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

[2FA]

Obviously there are more factors than just IPC and Clocks when it comes to CPUs, and unless you're a Nanoscale Engineer with experience in CPUs, don't claim or even act like you claim as much.

 

Performance from the i7-990X to the i7-5960X really only improved roughly 30% as be seen in this Anandtech article: http://www.anandtech.com/show/8426/the-intel-haswell-e-cpu-review-core-i7-5960x-i7-5930k-i7-5820k-tested/2 . The majority of the naysayers are quoting that 40% over Excavator and only thinking of it in terms of single-core. Imagine the multi-core potential of Zen, especially with its multi-threading. Consider DX12 and Vulkan, both are set to improve multi-core performance of processors in 3D applications. There are implications that Zen could be quite promising.

 

Just my two cents.

[Prysin]

It's obviously not good for overclocking.

 

The 4790K's in my dad's machine. I myself am running a 4.7GHz 2600K until Zen and Kaby Lake are out along with AI and Pascal so I can have my pick of upgrades. The stock cooler was perfectly fine and quiet when we were using it for MOBO testing before running the liquid cooling.

yeah... mobo testing...

 

as in idling..

 

it didnt hold up in well in gaming. Cinebench R15 turned it into a banshee.

[2FA]

it IS possible for AMD to sport 8 cores -> 16threads. because ZEN based FX will NOT feature onboard graphics.

So all the die space taken up by a traditional iGPU will be "free" to be used for more CPU cores.

 

Just look at X99... as soon as Intel removed the iGPU, they could slap more cores on there.

It is simply a matter of die area they can utilize for CPU cores... in Haswell and Skylake, the iGPU is nearly 40% of the die...

now 40% is A LOT of space...

that leaves 60% space left for Cache, cores, memory controller, video decoder/encoder and whatnot....

AMD confirmed that Zen will have dual threaded cores and that future generations could sport more threads per core.

[Notional]

Full stock clocks, about 80*, and not loud at all, and that was before we delidded.

 

And by stock clock you mean 4.0 GHz? 4.4 will increase it quite a bit. It of course depends on the case and the airflow in it too, but it is very loud, and by far the loudest part in my pc (glad I got a cheap CM AIO, it's SO much quieter). For someone claiming to have exceptional hearing, you should not be saying such things.

[patrickjp93]

 

Obviously there are more factors than just IPC and Clocks when it comes to CPUs, and unless you're a Nanoscale Engineer with experience in CPUs, don't claim or even act like you claim as much.

 

Performance from the i7-990X to the i7-5960X really only improved roughly 30% as be seen in this Anandtech article: http://www.anandtech.com/show/8426/the-intel-haswell-e-cpu-review-core-i7-5960x-i7-5930k-i7-5820k-tested/2 . The majority of the naysayers are quoting that 40% over Excavator and only thinking of it in terms of single-core. Imagine the multi-core potential of Zen, especially with its multi-threading. Consider DX12 and Vulkan, both are set to improve multi-core performance of processors in 3D applications. There are implications that Zen could be quite promising.

 

Just my two cents.

Amdahl's Law says hello. IPC will always be king even in multithreading for specifically this reason. The performance gained by throwing more cores at the problem is strictly limited by how much of the program is parallelizable, not to mention the overhead of actually launching threads and killing them which takes away from that potential.

 

 

And by stock clock you mean 4.0 GHz? 4.4 will increase it quite a bit. It of course depends on the case and the airflow in it too, but it is very loud, and by far the loudest part in my pc (glad I got a cheap CM AIO, it's SO much quieter). For someone claiming to have exceptional hearing, you should not be saying such things.

4.4 turbo boost single core, 4.3 for 2 cores, 4.2 for 4 cores, right on stock specs.

AIOs were always going to be cheaper.

Loud is still relative, and trust me nothing is as loud as the Thermalright cooler in my school desktop. Intel's stock cooler is half this volume.

 

 

yeah... mobo testing...

 

as in idling..

 

it didnt hold up in well in gaming. Cinebench R15 turned it into a banshee.

Sorry my experience says otherwise. It's always good to have a solid baseline so you can see what your end gains are when doing high end watercooling and overclocking. Under multithreaded Cinebench we were at 85 with no throttling, and that was the loudest it got, still quieter than the Thermalright Cooler I use at school.

[Prysin]

Sorry my experience says otherwise. It's always good to have a solid baseline so you can see what your end gains are when doing high end watercooling and overclocking. Under multithreaded Cinebench we were at 85 with no throttling, and that was the loudest it got, still quieter than the Thermalright Cooler I use at school.

well maybe you struck gold and got a chip using slightly lower stock voltage.... who knows.

[patrickjp93]

well maybe you struck gold and got a chip using slightly lower stock voltage.... who knows.

1.22 isn't exactly low.

[2FA]

Amdah's Law says hello. IPC will always be king even in multithreading for specifically this reason. The performance gained by throwing more cores at the problem is strictly limited by how much of the program is parallelizable, not to mention the overhead of actually launching threads and killing them which takes away from that potential.

That may be true, but we can't accurately predict as of this moment how Zen will perform. Better APIs are coming out that we haven't completely seen the impact of yet, afaik DX12 and Vulkan. We don't know the true extent of how well parallelized games and programs will become due to these APIs. You said it yourself that overheads exists which impact performance beyond that of just IPC. 

 

All I'm saying that is that stop treating this like a simple math equation, there are far more factors which impact performance than just IPC. *Note I didn't say how much they would impact because you can't accurately predict without actually knowing all of the details of the architecture.*

[Dabombinable]

what CPU do you own Patric?

 

because i own the Core i7 4790k... and the stock cooler does NOT keep my chip "cool and quiet"

Neither does the stock cooler for the FX 8350....

[patrickjp93]

That may be true, but we can't accurately predict as of this moment how Zen will perform. Better APIs are coming out that we haven't completely seen the impact of yet, afaik DX12 and Vulkan. We don't know the true extent of how well parallelized games and programs will become due to these APIs. You said it yourself that overheads exists which impact performance beyond that of just IPC. 

 

All I'm saying that is that stop treating this like a simple math equation, there are far more factors which impact performance than just IPC. *Note I didn't say how much they would impact because you can't accurately predict without actually knowing all of the details of the architecture.*

Oh yes you can. If you have QEMU and a lick of sense, you can easily predict with 0 margin for error other than AMD's actual projection of 40% IPC increase and the end clock speed. You can generate an exact range of performance possibilities tied to clock speeds.

 

No, it's a simple math equation where all you need are a finite number of facts. Since I know the least expensive way to launch a thread on another core, all it takes is summing up cycle latencies. If you know the:

• width and depth of the Out of Order Processing Engine (which is known at 8 wide by 14 deep, which is less than I expected),
• base case (Excavator instruction latencies),
• projection (40% IPC increase overall),

  constraints (some instructions can't execute in less than a number of cycles strictly greater than 1, and no instruction can execute in less than 1 or in a "mixed number" of numbers),

• memory interface (DDR4 quad channel),
• cache structure (3 layer, with the same L1 and L3 cache sizes as mainstream Intel I7s and double the L2 cache size at 512KB instead of 256),
• core count (clusters of 4 is confirmed at this point, where the 8-core SKUs will be MCMs)
• program environment (benchmarks, including all internal structure--which is why you should only use open source benchmarks--)

Then you can predict with 0 margin for error the performance per clock of any chip built on such an architecture as definable by the data above forgoing such issues as thermal throttling which are environmentally dependent. You can even take bandwidth and latency bottlenecks into account with ease. You can set this up as a virtual CPU in QEMU and test it yourself. Zen will have IPC roughly 3.2% lower than Haswell in the best case aggregate of open benchmarks. That will leave it needing 15% higher clocks over Skylake to be an even match in per-core performance.

 

Parallelism doesn't lay in the API. It lays in the design of the program overall. 

[DarkBlade2117]

I really hope people,just do not buy Zen the minute it is released. People get to hyped about stuff and are disappointed. People who have FX8s and 295x2 or 980Tis are gonna get Zen and think it brings a huge increase while on thenother hand people,with i7-4770ks or what not will not see a big improvement. This is theoretical, I have high hopes for Zen.

[2FA]

 

-snip-

As long as Zen isn't an overheating mess, then I could see AMD creating chips with that 15% higher clock rate. I never disputed that your numbers were wrong when solved for, all I said is that are unaccounted factors which could play a role in the real world performance of Zen and not the theoretical performance.

 

 

I really hope people,just do not buy Zen the minute it is released. People get to hyped about stuff and are disappointed. People who have FX8s and 295x2 or 980Tis are gonna get Zen and think it brings a huge increase while on thenother hand people,with i7-4770ks or what not will not see a big improvement. This is theoretical, I have high hopes for Zen.

I have a 4670k so even if Zen wasn't the INTEL SUPREME CRUSHORNATOR 2000, I would likely still get it for that multi-core performance.

[patrickjp93]

As long as Zen isn't an overheating mess, then I could see AMD creating chips with that 15% higher clock rate. I never disputed that your numbers were wrong when solved for, all I said is that are unaccounted factors which could play a role in the real world performance of Zen and not the theoretical performance.

But the point of the theoretical is that it is the perfect case. Almost nothing ever reaches the perfect case.

[2FA]

But the point of the theoretical is that it is the perfect case. Almost nothing ever reaches the perfect case.

And that is exactly my point, I take anything theoretical with a block of salt.

[patrickjp93]

And that is exactly my point, I take anything theoretical with a block of salt.

It's one thing to take it with a block of salt from a marketing team that product x will do better because of improvements A,B,C. You can accurately predict exactly the potential of those improvements separately and together, and with that you can establish a level of doubt. Anyone who predicts Zen beating Skylake and Kaby Lake hasn't run the numbers or seen the writing on the wall. With a 95W TDP ceiling for Zen's opening gambit, it's not reasonable to expect. Global Foundries is not known for having efficient transistors, and Samsung's 14nm is having growing pains for the high performance process. We can say with a high degree of definitiveness that Zen will fall short on overall performance. AMD will have to win on price. The only remaining question is how Intel responds. Give AMD plenty of berth to let it recover and give Intel time to push its MIC accelerator scheme while Nvidia and AMD duke it out, or bury AMD and go for the kill so Intel can buy Radeon instead of the longer gamble on being able to buy Nvidia?

[TechGod]

It's one thing to take it with a block of salt from a marketing team that product x will do better because of improvements A,B,C. You can accurately predict exactly the potential of those improvements separately and together, and with that you can establish a level of doubt. Anyone who predicts Zen beating Skylake and Kaby Lake hasn't run the numbers or seen the writing on the wall. With a 95W TDP ceiling for Zen's opening gambit, it's not reasonable to expect. Global Foundries is not known for having efficient transistors, and Samsung's 14nm is having growing pains for the high performance process. We can say with a high degree of definitiveness that Zen will fall short on overall performance. AMD will have to win on price. The only remaining question is how Intel responds. Give AMD plenty of berth to let it recover and give Intel time to push its MIC accelerator scheme while Nvidia and AMD duke it out, or bury AMD and go for the kill so Intel can buy Radeon instead of the longer gamble on being able to buy Nvidia?

May not perform as well as Skylake but I don't care. I will buy it if it get Haswell level performance

You are so pessimistic. My God.

[patrickjp93]

May not perform as well as Skylake but I don't care. I will buy it if it get Haswell level performance

You are so pessimistic. My God.

I'm realistic. Reality isn't all roses and sunshine.

[2FA]

 

It's one thing to take it with a block of salt from a marketing team that product x will do better because of improvements A,B,C. You can accurately predict exactly the potential of those improvements separately and together, and with that you can establish a level of doubt. Anyone who predicts Zen beating Skylake and Kaby Lake hasn't run the numbers or seen the writing on the wall. With a 95W TDP ceiling for Zen's opening gambit, it's not reasonable to expect. Global Foundries is not known for having efficient transistors, and Samsung's 14nm is having growing pains for the high performance process. We can say with a high degree of definitiveness that Zen will fall short on overall performance. AMD will have to win on price. The only remaining question is how Intel responds. Give AMD plenty of berth to let it recover and give Intel time to push its MIC accelerator scheme while Nvidia and AMD duke it out, or bury AMD and go for the kill so Intel can buy Radeon instead of the longer gamble on being able to buy Nvidia?

I thought they were using TSMC's 14nm?

 

You think if I went to the GloFo facility down the highway from me that if I asked they would tell me? lol

[patrickjp93]

I thought they were using TSMC's 14nm?

For Zen, GloFo. For Arctic Islands, it's believed to be TSMC's 16nm FF+

 

If AMD stopped being GloFo's customer, who would GloFo have left other than VIA and Xilinx? They have IBM now, but that's one specific foundry for 22nm FDSOI.

[TechGod]

I'm realistic. Reality isn't all roses and sunshine.

Answer me in the other thread about AMD's dual GPU.

[looncraz]

Oh yes you can. If you have QEMU and a lick of sense, you can easily predict with 0 margin for error other than AMD's actual projection of 40% IPC increase and the end clock speed. You can generate an exact range of performance possibilities tied to clock speeds.

 

No, it's a simple math equation where all you need are a finite number of facts. Since I know the least expensive way to launch a thread on another core, all it takes is summing up cycle latencies. If you know the:

• width and depth of the Out of Order Processing Engine (which is known at 8 wide by 14 deep, which is less than I expected),
• base case (Excavator instruction latencies),
• projection (40% IPC increase overall),

  constraints (some instructions can't execute in less than a number of cycles strictly greater than 1, and no instruction can execute in less than 1 or in a "mixed number" of numbers),

• memory interface (DDR4 quad channel),
• cache structure (3 layer, with the same L1 and L3 cache sizes as mainstream Intel I7s and double the L2 cache size at 512KB instead of 256),
• core count (clusters of 4 is confirmed at this point, where the 8-core SKUs will be MCMs)
• program environment (benchmarks, including all internal structure--which is why you should only use open source benchmarks--)

Then you can predict with 0 margin for error the performance per clock of any chip built on such an architecture as definable by the data above forgoing such issues as thermal throttling which are environmentally dependent. You can even take bandwidth and latency bottlenecks into account with ease. You can set this up as a virtual CPU in QEMU and test it yourself. Zen will have IPC roughly 3.2% lower than Haswell in the best case aggregate of open benchmarks. That will leave it needing 15% higher clocks over Skylake to be an even match in per-core performance.

 

Parallelism doesn't lay in the API. It lays in the design of the program overall. 

 

The margin of error in this calculation would be monstrous.  There are too many unknowns regarding internal ISA changes, cache latencies, changes in instruction latencies, mispredict penalties, mispredict and miss rates, and so on.

 

Using the Excavator as a source for instruction latencies limits you to a WORST CASE simulation, as Zen will undoubtedly improve upon many of these latencies by the pure fact of it being a wider design with updated logic.  There is no doubt AMD will focus on the most critical instructions, as well as the lowest hanging fruit - and many instructions are as fast as they can ever be (LEA, for example).  We also have 10 pipelines, not eight, though two are AGU and are limited to memory ops, they still have a meaningful impact on performance that must be taken into account.

 

My estimates have put Zen to be on par with Haswell almost spot on.  The lowest hanging fruit in Excavator* is worth 25% performance on their own, without adding two ALUs and two additional FPU pipes with a dedicated FPU or any of the numerous improvements that have been made we know nothing about (smart new uses for register renaming (to reduce mov penalties), stage bypasses [low penalty pipeline flush], improved prefetch and predictions, and more have all been either patented by AMD recently or have actually been stated by AMD as being in Zen).

 

*Module contention, the extra 5 for so stages before the EUs, the extra stage(s) after the EUs (WCC for integer, retirement for FPU to LSU plus the WCC - double wammy), the terrible caches, and so on...

 

Zen should match Haswell, IPC-wise.  I fear it will fail to match it clock-rate-wise.  I want ~Haswell IPC, 4GHz, 8 cores.  I will settle for Sandy/Ivy Bridge IPC, 4.2GHz, 6/8 cores, though, but obviously at a lower price.

[patrickjp93]

The margin of error in this calculation would be monstrous. There are too many unknowns regarding internal ISA changes, cache latencies, changes in instruction latencies, mispredict penalties, mispredict and miss rates, and so on.

Using the Excavator as a source for instruction latencies limits you to a WORST CASE simulation, as Zen will undoubtedly improve upon many of these latencies by the pure fact of it being a wider design with updated logic. There is no doubt AMD will focus on the most critical instructions, as well as the lowest hanging fruit - and many instructions are as fast as they can ever be (LEA, for example). We also have 10 pipelines, not eight, though two are AGU and are limited to memory ops, they still have a meaningful impact on performance that must be taken into account.

My estimates have put Zen to be on par with Haswell almost spot on. The lowest hanging fruit in Excavator* is worth 25% performance on their own, without adding two ALUs and two additional FPU pipes with a dedicated FPU or any of the numerous improvements that have been made we know nothing about (smart new uses for register renaming (to reduce mov penalties), stage bypasses [low penalty pipeline flush], improved prefetch and predictions, and more have all been either patented by AMD recently or have actually been stated by AMD as being in Zen).

*Module contention, the extra 5 for so stages before the EUs, the extra stage(s) after the EUs (WCC for integer, retirement for FPU to LSU plus the WCC - double wammy), the terrible caches, and so on...

Zen should match Haswell, IPC-wise. I fear it will fail to match it clock-rate-wise. I want ~Haswell IPC, 4GHz, 8 cores. I will settle for Sandy/Ivy Bridge IPC, 4.2GHz, 6/8 cores, though, but obviously at a lower price.

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.

[TheMechEngineer]

I really hope people,just do not buy Zen the minute it is released. People get to hyped about stuff and are disappointed. People who have FX8s and 295x2 or 980Tis are gonna get Zen and think it brings a huge increase while on thenother hand people,with i7-4770ks or what not will not see a big improvement. This is theoretical, I have high hopes for Zen.

 

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.

[alphaproject]

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.

[TechGod]

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.

This post is so ridiculous.