[Techpowerup] AMD "Zen" CPU Prototypes Tested, "Meet all Expectations"

[Glo.]

I think first gen Zen CPUs will bring Haswell-EP-ish performance levels.

 

[myocardia]

i7 5960X runs @3ghz and comes with tdp of 140W.

The 5960X is a 22nm CPU, though.

Why are people expecting Zen's 8 core to be better than this when it is said to be 95W?

I have no idea. The only person I've seen in this thread mention 4 Ghz said the quad core/8 threaded model. I personally will be buying one of the 16 thread Zens, as long as it has the rough IPC of Intel's Sandy Bridge CPUs, even if it has a speed of less than 3 Ghz, as long as it isn't WAY below 3 Ghz. That is about the speed and IPC I'm expecting the 8 core/16 thread Zen to have, btw.

I hope Zen, Artic Islands gpus and k12 are good enough to help AMD survive and keep competition in X86 and gpu market.

I thought that K12 had been officially canceled? Or has it just been unofficially canceled, since I haven't heard AMD mention it in years? BTW, welcome to Anandtech.

 

[wiseguyin]

Ok -- i fell for the headline too

 

[looncraz]

Anyway, everyone just always saying "search" or "it's huge" instead of giving me something to actually look at.

Compared to an FX83XX, Zen should have like what...60% to 65% more IPC?

Zen should be about 64.09% faster than Piledriver, according to my in-depth benchmarking then assuming a 40% straight improvement for Zen.

Of course, in the real world, it REALLY matter where the improvements are seen. An 80% improvement in one area and a 0% improvement in another adds to 40% :thumbsup:

Ze's design suggests we should see a performance profile closer to that of the Phenom IIs than of Bulldozer, so we should see > 64% improvement for Cinebench, and < 64% improvement for compression and a few areas where Zen can't really add to IPC (where Bulldozer had 1 cycle latency, we should even expect a possible reduction in IPC, though net throughput could still be improved... CPUs are complicated little beasts :wub.

I've been trying to get around to doing a full array of benchmarks on my wife's Phenom II x4 955 @ 3.5 to compare with my i5 2500k @ 3.5 and i5 4690k @ 3.5 results. Will do so soon-ish. I hope ;-)

In any event, if you take this comparison:

http://www.anandtech.com/bench/product/700?vs=1261

You can ignore the clock-speed differences. I've verified that, in real testing, the 4690k runs at 3.7GHz in full load, 3.9 when one OR two cores are loaded, and the FX-4300 runs at 3.8 in full load, and bobs between 3.9 and 4.0 when one core is in use. So, basically, for our uses, we can loosely consider these a same-ish-clock comparison :whiste:

If evenly apply the expect 64% improvement, we see that Zen just blasts past Haswell in many places, but stays well behind in other places. If you average it all, they come out neck and neck. Which should be expected from other methods.

Method #2:

Comparing Excavator to Penryn (Core 2), IPC is nearly identical. Haswell's IPC, according to Intel and a multi-bench mean I calculated, is 39% faster than Penryn. With Excavator and Penryn being nearly identical, and both having a 40% improvement in IPC, performance should be expected to be similar.

Method #3:

Comparing Excavator to Deneb (Phenom II), we see that Excavator has about an 8% higher IPC. Penryn has about an 8% higher IPC than Deneb. From there, you can add 40% to both to reach Haswell/Zen.

And there are other ways, of course

Everyone likes to look at the areas where Intel has improved most and compare it to where AMD improved none (except for Excavator): Cinebench.

Take away Cinebench results and the story changes quite a bit. And, of course, applications that actually take advantage of AVX.

 

[Dresdenboy]

This is the Chinese Whispers effect.

Same with Zen pipelines (where did those 256b FMACs reappear?) and other tidbits.

Been there, done that.

 

[LTC8K6]

In any event, if you take this comparison:

http://www.anandtech.com/bench/product/700?vs=1261

You can ignore the clock-speed differences. I've verified that, in real testing, the 4690k runs at 3.7GHz in full load, 3.9 when one OR two cores are loaded, and the FX-4300 runs at 3.8 in full load, and bobs between 3.9 and 4.0 when one core is in use. So, basically, for our uses, we can loosely consider these a same-ish-clock comparison :whiste:

If evenly apply the expect 64% improvement, we see that Zen just blasts past Haswell in many places, but stays well behind in other places. If you average it all, they come out neck and neck. Which should be expected from other methods.

Let's take this one instead, where we know the i3 does not have turbo, and is running it's 2 cores at 3.7Ghz under load, so we know it is at a clock speed disadvantage to the FX-4300 all the time:

http://www.anandtech.com/bench/product/700?vs=1197

Heck, let's look at this one too, where the i3 has a large clock speed handicap:

http://www.anandtech.com/bench/product/700?vs=1267

 

[myocardia]

Zen should be about 64.09% faster than Piledriver, according to my in-depth benchmarking then assuming a 40% straight improvement for Zen.

The problem with that line of thought is that AMD is a business. Like all other businesses whose success or failure is predicated on their products' performance, "40% higher" can only mean one thing: up to 40% higher, in some instances/scenarios. If Zen has 40% higher performance with one software, then AMD wasn't lying about it's performance.

If Zen actually averages anywhere near 25% higher IPC than Excavator, with anywhere near the efficiency that it should have, it will be a grand slam, runaway hit, at least in the consumer market.

 

[LTC8K6]

If Zen actually averages anywhere near 25% higher IPC than Excavator, with anywhere near the efficiency that it should have, it will be a grand slam, runaway hit, at least in the consumer market.

Why would it? Is 25% better than excavator all that impressive?

Heck, Zen also might use too much power, or it might cost too much.

It's 25% better IPC numbers may get overshadowed.

There's no way to tell at this point.

Have we even had a whiff of a desktop excavator chip?

 

[LTC8K6]

In September 2015, they could only "see the finish line" for Zen:

Suzanne Plummer, the veteran Austin chip engineer who heads the Zen team, exudes confidence about the project.

&#8220;It is the first time in a very long time that we engineers have been given the total freedom to build a processor from scratch and do the best we can do,&#8221; Plummer said. &#8220;It is a multi-year project with a really large team. It&#8217;s like a marathon effort with some sprints in the middle. The team is working very hard, but they can see the finish line. I guarantee that it will deliver a huge improvement in performance and (low) power consumption over the previous generation.&#8221;

Plummer has worked at AMD since 2002, when it acquired the startup she was working for, Alchemy Semiconductor.

 

[myocardia]

Heck, Zen also might use too much power

Read what you quoted.

 

[myocardia]

Why would it? Is 25% better than excavator all that impressive?

Is 500% better IPC gains than Intel has been averaging for years now impressive? I would say yes.

 

[frozentundra123456]

It is going to be very limited in the consumer market until they get an igp integrated. 2017 at least, maybe later? Even *if*, and I cant stress how big that if is, it comes even close to the constantly inflating expectations in these forums, the enthusiast market is what, maybe 5% or less of the total market? OEMs are certainly not going to want to go back to having to use motherboard graphics or add a discrete card for the consumer/enterprise/education user.

If they can get decent cpu performance and integrate HBM apus, at a reasonable cost, that could be an attractive package though.

 

[Fjodor2001]

Server and HEDT CPUs in 2016/2017, desktop APUs in 2017+. Should be lots of money to grab for AMD if Zen is successful.

Remember that the Server and HEDT markets have high margins too. And that AMD is a relatively small company, so even grabbing a decent share of those markets should be huge for them. Huge for the consumers too, to finally get some significant competition going again.

 

[Arachnotronic]

Server and HEDT CPUs in 2016/2017, desktop APUs in 2017+. Should be lots of money to grab for AMD if Zen is successful.

Remember that the Server and HEDT markets have high margins too. And that AMD is a relatively small company, so even grabbing a decent share of those markets should be huge for them. Huge for the consumers too, to finally get some significant competition going again.

Are you trying to promote AMD stock?

 

[Fjodor2001]

Are you trying to promote AMD stock?

Nah, just saying that there is quite a big potential for AMD if Zen is successful. Especially since they are (re)starting from such a low level.

 

[MiddleOfTheRoad]

I am pretty sure that AMDs new architecture will meet the expectation that any knowledgeable, unbiased, rational, reasonable person would have.

Exactly.

Of course -- Zen could retail for $75, offer quadruple the performance of Haswell, benchmark Cinebench with 1 billion points and operate with a TDP of 5 watts and Shintai will still call it a flop.

Gotta love it.

 

[mysticjbyrd]

It's on wikipedia, therfor it must be true.

https://en.wikipedia.org/wiki/Zen_(microarchitecture)

It's going to suck megawatts and kick all out ass, intel dummies will scramble to regain som level of formation in dispair and horror - an utterly hopeless maneuveur before getting squashed under the mighty force that is the Zen.

.
.
.
.
.
.
.
.
.
.
.
[giving you a chance to type alot of stuff, then erase it. your welcome.]

On another note, suppose everything is fine and good and even glofo get its act together, I am reading target TDP's for zen as low as 5 watts for dual cores and upper 95 for performance.. thats like an exact copy of what Intel is doing, stretching one uarch over the entire field. That cant be a good sign in cotext of "competing with Intel", I dont see the angle? How can they hope to copy, step by step, and still end up better?

Being different screws them from a software perspective. That's a lot of the problem now.

 

[looncraz]

Would that be the same AMD that sold the first ever $1,000+ personal-usage CPUs, and had median retail price of their CPUs that was well over $300, until Intel Conroed them, and forced them to lower their prices?

That was a supply/demand issue. It was what Intel tried to use as their preliminary defense when AMD sued them for antitrust violations.

If you can't make enough, but people are willing to pay more for the ones you have, you charge more so that you can make the most money. Simple.

If AMD could have kept up with demand, they would have STILL not overtaken Intel, because Dell, Gateway, HP, and IBM were all contracted to Intel and AMD could literally give away their CPUs and it would have cost those OEMs more (I worked for one of those companies, and we discussed the problem internally).

Intel used the same tactic as Microsoft. You couldn't get other operating system software on desktop computers because Microsoft would yank your volume discount, which was worth $50+ a computer.

 

[Magic Carpet]

Nah, just saying that there is quite a big potential for AMD if Zen is successful. Especially since they are (re)starting from such a low level.

Yeah, could be a bomb. Just hope my Haswell rig won't get de-valued overnight when Zen comes out

Voted for "no idea".

 

[LTC8K6]

Read what you quoted.

Trust the saleswoman?

Really?

 

[LTC8K6]

Is 500% better IPC gains than Intel has been averaging for years now impressive? I would say yes.

Sure it's impressive, but when you are starting from way back, getting 3/4 of the way to the front looks impressive, but you are still way behind.

We simply have no idea about Zen at this point.

 

[looncraz]

Let's take this one instead, where we know the i3 does not have turbo, and is running it's 2 cores at 3.7Ghz under load, so we know it is at a clock speed disadvantage to the FX-4300 all the time:

http://www.anandtech.com/bench/product/700?vs=1197

Heck, let's look at this one too, where the i3 has a large clock speed handicap:

http://www.anandtech.com/bench/product/700?vs=1267

I was trying to compare quad core CPUs

But the picture doesn't really change, the single-threaded benchmarks here show that adding 64% will lead to a pretty good matchup:

BENCH - FX-4300 - i3 4360

CB R15 - 92    - 149
CB R10 - 4114  - 7369
CB R11 - 1.02  - 1.7
3dPM   - 69.63 - 122.74

BENCH - Zen-4400 - i3 4360

CB R15 - 151    - 149
CB R10 - 6751   - 7369
CB R11 - 1.67   - 1.7
3dPM   - 114.26 - 122.74

And this doesn't take into account many factors that Zen will have going for it that may not be included in that 40% figure (though they may be, so we can't play with numbers more than this and have them mean anything). However, it doesn't take into account the possibility that Zen may not be able to reach these speeds in stock SKUs (more likely with a quad core, though).

Of interest, though, is that these particular programs will likely see the largest boosts. Bulldozer was weakest in these, but Phenom II was quite good at them in comparison to the rest of its performance.

http://www.anandtech.com/bench/product/362?vs=1197
There aren't much on Anandtech that give us a direct comparison, (and it's important to remember that this is a Quad core vs a dual core for multithreaded benchmarks... and they pretty much tie).

Yes, the Phenom II had higher IPC than Piledriver. That went away with Steamroller... except for Cinebench scores, which pretty much didn't change (or even got worse).

http://www.anandtech.com/bench/product/1200?vs=1262

We see a slower clocked 7850k beating a 6800k, while pulling less juice. Except in Cinebench, where pretty much nothing changed until Excavator, where Cinebench scores jumped about 10% at the same clocks.

Intel has focused on the very areas AMD has weaknesses, making the situation look worse than it is when comparing those areas (and it just happens that the favorite benchmarks used by reviewers are exactly the areas where Intel focused improving performance...it's not cheating, of course, since the performance is there, but it's giving a skewed comparison).

My own fixed-clock (3.5GHz) Sandy Bridge vs Haswell benchmarks:

All at 3.5GHz
BENCH - 4690k > 2500k %

CB R10 ST - 16.65%
CB R10 MT - 17.51%
CB R11 ST - 4.11%
CB R11 MT - 6.14%
CB R15 ST - 12.2%
CB R15 MT - 16.3%

Octane ST - 5.23%
Octane MT - 4.87%
WebXPRT ST - -5.28%
WebXPRT MT - -2.6%
Silverbench ST - 1.16%
Silverbench MT - -2.7%

I had to repeat these results recently to verify that performance really did decline (Firefox 42.0, nearly identical systems, even motherboard brand (AsRock) and installed software). My previous results, based on web findings, and a previous Haswell system I built, said the same (to within about 0.5%).

Non-ST benches were run on a single core by using CPU affinity in Windows 7 x64. This results in close to the purest result you can get as other tasks have other core to use and the core (core 3) can perform only one task.

 

[looncraz]

The problem with that line of thought is that AMD is a business. Like all other businesses whose success or failure is predicated on their products' performance, "40% higher" can only mean one thing: up to 40% higher, in some instances/scenarios. If Zen has 40% higher performance with one software, then AMD wasn't lying about it's performance.

If Zen actually averages anywhere near 25% higher IPC than Excavator, with anywhere near the efficiency that it should have, it will be a grand slam, runaway hit, at least in the consumer market.

This is a VERY real concern, no doubt. We really don't know what they mean by 40%, beyond the claim of IPC. With Lisa Su in charge, though, I'd suspect she truly means it as an instruction retirement rate improvement. With double the ALUs, faster caches, no module overhead, four dedicated FPU pipelines, I'd expect some pretty hefty improvements.

 

[Shamrock]

If its faster than my i7 4790K (in games), I'll buy it.

 

[Boze]

i7 5960X runs @3ghz and comes with tdp of 140W. Why are people expecting Zen's 8 core to be better than this when it is said to be 95W? Also Intel's foundry is definitely more experienced than that at GF. So no help to AMD there either.
I hope Zen, Artic Islands gpus and k12 are good enough to help AMD survive and keep competition in X86 and gpu market.

Nice to see at least some realism in this thread.