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

[cbn]

AMD always talks of averages, they said 4% for EXV, was it up to 4%..?

It was 9-13%, so even people who talk of track record are blatlantly ignoring this record and are instead posting completely made up assumptions that are contradicted by the facts.

I think Zen will be a good CPU for high core count Servers and mobile (and to a lesser extent desktop as well).

My biggest concern actually is how they configure the chips.

For example, based on what we are seeing happen with Carrizo (eg, 15W Carrizo + 15W dGPU favored over 35W Carrizo) I hope this time they favor the CPU (for the APU configurations). And go conservative on the iGPU for at least of one the core configurations.

For desktop, I do hope they get the L3 cache right this time. That was ballooning the die quite a bit in Vishera (but didn't seem to add much in the way of performance).

 

[Vesku]

Intel seems to get some info about AMD stuff from various sources after a tapeout, at the very least OEMs probably talk to Intel about how they can counter offer on AMD proposals. I don't think it's coincidence that we are hearing Cannonlake will up the core count per SKU and that it should be available around the same time as Zen.

 

[svenge]

Lets keep it simple, lets keep ti technical,

Does the Intel Sales Team (IST) [blah blah rant rant]

Your "Intel Sales Team" statment is quite ironic, especially considering that you're literally the only one here who actually makes his living selling second-rate processors in a third-world economy. I think your declining fortunes have finally started deteriorating your thought processes...

 

[cbn]

Intel seems to get some info about AMD stuff from various sources after a tapeout, at the very least OEMs probably talk to Intel about how they can counter offer on AMD proposals. I don't think it's coincidence that we are hearing Cannonlake will up the core count per SKU and that it should be available around the same time as Zen.

So Intel has 8C/16T + iGPU, 4C/8T + iGPU, 2C/4T + iGPU (?) for 10nm.

What does AMD launch for Zen APU?

4C/8T + "big" iGPU?

2C/4T + "medium" iGPU?

(Or maybe instead of that 2C/4T + medium iGPU they instead launch a 4C/8T with "small iGPU" and bin as 3C/6T for 5W? Or maybe worst case scenario take the 2 best of 4 cores for 5W?)

 

[mrmt]

This is what happens again and again in these threads: we pretend to talk specs while we call each other names. Both sides. It is incredibly tiresome for people wanting to learn more about any upcoming product to filter this noise constantly.

I would be happy to talk about it if we had anything beyond AMD kool-aid about this upcoming product.

 

[itsmydamnation]

I would be happy to talk about it if we had anything beyond AMD kool-aid about this upcoming product.

Then why post, yeah maybe you should see yourself out . We know bits and pieces but the biggest problem with the nay sayers is they cant tell you whats good or whats bad about bulldozer, just that bulldozer suxors. Go read what agner has to say about things like Bulldozers predictors, prefetches and decode or the tests there have been on the load store system.

Then look at the things they haven't bothered/been able to fix, execution width, mispredict latency, L1i associativity, L2 latencies*. Now what matters for "IPC", recovering from stalls and racing to stalls and thats whats poor in CON cores and they never really addressed them (likely fundamental issues).

Now look at what they actually improved from bulldozer to ex,
Load/Store system update in every revision,
uop cache,
improved branch prediction
increases in TLB's
implemented DVFS

All these things can and likely will be used as a base for Zen along with Decode, PRF and prefetches (which along with predictors are still shared in Excavator). Scheduler direction will be interesting as jaguar and con cores are different there.

Now look at what has been said about zen, we have more execution ports ( good for loops, but also for things like branches which used the same port as mul on con cores), some instructions have moved form AGU to ALU (freeing up AGU time) . We know there is a complete new cache system. Thats two of con cores big problems right there. The FP units latency has dropped as FMA looks to be some kind of bridge(another sore point vs intel as they still have a non FMA FP adder).

On the guessing side we expect a shorter pipeline ( no more mid 4ghz targets) will help mispredict latency, larger PRF ( cuz SMT) which means the core can be more aggressive in predictive Loads when running 1 thread.

So you take the good work that has been done over 4 years and integrate that into a core that addresses the CON cores weakness. Thats very different from STARS to con core as almost everything fundamentally changed.

and just to give a link:
http://www.notebookcheck.net/fileadmin/_processed_/csm_carrizo_excavator_features_e7874bf3c5.jpg

4 to 15% higher IPC

Some good selectively memory from the select few........

i think there is room for optimism, i dont know how some people in here muster the will to drag themselves out of bed, "realist" is the excuse, cant justify that "realist" position with technical facts...........

*david kanter has commented that the arrays themselves are actually quite fast but CMT plus cache policy creates layout issues and WCC doesn't help latency only throughput.

 

[TurtleCrusher]

you will tell us that your livelihood does depend on Intel doing well but that, because you are so incredibly objective

That's exactly the case. Intel is by the biggest contract union electrician employer in Oregon. Thousands of our living-wage jobs heavily rely on Intel expanding their fabs out in Hillsboro year around. If Intel slumped and halted construction, that'd put me and thousands of skilled electricians on the books. D1X alone has brought billions in cash flow to the metro area.

So yeah, my livelihood does depend on Intel to do well, far more than most people here, and yet I remain open minded. If that's too complicated for you to understand then I'll break it down for you.

 

[ShintaiDK]

So yeah, my livelihood does depend on Intel to do well, far more than most people here, and yet I remain open minded. If that's too complicated for you to understand then I'll break it down for you.

Believing a hype based on a single unspecified metric from a company who notoriously lied with every new major performance CPU uarch the last 10 years?

You call that open minded?

Everything point to that AMD cant deliver what some people fanatically expect.

 

[ShintaiDK]

There has seemingly been nothing unreasonable said about Zen and it's performance. The IPC increase from Northwood to Conroe was what, 100%? a 40% claim isn't crazy.

The difference was more like ~40%. It also gave a clock regression.

Even more important, Conroe was a relatively small leap over Dothan/Yonah. You know, the Pentium-M and Core CPUs that kept K8 out of laptops.

If this is how you remember history, then I can see why you think Zen will perform as hyped.

 

[mrmt]

Then why post, yeah maybe you should see yourself out . We know bits and pieces but the biggest problem with the nay sayers is they cant tell you whats good or whats bad about bulldozer, just that bulldozer suxors.

Nope, the reasons are:

- The company is underfunded;

- It lied about every single product I can remember with a straight face, so they have a well deserved credibility issue;

- The company is only talking about upsides and not a thing about downsided. For example, how much power consumption we'll get?

- Globalfoundries.

 

[Gikaseixas]

The company is underfunded

Only AMD and it's partners know how much is needed to finance this project. AMD generally spends less than it's competitors and that might not be a negative. While budget size is critical, it is not only determinant key performance indicator by any means

It lied about every single product I can remember with a straight face

Really, every? WOW

The company is only talking about upsides and not a thing about downsided

Why would any company talk bad about it's own produts??? And what makes you think that Zen's power consumption numbers are worst compared with the competition's??? It's all speculation and yet you decided it is a downside.

Globalfoundries

Completely agree here, these guys at GF will most likely f...up once again

 

[Abwx]

Completely agree here, these guys at GF will most likely f...up once again

No need to jump on the deffamation waggon, numbers are known, they did better than Intel actualy..

GF s 14nm LPP LVT is 20% more efficient at 2.4GHz than the process used for SKL, and the LPP sLVT will extend this lead even further while allowing frequencies well over 3GHz.

 

[ShintaiDK]

Only AMD and it's partners know how much is needed to finance this project. AMD generally spends less than it's competitors and that might not be a negative. While budget size is critical, it is not only determinant key performance indicator by any means

You do not get what you didn't pay for. So its certainly not positive.

 

[frozentundra123456]

Believing a hype based on a single unspecified metric from a company who notoriously lied with every new major performance CPU uarch the last 10 years?

You call that open minded?

Everything point to that AMD cant deliver what some people fanatically expect.

Exactly. Those who are skeptical do not have to "prove" anything. It is up to those making the ever inflating claims about a future product, from AMD or Intel, to "prove" the claims are accurate. I havent seen any benchmarks from objective review sites showing AMD reached this theoretical expected 40% IPC increase.

In fact, I think it is highly likely that AMD can reach this IPC improvement in some selected "best case" benchmark. It is not really that difficult if you pick a benchmark that Vishera is particularly bad at, and it may (most likely) will be at a sacrifice in clockspeed. We have absolutely no idea, despite page after page of theoretical speculation, what overall level of performance and power usage it will achieve.

 

[Abwx]

Exactly. Those who are skeptical do not have to "prove" anything. It is up to those making the ever inflating claims about a future product, from AMD or Intel, to "prove" the claims are accurate. I havent seen any benchmarks from objective review sites showing AMD reached this theoretical expected 40% IPC increase.

In fact, I think it is highly likely that AMD can reach this IPC improvement in some selected "best case" benchmark. It is not really that difficult if you pick a benchmark that Vishera is particularly bad at, and it may (most likely) will be at a sacrifice in clockspeed. We have absolutely no idea, despite page after page of theoretical speculation, what overall level of performance and power usage it will achieve.

You are making "predictions" based on a psychological perception of yours while other people here look at the published caracteristics for their previsions.

As said ad nauseam a Zen core is twice a EXV core for integer and twice a module for FP, so in this respect 40% is not surprising at all and is even modest.

 

[Erenhardt]

Exactly. Those who are skeptical do not have to "prove" anything. It is up to those making the ever inflating claims about a future product, from AMD or Intel, to "prove" the claims are accurate. I havent seen any benchmarks from objective review sites showing AMD reached this theoretical expected 40% IPC increase.

In fact, I think it is highly likely that AMD can reach this IPC improvement in some selected "best case" benchmark. It is not really that difficult if you pick a benchmark that Vishera is particularly bad at, and it may (most likely) will be at a sacrifice in clockspeed. We have absolutely no idea, despite page after page of theoretical speculation, what overall level of performance and power usage it will achieve.

For the good of all here, based on your post, I advice you to leave any Zen thread until it is finally released. Don't expect any reviews until that time. Any discussion with you on this matter is futile until then.
See you soon™.

 

[frozentundra123456]

Until the mods say otherwise, I will post in any thread I wish, thank you.

 

[tential]

The same way I am not expecting AMD to meet the 40% improvement in IPC, I'm not expecting intel to make a huge IPC jump, or to be on time for any of their processors as they've continually delayed recently too.

It isn't just a bash AMD. AMD has set expectations, and so has intel.

If Elon Musk told me he'd give me a $20,000 Tesla in 2018, I'd schedule him in for 2020.

 

[Headfoot]

I don't see why 40% IPC is so hard to believe -- it will almost certainly come at the expense of core frequency. 40% IPC while clocking at the same rates is what is hard to believe. AMD's last few releases have actually delivered the IPC they said it would, decent amounts, but each time the clock ceiling came down. I find it hard to believe that AMD will deliver both much higher IPC and higher clocks. Both the current Intel architectures scale well (Skylake, Has-E) so id be very surprised if this got surpassed

 

[Abwx]

Until the mods say otherwise, I will post in any thread I wish, thank you.

Do what you want but it s obvious that your posts bring no added value to this thread, quite the contrary, all you re posting is the forever fud that seems the specialty of some anti AMD members by there.

As said psychological perceptions have nothing to do in a technical debate, so either you are able to produce technicaly substancied arguments or else all you ll achieve is to reduce the signal/noise ratio in the thread.

 

[coercitiv]

For the sake of the argument, what is the minimum increase a Zen core must gain in IPC in order to reach the same throughput as XV? (considering SMT brings smaller benefits than CMT)

 

[AtenRa]

I don't see why 40% IPC is so hard to believe -- it will almost certainly come at the expense of core frequency. 40% IPC while clocking at the same rates is what is hard to believe. AMD's last few releases have actually delivered the IPC they said it would, decent amounts, but each time the clock ceiling came down. I find it hard to believe that AMD will deliver both much higher IPC and higher clocks. Both the current Intel architectures scale well (Skylake, Has-E) so id be very surprised if this got surpassed

Yes, by using inferior node process than the last one,

Trinity was made using the 32nm SOI, then Kaveri went to 28nm Planar and Carrizo to 28nm Planar HDL.

With ZEN we are going to from 28nm HDL to 14nm FinFets. So we could have both higher IPC and increase in clocks. Especially in low voltage mobile chips like current 15-30W TDP Carrizo.

 

[AtenRa]

For the sake of the argument, what is the minimum increase a Zen core must gain in IPC in order to reach the same throughput as XV? (considering SMT brings smaller benefits than CMT)

As i said before, if you have 40% increase in IPC then your second thread scaling from SMT should be at least 60% to keep the same throughput as Excavator Module at the same clocks. That assumes you have 100% scaling from the CMT for simplicity reasons.

Im starting to believe that AMD claim of 40% increase of IPC is not about Single Thread Performance but general Integer instruction (ADD, MUL etc) latencies perhaps ???

 

[Abwx]

For the sake of the argument, what is the minimum increase a Zen core must gain in IPC in order to reach the same throughput as XV? (considering SMT brings smaller benefits than CMT)

I guess that you re talking of a module throughput, in that case the IPC requirement including HT must be 90%, that s why i said that the 40% are related to ST perf without HT.

Now if we look at what is known from this uarch the theoricaly possible improvement in throughput over an EXV core is 100% in integer.

For FP it s different since the FPU is no more shared and that Zen will have a FPU twice as powerfull (theoricaly) as the one in EXV, so on a core per core basis in MT environment the peak throughput of a Zen core is theoricaly 4x the one of an EXV core.

 

[BigDaveX]

Yes, by using inferior node process than the last one,

Trinity was made using the 32nm SOI, then Kaveri went to 28nm Planar and Carrizo to 28nm Planar HDL.

With ZEN we are going to from 28nm HDL to 14nm FinFets. So we could have both higher IPC and increase in clocks. Especially in low voltage mobile chips like current 15-30W TDP Carrizo.

I think Headfoot was speaking from a historical perspective. K10 managed to get single-thread performance to about Conroe's level, but ended up being released at such low clock speeds that it didn't have a prayer of beating the Core 2 Quad (even before you factor in the TLB bug and the fact that Intel had moved onto Penryn). And before that, K8 delivered spectacularly on IPC, but launched at clockspeeds that weren't quite where they needed to be; though on that occasion AMD handled the situation smartly, by focusing on getting Opteron established in the server market and abandoning the Athlon XP to getting smacked around by Northwood-C until Athlon 64 was ready to come in.