(Rumour) - Zen chips have already been tested, met all expectation, no significant bottlenecks

[shawno]

This post is so ridiculous.

Have to say its all mainly true.

If amd release a processor equal with intels haswell, then whos going to spend x amount of money transitioning from haswell to zen? Theyd be doing it purely for the brand.

Id say that even intel would find it hard selling skylake as the performance gained from upgrading from haswell is barely seen.

Fury still gets its mentions but the fury nano suits a niche market not mainstream so it's presstime is pretty limited.

[MageTank]

On my phone, so sorry if I don't directly quote people. To clarify my previous statement, I'll start with what we've seen so far.

So far, we've seen a Four FPU block chart for Zen. Next, we see rumors Of AMD using 1 FPU per core. Then we see "up to 95w TDP". Yes. The magical words "up to". By now, anyone that has dealt with ISP's should be shaking. We have not seen any evidence yet, of a 8c, 16t Zen sku. Even if we do, that 95w TDP is an impossible dream for what they are hyping. Physics just won't allow it. There is a reason the X99 chips have 135w TDP's. Even Broadwell X99 will have high TDP. Unless one of you can point me towards reliable evidence of a 8c, 16t Zen sku, I am inclined to believe it won't exist. Though, I would love to be wrong.

[Dabombinable]

On my phone, so sorry if I don't directly quote people. To clarify my previous statement, I'll start with what we've seen so far.

So far, we've seen a Four FPU block chart for Zen. Next, we see rumors Of AMD using 1 FPU per core. Then we see "up to 95w TDP". Yes. The magical words "up to". By now, anyone that has dealt with ISP's should be shaking. We have not seen any evidence yet, of a 8c, 16t Zen sku. Even if we do, that 95w TDP is an impossible dream for what they are hyping. Physics just won't allow it. There is a reason the X99 chips have 135w TDP's. Even Broadwell X99 will have high TDP. Unless one of you can point me towards reliable evidence of a 8c, 16t Zen sku, I am inclined to believe it won't exist. Though, I would love to be wrong.

Skylake's enthusiast/workstation socket CPU might be close, but they'll still be far from 95W.

[Tomsen]

All these predictions are based on such an abstract value, IPC. IPC is nothing more than a marketing term now.
Yes, it is easy to apply third grade math on, to try extrapolate zens possible performance.
But again, the math are based on marketing. Not actual valid information.
 
@looncraz
+1
 
Only miniscule amount of actual information about zen-uarch is released. By far not enough to base an actual conclussion on its performance.
Zen is a wider core than haswell. Doesn't necessarily mean it will be better.

 
 
@MageTank

Zen does have 4 FP units. 2x 128bit  FMUL + 2x 128bit FADD.

The one FPU per core, is to clarify that it no longer shares the FPU with another core. Like with bulldozer.

 

What are AMD hyping?
Seems like the hype is comming from the community, not AMD.

 

8-core 95w is doable. However, it doesn't really say much if the cores are running at 2GHz.

I think we can expect stock at around ~3.5GHz.

 

We will probably see you wrong. Only the future can tell.

[MageTank]

All these predictions are based on such an abstract value, IPC. IPC is nothing more than a marketing term now.

Yes, it is easy to apply third grade math on, to try extrapolate zens possible performance.

But again, the math are based on marketing. Not actual valid information.

 

@looncraz

+1

 

Only miniscule amount of actual information about zen-uarch is released. By far not enough to base an actual conclussion on its performance.

Zen is a wider core than haswell. Doesn't necessarily mean it will be better.

.

 

 

@MageTank

Zen does have 4 FP units. 2x 128bit  FMUL + 2x 128bit FADD.

The one FPU per core, is to clarify that it no longer shares the FPU with another core. Like with bulldozer.

 

What are AMD hyping?

Seems like the hype is comming from the community, not AMD.

 

8-core 95w is doable. However, it doesn't really say much if the cores are running at 2GHz.

I think we can expect stock at around ~3.5GHz.

 

We will probably see you wrong. Only the future can tell.

 

That 40% IPC number came from AMD. You yourself just called IPC marketing. What is marketing used for? Creating hype to sell products.

 

I want to be wrong about the 8c, 16t part. It would do AMD a world of good to be able to compete with Z97 and X99 at the same time. I just feel we won't see that specific Sku ( if it exists) until it's too late

[Tomsen]

That 40% IPC number came from AMD. You yourself just called IPC marketing. What is marketing used for? Creating hype to sell products.

I want to be wrong about the 8c, 16t part. It would do AMD a world of good to be able to compete with Z97 and X99 at the same time. I just feel we won't see that specific Sku ( if it exists) until it's too late

I'm fully aware of the origin of the IPC increase. Does that AMD marketing slide make a clear idea on zen performance?

The problem I see, is when the community take marketing as technical details, and think they can produce a realistic guestimate within x% error, and thinking their math is bullet-proof. AMD might have made a spark, that esculated to a forest fire, because of the community.

 

I cant think of a reason why AMD would not offer a 8c/16t model at launch, with 4c/8t (and maybe/hopefully 6c/12t). Maybe the 8c/16t will be more expensive than many might have thought (I dont think they go after bulldozers-pricing). I also really dont believe we will see 8c/16t in MCM.

[DarkBlade2117]

It's like when the Titan X came out and everyone was buying it left right and centre expecting to get 100fps in Crysis 3.

But the gaming community is just dumb in general and they'll just go straight for the newest or most expensive stuff in hope that it's the best.

Exactly, people want newer and better but... They never do the research to see the actual gains they will see from it. Like some mentioned earlier, he would get Zen even if it barely held ul agisnt Haswell. Why?....

[ManWithBeard1990]

Mate, even the 8370e has a TDP of 95 watts. The question isn't so much whether they can make it happen, but rather how much it'll throttle in the process... I hope not a lot. What's more worrying about a low TDP like this is AMD traditionally used thermal paste for the integrated heat spreader for everything up to 95W and solder for everything above it, so overclocking will probably not be as good as it used to be.

[patrickjp93]

I'm fully aware of the origin of the IPC increase. Does that AMD marketing slide make a clear idea on zen performance?

The problem I see, is when the community take marketing as technical details, and think they can produce a realistic guestimate within x% error, and thinking their math is bullet-proof. AMD might have made a spark, that esculated to a forest fire, because of the community.

 

I cant think of a reason why AMD would not offer a 8c/16t model at launch, with 4c/8t (and maybe/hopefully 6c/12t). Maybe the 8c/16t will be more expensive than many might have thought (I dont think they go after bulldozers-pricing). I also really dont believe we will see 8c/16t in MCM.

The math is bulletproof. It all comes down to how honest AMD is being. I take it you never passed any proof-based math course if that's your attitude. You can prove it's bulletproof. As to whether or not the common person understands why a proof is valid...most people here don't understand mathematical induction and proof by contradiction, much less more advanced techniques.

 

We're not getting a 6-core SKU. Zen's been designed in 4-core modules (we have the chip block diagrams straight from AMD confirming this). We're also not getting a single-chip 8-core SKU, only MCMs.

[patrickjp93]

Mate, even the 8370e has a TDP of 95 watts. The question isn't so much whether they can make it happen, but rather how much it'll throttle in the process... I hope not a lot. What's more worrying about a low TDP like this is AMD traditionally used thermal paste for the integrated heat spreader for everything up to 95W and solder for everything above it, so overclocking will probably not be as good as it used to be.

The 8370e doesn't have a dedicated FPU and AVX unit per core. It's a whole new ballgame on a node with hot transistors compared to the competition (Samsung's licensed tech, designs vs. TSMC and Intel).

[cesrai]

the only company with shit stock coolers are intel. Do not even try to argue this.

Intel cooler is shit. The end.

No they aren't, my pentium cooler is quieter than my Athlon stock cooler.

You're the only guy I know that calls Intel stock coolers shitty. They do their job well and they are quiet.

[Tomsen]

The math is bulletproof. It all comes down to how honest AMD is being. I take it you never passed any proof-based math course if that's your attitude. You can prove it's bulletproof. As to whether or not the common person understands why a proof is valid...most people here don't understand mathematical induction and proof by contradiction, much less more advanced techniques.

 

We're not getting a 6-core SKU. Zen's been designed in 4-core modules (we have the chip block diagrams straight from AMD confirming this). We're also not getting a single-chip 8-core SKU, only MCMs.

When the math is based on marketing, very abstract/vague figures are giving, you think you can turn that into actual data?

You are a marketing whisper

 

I'm aware of that. Hence my careful statements regarding the 6c/12t SKU. They could implement some kind of method to disable cores in the modules, to increase yield. Why on earth would AMD manufacturer small 4c CPU dies?

Have great doubts that we will see 8c in MCM. 

[patrickjp93]

When the math is based on marketing, very abstract/vague figures are giving, you think you can turn that into actual data?

You are a marketing whisper

I'm aware of that. Hence my careful statements regarding the 6c/12t SKU. They could implement some kind of method to disable cores in the modules, to increase yield. Why on earth would AMD manufacturer small 4c CPU dies?

Have great doubts that we will see 8c in MCM.

No, I'm saying if the marketers tell the truth, this is the best you could possibly get. It's the same concept of Big-O notation in asymptotic time and memory complexity analysis for algorithms. It tells you the maximal case, but that's not precise enough for a full report, which is why we also have Big Omega for the minimal and Big Theta for the average case, and we even have amortized analysis for a case by case basis analysis. I'm saying the math is bulletproof. How you choose to interpret it and use it is the only variable.

Because 4C dies increase yields and reduce costs, even if there is overhead associated with communication between the modules. It's IBM's very effective strategy instead of using monolithic large dies to control costs and maximize margins per chip. There's no reason to abandon what works if there is no better alternative.

[Castdeath97]

AMD already lost so much money..... does Zen really matter at this point?

 

Just like the hyped Fury that no one talks about anymore.... I see more threads about the 980ti but hardly any about the overpriced Fury cards now that the hype has all but fizzled.

 

Zen would have to beat any current Intel CPU by at least 10-15% to make anyone care about it. 8 cores sounds fantastic if things can use those cores. If they can't... then it's just a waste. Just recently games have been shown to use quad cores the best. This was on LTT. Even having more cores didn't matter in games. Dual cores finally are a weak spot but 16 cores.... Hmm...

 

Nano was also a bust because it's price point was ludicrous. People saying it's decent at $580 (With rebate are high)

 

We shall see I guess.

The R9 fury isn't overpriced, in fact it offers better value than the 980. I agree in the fury X.

[Tomsen]

No, I'm saying if the marketers tell the truth, this is the best you could possibly get. It's the same concept of Big-O notation in asymptotic time and memory complexity analysis for algorithms. It tells you the maximal case, but that's not precise enough for a full report, which is why we also have Big Omega for the minimal and Big Theta for the average case, and we even have amortized analysis for a case by case basis analysis. I'm saying the math is bulletproof. How you choose to interpret it and use it is the only variable.

Because 4C dies increase yields and reduce costs, even if there is overhead associated with communication between the modules. It's IBM's very effective strategy instead of using monolithic large dies to control costs and maximize margins per chip. There's no reason to abandon what works if there is no better alternative.

There are no "truth" to the marketing game. The "truth" have many sides, and all can be considered the "truth", and the "truth" can be deceptive.

As said, don't even try taking marketing information as technical details.

 

The problem is, that if there are fault in a die, (and there is no way to implement a method to disable cores within the module), you wasted a die.

There are no fallbacks. 8 core per die seems much more reasonable. If a module is not functioning, cut it off. There you get the 4c CPU. Also it increase cost of 8c, 16c, .. SKU.

I'm not saying they wont be using MCM. I'm saying it is rather ridiculous in that exact config. But we will see. Only time can tell.

[patrickjp93]

There are no "truth" to the marketing game. The "truth" have many sides, and all can be considered the "truth", and the "truth" can be deceptive.

As said, don't even try taking marketing information as technical details.

 

The problem is, that if there are fault in a die, (and there is no way to implement a method to disable cores within the module), you wasted a die.

There are no fallbacks. 8 core per die seems much more reasonable. If a module is not functioning, cut it off. There you get the 4c CPU. Also it increase cost of 8c, 16c, .. SKU.

I'm not saying they wont be using MCM. I'm saying it is rather ridiculous in that exact config. But we will see. Only time can tell.

No way to disable cores in a module? Intel had 3-core MCMs built of Core 2 Duo dies. It's very doable.

[Tomsen]

No way to disable cores in a module? Intel had 3-core MCMs built of Core 2 Duo dies. It's very doable.

I'm aware of that, thanks.

It should have read as if AMD haven't implemented a method to reliably disable core(s) within the 4c module. Not as there are no way to implement it.

Which would make a 6c/12t SKU doable.

[patrickjp93]

I'm aware of that, thanks.

It should have read as if AMD haven't implemented a method to reliably disable core(s) within the 4c module. Not as there are no way to implement it.

Which would make a 6c/12t SKU doable.

I just don't see that SKU coming in the immediate term. Yields would have to be sufficiently bad as to provide a supply of them.

[Tomsen]

I just don't see that SKU coming in the immediate term. Yields would have to be sufficiently bad as to provide a supply of them.

Volume would just have to be high enough, for long enough time.

Maybe they wont have enough at launch, at will wait it out. We will see. 

[alphaproject]

The R9 fury isn't overpriced, in fact it offers better value than the 980. I agree in the fury X.

Yea, I forgot the X part.... But Nano is so case specific and expensive.

 

So the regular Fury was the only decent Fury card IMO @ 550.

[Dabombinable]

No they aren't, my pentium cooler is quieter than my Athlon stock cooler.

You're the only guy I know that calls Intel stock coolers shitty. They do their job well and they are quiet.

With the original TIM and the cooler being used on my i5 4440, my entire rig was silent (Gigabyte GTX 650ti OC 2GB Windforce Rev 2.0-fans never go above idle the cooler is that good, plus the rear case fan running at minimum RPM)

[looncraz]

The ISA is known. It's guaranteed to be ISA-compatible with Haswell. Cache latencies are pretty much known and given at this point. The only variable there will be data retention times which I don't expect to exceed 300 milliseconds. Penalties for misprediction have also already been taken into account.

No, a 40% IPC improvement over Excavator is a best-case simulation since that is exactly what AMD promised and since no semiconductor company ever lives up to its promises. With a 3-tier associative cache, we know within a tight range what AMD will come up with for cache timings and misprediction penalties.

8 cores at 4GHz in 95W with AVX2 capability on 14nmFF from GloFo. Are you raving mad? It'll never happen.

 

Sorry for the lateness of my reply, this forum doesn't notify me of comments by email...

 

The INTERNAL ISA is *NOT* known.  Do you know of what I speak?  It's the language used inside the CPU, not the one used to communicate WITH the CPU.

 

We know AMD uses FastPath and is improving upon it for Zen, we don't know what those improvements are - this effects specific instruction latencies and resource utilization (including ALU/AGU/scheduler entries/etc...).

 

Cache latencies are known how?  We know absolutely nothing about what to expect from Zen's cache latencies, AFAICT.  Tell me, what are the L1I/D latencies?  What about the L2 read/write latencies? L3 latencies?  Give me a link to show me from where these numbers are derived, because just a few cycles difference can change the picture for latency-critical executions.

 

Next, how do you even know what the misprediction penalties will be?  What is the nominal misprediction penalty? Is it 13 cycles? 17 cycles? 30 cycles?  What is the 95% worst-case penalty?  The reason you give means nothing.  k10, 15h, i5/7, and many more CPUs all use three-tiered caches and have VERY different penalties and latencies... you can draw no conclusions from this information.

 

Seriously, how many ways does the L2 have? What is the write policy? Sure, we can make assumptions, but these are not known quantities and each adds a degree of uncertainty.

 

Finally, I never said anything about 95W for the 8-core CPU.  They could probably do it, though.  I mean, you can currently buy 95W AMD 8-core CPUs (FX-8320E, 3.2GHz~4.0GHz), so not even that is out of the question for Zen, though I doubt they will actually deliver such a thing.  I expect 125~150W for a high end 8-core with SMT at reasonable clock speeds.

[looncraz]

On my phone, so sorry if I don't directly quote people. To clarify my previous statement, I'll start with what we've seen so far.

So far, we've seen a Four FPU block chart for Zen. Next, we see rumors Of AMD using 1 FPU per core. Then we see "up to 95w TDP". Yes. The magical words "up to". By now, anyone that has dealt with ISP's should be shaking. We have not seen any evidence yet, of a 8c, 16t Zen sku. Even if we do, that 95w TDP is an impossible dream for what they are hyping. Physics just won't allow it. There is a reason the X99 chips have 135w TDP's. Even Broadwell X99 will have high TDP. Unless one of you can point me towards reliable evidence of a 8c, 16t Zen sku, I am inclined to believe it won't exist. Though, I would love to be wrong.

 

It's not impossible at all, just clock it to 2.5GHz

 

X99 chips also have extra PCI-e lanes, a quad channel memory controller, and other power-hungry bits.

 

AMD already has a 95W 8-core CPU.  On 32nm.  Shrink that to 16/14nm FF+/LPP and power usage drops immensely.  Get rid of the huge, power-hungry, L2 and drop even more watts off the rating.  Engage in more power gating and more restrictive turbo configurations (compared to Intel) and it's easy to see how AMD could produce a Zen-based 8-core CPU with only a 95W draw.  I doubt they will do so for their mainstream CPUs (likely will be the 'E' models such as they have now), but they could certainly do so.

[Jeffrey Chiu_081]

http://wccftech.com/amd-zen/

 

-better source + images 

[MageTank]

It's not impossible at all, just clock it to 2.5GHz

 

X99 chips also have extra PCI-e lanes, a quad channel memory controller, and other power-hungry bits.

 

AMD already has a 95W 8-core CPU.  On 32nm.  Shrink that to 16/14nm FF+/LPP and power usage drops immensely.  Get rid of the huge, power-hungry, L2 and drop even more watts off the rating.  Engage in more power gating and more restrictive turbo configurations (compared to Intel) and it's easy to see how AMD could produce a Zen-based 8-core CPU with only a 95W draw.  I doubt they will do so for their mainstream CPUs (likely will be the 'E' models such as they have now), but they could certainly do so.

Yeah, i see what you mean. Though, i thought it was rumored that Zen was supporting quad channel memory as well? Though, it was a rumor (and an old one, at that) so i won't be putting much faith into that being true. 

 

All i know is, too many assumptions are being made on nothing but rumors. None of us can possibly know what Zen will do, until we see it. We have several pieces of the puzzle, but all of the pieces we are missing, are the most important pieces to figure out exactly what we are looking at. Everything discussed in this thread is just guesswork at the moment. Nothing more.

 

I do however, have faith that Zen will succeed IF it can match Haswell. It does not need to match Skylake, because frankly, Skylake is only 5% faster than Haswell in the areas that really count. All other area's can be contributed to DDR4. With that 11% number from before, coming from compression (Raw memory bandwidth). Zen will gain this boon as well, with DDR4. Then again, i am only adding more guesses to an already enigmatic subject. I'll stop talking now.