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

[tential]

With AMD its better to have low expectations. I expect they will disappoint with performance. AMD has become all talk and not much about execution. Nvidia and Intel are hammering AMD to its imminent death. As an underdog supporter I still hope AMD can bring out competitive products in 2016/2017 and recover towards long term sustainable profitability.

You got me all hyped up on Fury X/Fiji too....
Then Fury X launched.... I wanted to cry....

 

[Dresdenboy]

You got me all hyped up on Fury X/Fiji too....
Then Fury X launched.... I wanted to cry....

Do you consider DX11 or DX12 performance, and at low or high resolution, and with slow or fast CPU?

 

[Jacky60]

Its legitimately disappointing to see how many people think its gonna be a bulldozer attempt. I expected more objectivity from the members of this forum.

What does this even mean. AMD have failed to compete and under-delivered in the CPU space for some time now especially when it comes to IPC. Yes they are good value but in absolute performance terms they're way behind. Bulldozer was expected to change that but under performed expectations by a big margin. They've recently copied this business model into the GPU space with Fury-x (not a bad product but certainly not as good as hoped for) being over priced considering it's limited RAM and almost complete lack of overclockability.
Being objective means being unbiased and if a manufacturer consistently fails to deliver on expectations then being unbiased means expecting more of the same i.e. under-delivering on expectations again. I really want AMD to do better and compete but lack of R&D budget and various other financial restrictions in addition to Intel and Nvidias dirty tricks mean I don't expect a return to the glory days any time soon.

 

[Dresdenboy]

What does this even mean. AMD have failed to compete and under-delivered in the CPU space for some time now especially when it comes to IPC. Yes they are good value but in absolute performance terms they're way behind. Bulldozer was expected to change that but under performed expectations by a big margin. They've recently copied this business model into the GPU space with Fury-x (not a bad product but certainly not as good as hoped for) being over priced considering it's limited RAM and almost complete lack of overclockability.
Being objective means being unbiased and if a manufacturer consistently fails to deliver on expectations then being unbiased means expecting more of the same i.e. under-delivering on expectations again. I really want AMD to do better and compete but lack of R&D budget and various other financial restrictions in addition to Intel and Nvidias dirty tricks mean I don't expect a return to the glory days any time soon.

Ok, and your view is not biased by the fact, that an uarch family spans many years, thus creating the impression you got over multiple years in a row? Same for VLIW4 and GCN.
How about synthesized Bobcat vs. the Atoms back then?

 

[Jacky60]

Ok, and your view is not biased by the fact, that an uarch family spans many years, thus creating the impression you got over multiple years in a row? Same for VLIW4 and GCN.
How about synthesized Bobcat vs. the Atoms back then?

I am biased, I'm biased in favour of AMD and make no secret of it. I'm saying that if a manufacturer fails to deliver on expectations or rumoured (perhaps hyped) performance consistently then to expect more of the same is not bias it's rational. If I tell you three lies in succession do you think you would be biased anticipating I might lie a fourth time? It's a simple concept, basing expectations upon past experience rather than treating every possible new development of a product line as a unique event with no connection to the past.
I would expect Nvidia to try and cheat benchmarks, use GW etc to make AMD performance look poorer for games, I would expect Intel to bribe/ coerce OEMs not to use AMD CPUs why? Because both companies have a record of doing such things. I expect AMD's next iteration of it's architecture in the CPU space to be underwhelming, why? Because it's been like that for about a decade.

 

[looncraz]

I am biased, I'm biased in favour of AMD and make no secret of it. I'm saying that if a manufacturer fails to deliver on expectations or rumoured (perhaps hyped) performance consistently then to expect more of the same is not bias it's rational. If I tell you three lies in succession do you think you would be biased anticipating I might lie a fourth time? It's a simple concept, basing expectations upon past experience rather than treating every possible new development of a product line as a unique event with no connection to the past.
I would expect Nvidia to try and cheat benchmarks, use GW etc to make AMD performance look poorer for games, I would expect Intel to bribe/ coerce OEMs not to use AMD CPUs why? Because both companies have a record of doing such things. I expect AMD's next iteration of it's architecture in the CPU space to be underwhelming, why? Because it's been like that for about a decade.

With the exception of Fury X (where they used benchmark settings that favored Fury X over 980Ti and one AMD rep said it would be an overclocker's dream), AMD's claims have been pretty much spot on. JF-AMD wasn't representing AMD's public claims, he was making claims of his own, so he doesn't count.

The problem is that people take the slightest amount of information and take it too far. And, they're always rooting for the underdog to slay Goliath, so they keep trying to predict it ill happen. Zen will max out around Haswell in IPC, period. We don't know how it clocks, we don't know how well its caches work.

If Zen has a slower cache system than Haswell, it will be slower than Haswell. Pretty simple. I'm more interested in the efficiency of its SMT implementation, as that is a good way for AMD to add value in lower SKUs without adding die area.

I will not be surprised when a Zen 4c/8t CPU is competing against the Skylake i5s, losing in single thread, winning in multi-thread. I will not be surprised when they offer every core count - with SMT available at an extra cost. AMD would be incredibly stupid to shakeup CPU pricing, they can't afford to play on price. They CAN afford to play Intel's game exactly as Intel is doing.

 

[Dresdenboy]

I am biased, I'm biased in favour of AMD and make no secret of it. I'm saying that if a manufacturer fails to deliver on expectations or rumoured (perhaps hyped) performance consistently then to expect more of the same is not bias it's rational. If I tell you three lies in succession do you think you would be biased anticipating I might lie a fourth time? It's a simple concept, basing expectations upon past experience rather than treating every possible new development of a product line as a unique event with no connection to the past.

I understand your view. I just wanted to bring in any typical long term effects based on decisions made a long time ago by people, who are no longer in the company. Same goes for statements.

I would expect Nvidia to try and cheat benchmarks, use GW etc to make AMD performance look poorer for games, I would expect Intel to bribe/ coerce OEMs not to use AMD CPUs why? Because both companies have a record of doing such things. I expect AMD's next iteration of it's architecture in the CPU space to be underwhelming, why? Because it's been like that for about a decade.

AMD could have an edge in console ports.

Of course it would be very bad, if the mostly CPU veterans who worked on Zen didn't find a way to create something cool after beeing given full freedom in designing it and after years of evolution of their internal tools and work processes and grown experience.

 

[Jacky60]

My estimate is 20% less than I'd expect or like, multi and single threaded. At best, Haswell E performance in 12-18 months time.
I hope and want to be pleasantly surprised but I don't expect to be.

 

[jpiniero]

Seeing Intel's investor meeting... It pretty much confirms how badly Zen would need to be competitive in Data Center (read: perf/watt) for AMD to survive. No other metric matters. They simply won't make enough with how declining the consumer pc market is even if they gained share.

 

[looncraz]

My estimate is 20% less than I'd expect or like, multi and single threaded. At best, Haswell E performance in 12-18 months time.
I hope and want to be pleasantly surprised but I don't expect to be.

Honestly, even with what little we know, it'd be pretty difficult to make that design only 20% faster than the disaster that is the Bulldozer family. Sure, if they were starting from a truly blank slate and had no existing IP from which to propel themselves I'd never expect Haswell-like IPC.

However, AMD could have, literally, taken the ALUs from Bulldozer, slapped on two ALUs, and split up the specialization hardware between the four total ALUs... and we'd see a 30%+ increase in integer IPC provided the front-end can keep them fed (which shouldn't be that difficult considering each ALU can do LEA, mov[x], cmp, rotates, shifts, etc.).

Zen should be able to schedule 8 instructions a cycle and execute 10 (4x ALU, 2x AGU, 4x FPU). Considering how much difficulty Bulldozer* had even managing two ops per cycle, this is an enormous shift. However, naturally, there are consequences in adding more execution units, and scaling is never what you want it to be.

The reality is that we will see the core bursting through instructions and the cache system, no matter how good, holding up the show. This is great for power efficiency, but means that the design's performance pretty much hinges on the one area AMD has never really excelled at: cache performance.

They built this in only three years, which means they used as much existing IP as they could (and this fact was repeated many times). They didn't build new multiplication or division units, new decoders, new memory controllers, new predictors, etc. They just bolted together the best of what they had with some binding logic.

What's really interesting is the division of labor in the Zen core. In some ways, Zen will actually be barely any faster than Excavator as a result of anything we know about it (still can only do one multiplication and one division at a time, for example).

http://looncraz.net/ZenAssignments.html

 

[looncraz]

Seeing Intel's investor meeting... It pretty much confirms how badly Zen would need to be competitive in Data Center (read: perf/watt) for AMD to survive. No other metric matters. They simply won't make enough with how declining the consumer pc market is even if they gained share.

When you have such a small pie of a market it doesn't take much growth to help your bottom line. What AMD really needs is to increase their margin, and Zen should be able to do that healthily.

 

[jpiniero]

When you have such a small pie of a market it doesn't take much growth to help your bottom line. What AMD really needs is to increase their margin, and Zen should be able to do that healthily.

They need both sales and margin really to cover the high costs. You can see it with the GPU side... the 390/Fury raising prices scheme has been a complete failure. They've gotten sales when they offered tremendous value (like the closeout pricing on the 290/X recently) but they probably didn't make very much on each sale.

 

[Arachnotronic]

When you have such a small pie of a market it doesn't take much growth to help your bottom line. What AMD really needs is to increase their margin, and Zen should be able to do that healthily.

When you have such a small pie of the market and you are nearly as old as your largest competitor which has essentially all of the pie, there's usually a reason for that.

 

[NostaSeronx]

AMD "Zen" CPU Prototypes Tested, "Meets all expectations from our imaginary customer base"

2017 Reviews;
Too late, too slow.[Cannonlake monolith looms over Zen]
Intel is better.[Has AVX512 cheeseburger, better CB16, x264, 7zip, etc scores.]
Expensive, it is actually cheaper to get Intel this time.[Mobile<->HPC]
Unexplained performance deterioration when SMT is in use.[I wonder why]
Higher power consumption than Intel.[Again]

 

[looncraz]

AMD "Zen" CPU Prototypes Tested, "Meets all expectations from our imaginary customer base"

2017 Reviews;
Too late, too slow.[Cannonlake monolith looms over Zen]
Intel is better.[Has AVX512 cheeseburger, better CB16, x264, 7zip, etc scores.]
Expensive, it is actually cheaper to get Intel this time.[Mobile<->HPC]
Unexplained performance deterioration when SMT is in use.[I wonder why]
Higher power consumption than Intel.[Again]

Those would be the headlines if even the following were true:

Releases in January 2017, 1% slower than Haswell.
Scores are 1% lower in CB16, x264, 7zip, etc, but 25% lower in AVX512 which no one uses.
Costs 0.025% more than the "equivalent" Intel CPU, but offers more features for the same price.
SMT deterioration is < 5%, scales 15% - still better than Intel's original.
Uses 5W more at a given performance tier.

No matter what, many reviewers will paint the most negative picture they can. See it all the time.

 

[NostaSeronx]

Releases in January 2017, 1% slower than Haswell.

My simulation puts it at less than Ivy Bridge.

Scores are 1% lower in CB16, x264, 7zip, etc, but 25% lower in AVX512 which no one uses.

Scores won't be 1% but more around 30% lower in AVX2, and 2x lower in AVX512. You seem to have missed the "Native" movement. AVX512 will be used a lot more than you think.

Costs 0.025% more than the "equivalent" Intel CPU, but offers more features for the same price.

10nm FinFETs => more chips => lower cost. The equivalent Intel product will be two full shrinks ahead. Cannonlake 8-core SoC = Zen 8-core CPU, Lisa Su: "AMD is not a low price company anymore."

SMT deterioration is < 5%, scales 15% - still better than Intel's original.

The SMT performance hit will be around 30+% to a maximum 50%.

Uses 5W more at a given performance tier.

We are looking at a 4x pJ differential between 20nm FinFETs @ SamFoundries and 10nm FinFETs @ IntelFoundries. A 25w Cannonlake 8-core SoC can be compared to a 95w Zen 8-core CPU and beating it.

 

[dark zero]

I tought that my gloom was the worst.... I was totally wrong.

Let me be clear. If AMD dies, VIA quits (or even Intel stops VIA due the Chinese influence) and Rockchip and Spreadtrum will abort that Frankenstein permanently.

So x86 will be Intel exclusive and even Mobo manufacturers will see that they will fear that Intel returns to the Mobo market too. Some techs like nVILINK won't appear on x86 and even PCI Express will be changed to the point on only favoring Intel.

In few words, x86 will end like Power PC. The best, but elitist and expensive. Maybe the time for another tech to become mainstream like ARM finally comes.

Maybe that is good since it will deliver to another gaming crash and only the best will survive. That won't sound bad after all.

 

[AtenRa]

10nm FinFETs => more chips => lower cost.

As with 22nm and especially 14nm, 10nm will be more expensive when it will arrive than its predecessors.

The equivalent Intel product will be two full shrinks ahead. Cannonlake 8-core SoC = Zen 8-core CPU, Lisa Su: "AMD is not a low price company anymore.

There is no DESKTOP Cannonlake in 2017.

 

[NostaSeronx]

As with 22nm and especially 14nm, 10nm will be more expensive when it will arrive than its predecessors.

Except, it isn't. The price is negated by the shrink. 193nm-TP is cheaper than EUV at Intel's 10nm node.

There is no DESKTOP Cannonlake in 2017.

Cannonlake's smallest SKU launches Q4 2016, mobile is Q1-Q2 2017, desktop is late Q2 2017, with actual launch mid-Q3 2017.

You don't go against the x86 king of SMT and FinFETs with SMT and FinFETs.

 

[AtenRa]

Cannonlake's smallest SKU launches Q4 2016, mobile is Q1-Q2 2017, desktop is late Q2 2017, with actual launch mid-Q3 2017.

Desktop KabyLake release is Q4 2016, i wouldnt count for Desktop CannonLake release earlier than Q1 2018.

 

[Arachnotronic]

Except, it isn't. The price is negated by the shrink. 193nm-TP is cheaper than EUV at Intel's 10nm node.Cannonlake's smallest SKU launches Q4 2016, mobile is Q1-Q2 2017, desktop is late Q2 2017, with actual launch mid-Q3 2017.

There won't be any Cannonlake in Q4'16 of any kind...

I think we'll see desktop Cannonlake in Q4'17-Q1'18.

 

[DrMrLordX]

AMD "Zen" CPU Prototypes Tested, "Meets all expectations from our imaginary customer base"

2017 Reviews;
Too late, too slow.[Cannonlake monolith looms over Zen]
Intel is better.[Has AVX512 cheeseburger, better CB16, x264, 7zip, etc scores.]
Expensive, it is actually cheaper to get Intel this time.[Mobile<->HPC]
Unexplained performance deterioration when SMT is in use.[I wonder why]
Higher power consumption than Intel.[Again]

. . . what?

My simulation puts it at less than Ivy Bridge.

YOUR simulation? Who gives a fig about YOUR simulation?

A 25w Cannonlake 8-core SoC can be compared to a 95w Zen 8-core CPU and beating it.

You don't know that. You don't even know what Cannonlake will be like, nor do you seem to know when Intel will actually bring it to market . . .

Cannonlake's smallest SKU launches Q4 2016, mobile is Q1-Q2 2017, desktop is late Q2 2017, with actual launch mid-Q3 2017.

Since when? Intel has confirmed that their 10nm process won't go onto the market until 2017 at the earliest in ANY form factor. Including mobile.

You don't go against the x86 king of SMT and FinFETs with SMT and FinFETs.

If that's the case, then you don't go against them with anything, at all, period. CMT didn't work.

 

[TechGod123]

AMD "Zen" CPU Prototypes Tested, "Meets all expectations from our imaginary customer base"

2017 Reviews;
Too late, too slow.[Cannonlake monolith looms over Zen]
Intel is better.[Has AVX512 cheeseburger, better CB16, x264, 7zip, etc scores.]
Expensive, it is actually cheaper to get Intel this time.[Mobile<->HPC]
Unexplained performance deterioration when SMT is in use.[I wonder why]
Higher power consumption than Intel.[Again]

I really wish that fanboy rule could be broken. This is bloody ridiculous.

 

[myocardia]

YOUR simulation? Who gives a fig about YOUR simulation?

I do. For every moronic loon that comes in here, making retarded predictions of Zen having Skylake or higher IPC, there should be someone else laughing at and making fun of them, as Nosta was doing here.

 

[frozentundra123456]

Yea, kind of ironic that those making the biggest criticisms of other's "simulations", satirical or not, are the first to promote/accept the AMD simulation that showed 40% ipc gain.