four threads, yeah, but four physical cores? No ). I think it's premature to think Zen will make the quad-core the norm, much less beyond quad-core. I haven't seen anything yet on final Zen SKUs and whether a quad- or dual-core CPU will be their low-end. Remember that games need to sell a lot of copies. Making them only available for gamers with at least i5-level performance cuts out a lot of potential consumers. And most people, not just gamers, still can't afford ~$200 CPUs.

 

Quad core is already in most of AMD's APUs under $100. It isn't so far-fetched to think that Zen 4C8T might have a ticket price not much above that, especially if there is an FX model which physically lacks an IGP to inflate the die size. Without an IGP, such a Zen chip should land under 90sqmm, which should be much cheaper to produce than AMD's current 28nm chips.

AMD has a lot more to gain from taking a modest increase in gross margin to ensure that Zen gets out there in bulk than nuking most of their potential sales by being greedy and asking nearly full Intel price for roughly the same performance.
 


Just clarify, 2 core, hyper threaded CPU's are the norm right now?
 

He already clarified: "four threads, yeah, but four physical cores? No"

Though AMD has been doing quad cores even at the low end (though many disagree that having two integer cores with independent integer cores sharing an FP ALU qualifies as two cores) for quite a few years now in an attempt to make up for its poor IPC. Ideally, Zen 4C8T should pick up roughly where the FX8xxx left off.
 
Yes, my general feeling is that i3-level chips are about the mainstream in desktops right now. I could be dead wrong, but that's the vibe I get from everyone I talk to. Keep in mind, most people on TH's forum are above mainstream. Now, if AMD can push a good quality 4C/8T chip out the door that can rival the i5 in price, they very well could accelerate market adoption of it, and that could do fabulous things for software development.
 
Someone will probably explain a reason why it is not feasible but something that would be really sick is not only the chips performing close to the Intel ones but actually being capable of running 2 on the same board. Intel's i7s are not designed to run in dual CPU configurations but AMD allowing their highest end consumer CPUs to run dual configurations would be absolutely stellar... but I just cannot see it happening for whatever reason. Let them Opterons keep their quad CPU configurations at premium prices.
 
Status
Not open for further replies.