AMD CPUs, SoC Rumors and Speculations Temp. thread 2

[Vogner16]

If 14nm and 16nm don't mean what their name implies than whats the difference in debating between the two?
plz watch
https://www.youtube.com/watch?v=ktHciPVAZaU

Well people on this forum already know that but that doesn't mean they aren't competitors which means the latest fabrication from both companies and Intel will be compared. IMO and probably the industry as a whole Intel has the best followed by TSMC and global foundary's is in last. Also its just my opinion on the manner but i personally think Globalfoundries is a joke(most here probably know that's what i think) ever since bulldozer constant delays over not being able to allow the frequency Amd wanted.

Now in 2016 i wonder how Amd's latest design would perform on TSMC and i also wonder if we would see a true high-end card from Amd instead of this mid-range stuff. Also i haven't been impressed with overclocking with the 400 series either.

It has to be about money or something but i can't understand why they keep going to Global foundry.

The answer is "WSA". When AMD sold its foundry business, they had to sign an agreement with Global Foundries that obligates AMD to fabricate a given volume of chips on Globalfoundries.

The WSA was renegotiated recently and now AMD uses Globalfoundries 14nm for CPUs, GPUs, and APUs

http://www.amd.com/en-us/press-releases/Pages/amd-amends-wafer-2014apr1.aspx

If I am not mistaken the WSA expires on 2028.

This is correct,

but why was the WSA renegotiated? zen APU's!

 

[jdwii]

http://www.anandtech.com/show/10578/amd-zen-microarchitecture-dual-schedulers-micro-op-cache-memory-hierarchy-revealed

http://www.tomshardware.com/news/amd-zen-microarchitecture-summit-ridge,32508.html

http://www.extremetech.com/computing/233977-amd-unveils-new-details-for-its-next-generation-2017-zen-cpu-core

Nice written articles for what happen today not sure why no one posted this yet ha ha.

Anyways i'm impressed and the fact that it completed the job at the same amount of cores and clock rate as the 6900K makes me generally impressed.

I do find some of the comments generally foolish or absurd with the constant Amd hate or Amd super love. Barely anyone who just looks at it from a technical standpoint.

Something i would like to have explained a bit more is about the dual schedulers is that a advantage or disadvantage?

 

[juanrga]

It is being discussed in the Zen thread. I will copy and paste

Blender is FP heavy benchmark.

@ 3GHz Broadwell 8C gives a maximum throughput of 768 GFLOPS.
@ 3GHz Zen 8C gives a maximum throughput of 384 GFLOPS.

Therefore it is evident that they are playing with compiler settings to get Zen achieve the same performance than Broadwell-E. Moreover, Blender code mix is unusual and must be favoring the four half FP pipes of Zen over the two full pipes on Broadwell microarchitecture.

Excellent marketing move!

There is no absolute answer to your question about schedulers. It depends on the rest of the microarchitecture. In general single scheduler is better because optimizes the scheduling of operations to all the execution units; however, a dual scheduler reduces latency and allows for higher clocks (if everything else is the same; I am not saying that Zen will hit higher clocks than Intel, in fact it will not).

It is worth mentioning that I predicted dual schedulers. When Desdrenboy published the GCC patch leak for Zen, he draw a hypothetical diagram for Zen with three schedulers. I said him that integer and memory units would be unified in a single scheduler. He didn't trust me. :-D

 

[-Fran-]

Let me start by saying that back self-patting is a very disgusting and annoying thing to do, everyone.

Secondly, Zens' Blender benchies do paint a positive picture for me, but not the holy grail either. It does paint a good picture for servers, since the mixed nature of Blender reflects most of the commonly used software loads in all-purpose servers. We'll need to see some specific testing pitting the two in SPEC or any other decent bench suite aimed for the enterprise sector. On the consumer side, it's kind of a let-down so far. I will behave decently, that is great, but clocks is the great unknown for Zen. At 95W TDP, with the speculated ~3.4Ghz Turbo and ~3Ghz Base it's going to be a let down compared to any Intel offering, including the high clocked i3s even. I'm willing to bet, for games at least, it will be *at best* on parity with my i7 2700K at 4.6Ghz, if not still behind. In all fairness, all stock offerings are behind it, including new Intel CPUs. Then you have OC, where we haven't seen any information, but starting at 95W and seeing how the node is tuned for efficiency, every voltage step up is going to hurt a lot in terms of TDP. Moving it to 4Ghz, if what we've seen is remotely true, is going to be painful.

Oh well, Benchies can't get soon enough.

Cheers!

 

[ComputerSecurityGuy]

I might point out this is partially the issue of heat dissipation from massive eight core dies. Intel's octa core consumer chips, the i7-5960X and i7-6900K, are clocked at 3.0 and 3.3gHz respectively.

 

[juanrga]

I might point out this is partially the issue of heat dissipation from massive eight core dies. Intel's octa core consumer chips, the i7-5960X and i7-6900K, are clocked at 3.0 and 3.3gHz respectively.

But Intel cores are much bigger and had higher TDP rating because have 2x256bit SIMD units, whereas Zen has 2x128bit. About one year ago I predicted Zen wold measure about 4mm² and some people analyzed the Summit die recently to get about that number. Broadwell is about twice bigger.

 

[jdwii]

Let me start by saying that back self-patting is a very disgusting and annoying thing to do, everyone.

Secondly, Zens' Blender benchies do paint a positive picture for me, but not the holy grail either. It does paint a good picture for servers, since the mixed nature of Blender reflects most of the commonly used software loads in all-purpose servers. We'll need to see some specific testing pitting the two in SPEC or any other decent bench suite aimed for the enterprise sector. On the consumer side, it's kind of a let-down so far. I will behave decently, that is great, but clocks is the great unknown for Zen. At 95W TDP, with the speculated ~3.4Ghz Turbo and ~3Ghz Base it's going to be a let down compared to any Intel offering, including the high clocked i3s even. I'm willing to bet, for games at least, it will be *at best* on parity with my i7 2700K at 4.6Ghz, if not still behind. In all fairness, all stock offerings are behind it, including new Intel CPUs. Then you have OC, where we haven't seen any information, but starting at 95W and seeing how the node is tuned for efficiency, every voltage step up is going to hurt a lot in terms of TDP. Moving it to 4Ghz, if what we've seen is remotely true, is going to be painful.

Oh well, Benchies can't get soon enough.

Cheers!

Personally to be honest with you I'm not that concerned about the clock rate its going to be lower then Intel and it will have less performance per core over that reason but if Amd did everything right it should be cheaper to produce as well. Like with any other X86 design that's made right IPC is more important in most cases as long as the clock rate isn't to low. 3Ghz is still considered a speedy design.

Can always improve clock speeds with fabrication tweaks and architecture tweaks but IPC don't expect much. Juan even said overclocks on even the average zen could do 4.5Ghz or so that's not to bad of course power consumption will go through the roof probably. I'm still conerned with IPC that one blender test shows how it would perform in FP but what about integer tasks(90+% of everything uses).

Hate the wait for benchmarks i mean why does Amd even wait like this if they already have samples i remember Intel doing paper launches, its not like it does much to Amd if they did release it people with intelligence will look at the benchmarks before buying it anyways and it surely isn't going to make their current stuff sell worse.

 

[-Fran-]

jdwii, let me provide some context on what I am saying:

http://www.tomshardware.co.uk/skylake-intel-core-i7-6700k-core-i5-6600k,review-33276-7.html

Check the stock i7 6700K and the i7 5930K (6C/12T). Given the test is obviously different since the rendering used must be, I'll put the Zen CPU right between the 5960X and the 5930K (maybe even slightly below the 5960X?). Then look at Cinebench single threaded. You can extrapolate a bit from those two pieces of information and making some educated guesses you might see what I see.

That is why I'm a bit worried about the clocks. I mean, I will welcome with open arms a better uArch that doesn't need to clock to ~5Ghz to be viable, but at ~3Ghz does not appeal to me one bit as a regular consumer John Doe TBH hitting those numbers. On the upside, if that bench is indicative, then it will double the performance of the FX9590 at 95W (vs 220W!).

Cheers!

 

[ComputerSecurityGuy]

Can we just settle the fact that oddly enough, Zen is looking to have comparible clocks to Intel octa-core chips and for that matter, Broadwell-E. Zen at 3.2gHz or so is faster than the i7-5960X and a wee bit slower than the i7-6900K. Poor overclocking ability is likely, but Broadwell-E is not exactly good at same. My conclusion is that Zen is pretty much on par with Broadwell, and *maybe* more efficient (but with AMD being so unpredictable, I am going to wait for power consumption in real life to be figured out.).

 

[jdwii]

Can we just settle the fact that oddly enough, Zen is looking to have comparible clocks to Intel octa-core chips and for that matter, Broadwell-E. Zen at 3.2gHz or so is faster than the i7-5960X and a wee bit slower than the i7-6900K. Poor overclocking ability is likely, but Broadwell-E is not exactly good at same. My conclusion is that Zen is pretty much on par with Broadwell, and *maybe* more efficient (but with AMD being so unpredictable, I am going to wait for power consumption in real life to be figured out.).

Based on just that blender test alone the FP performance is doing quite well and the 8 core Zen is supposed to have a 95 watt TDP so i call that an amazing feat since the Intel part uses more power even. But that is for just FP and that's probably just 256Bit not 512 like Intel. However with Zen+ or even a newer Zen adding something such as a capable 512bit FPU is possible i mean isn't that basically what intel did?

But again not having that capability means nothing to me as a consumer personally, however for the server business and some other high-end tasks it does matter even more so when people buy servers and keep them the same for 5 years+.

I personally want to see how it does in more memory intense integer applications which is what 90+% of applications/consumers use. I am however amazed over what Amd even did with their FP performance however as that was their weakness with the bulldozer family.

Other news Amd stock is doing quite well and based on paper alone Lisa Su is doing a great job. I also said that i hoped Amd themselves were not blind like their die hardcore fanbase and i think they proved they aren't. Amd isn't claiming single threaded performance doesn't matter over directx 12 or magic or something they realize the truth... it all matters. Wide cores with as many cores as possible and a high clock speed is and well remain the future for X86/X64 CPUs.

Edit I also want to see how their 4 core Zen comes out when it comes to frequency hoping for at least 3.4Ghz+. Again i still make my claims of sandy-ivy based IPC in integer tasks.

 

[Ags1]

Zen has 1.4 times more IPC than BD, but BD clocked at 4.2GHz, 1.4 times faster than than Zen.

So hasn't AMD stood still, at least in terms of single core performance?

 

[-Fran-]

Zen has 1.4 times more IPC than BD, but BD clocked at 4.2GHz, 1.4 times faster than than Zen.

So hasn't AMD stood still, at least in terms of single core performance?

Well, you also have to take into account "modern bits" of the design itself. At least for the server world, BD and PD were never a great design because of power. Now Zen can do the same, but at a fraction of the power. Hence why my fear is only for the regular consumer segment. I'm sure even the "pro-sumer" will be happy with Zen.

We just have to sit tight and wait for more benchies

Cheers!

 

[juanrga]

Juan even said overclocks on even the average zen could do 4.5Ghz or so that's not to bad of course power consumption will go through the roof probably.

What I have said is that I doubt that average Zen chip will hit 4.5GHz even with watter cooling. Golden chips could do it.

 

[juanrga]

Zen has 1.4 times more IPC than BD, but BD clocked at 4.2GHz, 1.4 times faster than than Zen.

So hasn't AMD stood still, at least in terms of single core performance?

AMD used a heavily multithreaded benchmark and comparison to 8-core Broadwell for obvious reasons.

I am anxiously waiting for comparisons of Zen to FX-9590. The 9590 has ~50% higher clocks.

 

[cdrkf]

@ags1, zen I'd said to have 40% greater ipc than excavator, which is 3 generations newer than bulldozer and already had a 15 to 20% ipc advantage. Compared to first gen bulldozer is expect more like 60% more ipc, significantly more in floating point math as well as zen has double the fp units.

 

[jdwii]

Juan even said overclocks on even the average zen could do 4.5Ghz or so that's not to bad of course power consumption will go through the roof probably.

What I have said is that I doubt that average Zen chip will hit 4.5GHz even with watter cooling. Golden chips could do it.

Ah sorry for the confusion, overall its using a specialize finfet technology with clock gating that is why it will be a poor overclocker its designed specifically within a certain frequency.

Something i noticed with owning piledriver and then owning haswell even if i clocked my CPU down to 2.5Ghz it would still outperform my older 8350 in most things despite having on paper similar single core performance. Other areas in the design that have improved for example Amd's poor branch predictor now has many improvements this goes beyond just IPC.

The technical aspects of a CPU is more then just frequency+IPC even if that is the most important. CPI for example is important to.

 

[juanrga]

Juan even said overclocks on even the average zen could do 4.5Ghz or so that's not to bad of course power consumption will go through the roof probably.

What I have said is that I doubt that average Zen chip will hit 4.5GHz even with watter cooling. Golden chips could do it.

Ah sorry for the confusion, overall its using a specialize finfet technology with clock gating that is why it will be a poor overclocker its designed specifically within a certain frequency.

Something i noticed with owning piledriver and then owning haswell even if i clocked my CPU down to 2.5Ghz it would still outperform my older 8350 in most things despite having on paper similar single core performance. Other areas in the design that have improved for example Amd's poor branch predictor now has many improvements this goes beyond just IPC.

The technical aspects of a CPU is more then just frequency+IPC even if that is the most important. CPI for example is important to.

CPI is just the inverse of IPC. Say there is a piece of code with 3360 instructions executed in 10000 cycles, then

CPI = 10000/3360 = 2.97
IPC = 3360/10000 = 0.33

We usually tend to write performance as product of IPC and frequency and take both as independent quantities. However, in reality IPC has a small dependence on the frequency. Therefore when underclocking the chip from 4GHz to 2.5GHz you are increasing the effective IPC gap between both microarchitectures from 60% to something as 65--70% probably (I don't know the exact number).

 

[jdwii]

Juan even said overclocks on even the average zen could do 4.5Ghz or so that's not to bad of course power consumption will go through the roof probably.

What I have said is that I doubt that average Zen chip will hit 4.5GHz even with watter cooling. Golden chips could do it.

Ah sorry for the confusion, overall its using a specialize finfet technology with clock gating that is why it will be a poor overclocker its designed specifically within a certain frequency.

Something i noticed with owning piledriver and then owning haswell even if i clocked my CPU down to 2.5Ghz it would still outperform my older 8350 in most things despite having on paper similar single core performance. Other areas in the design that have improved for example Amd's poor branch predictor now has many improvements this goes beyond just IPC.

The technical aspects of a CPU is more then just frequency+IPC even if that is the most important. CPI for example is important to.

CPI is just the inverse of IPC. Say there is a piece of code with 3360 instructions executed in 10000 cycles, then

CPI = 10000/3360 = 2.97
IPC = 3360/10000 = 0.33

We usually tend to write performance as product of IPC and frequency and take both as independent quantities. However, in reality IPC has a small dependence on the frequency. Therefore when underclocking the chip from 4GHz to 2.5GHz you are increasing the effective IPC gap between both microarchitectures from 60% to something as 65--70% probably (I don't know the exact number).

That's it then, i was just saying its more then just frequency + IPC. Other things come into play such as even the cache improvements and memory controller IPC technically just stands for instructions per cycle(which instructions exactly is being used matters to)

Most programs don't even use above 2IPC i think but the branch predictor and other things matter more that increase the performance per cycle. Amd always had a worse branch predictor which makes the CPU do more work as you already know.

I really want to see how Zen performs for integer based applications.

 

[-Fran-]

http://www.anandtech.com/show/10631/amd-amends-globalfoundries-wafer-supply-agreement-through-2020

Interesting how AMD is owning mistakes and paying green blood for them. At least they now have a bit more freedom for Fabbing and it should be interesting to see how it pans out.

Also, reading between the lines, if they went ahead with this payment, that means the Quarter is not going that bad. It will be interesting to see the next Q financial report for AMD.

EDIT: After a closer read, GloFo seems to have a better end of the deal, lol.

Cheers!

 

[jimmysmitty]

I still think GloFlo was one of the worst mistakes AMD made, not as bad as their untimely purchase of ATI but still a bad choice.

It has not benefitted them much and it looks like GloFlos 14nm might be inferior to TSMCs 16nm. Guess we will see how it fares against Intels 14nm.

 

[ComputerSecurityGuy]

If the TDP and clock figures are accurate, not bad. If Zen has a 95W TDP and a 3.2gHz clock or so (that's my best guess at final chip speed) that makes GloFo presumably reasonably competitive with Intel.

 

[jimmysmitty]

It will depend though. Remember there was a time when Intels TDP was a vast over estimation (I remember the QX9650 needing 4GHz to hit 135w power draw with a 130w TDP) and AMDs was an average estimation. I will have to wait and see power numbers and thermals before I judge.

It is even hard to judge with Polaris because a GPU is a very different beast power wise so GloFlos 14nm might be decent for CPUs but bad for GPUs.

 

[ComputerSecurityGuy]

Yeah, I think microarchitecture is bigger here. The only good comparison we have, courtesy of Apple, is that TSMC has a slightly more efficient process than GloFo. Anything else has issues with the simple fact that not all microarchitectures are the same.

 

[juanrga]

I still think GloFlo was one of the worst mistakes AMD made, not as bad as their untimely purchase of ATI but still a bad choice.

It has not benefitted them much and it looks like GloFlos 14nm might be inferior to TSMCs 16nm. Guess we will see how it fares against Intels 14nm.

Intel 14nm > TSMC 16nm > Samsung/Glofo 14nm.

 

[juanrga]

If the TDP and clock figures are accurate, not bad. If Zen has a 95W TDP and a 3.2gHz clock or so (that's my best guess at final chip speed) that makes GloFo presumably reasonably competitive with Intel.

If Glofo was competitive, companies wouldn't be avoiding it like the plague and AMD wouldn't be paying $300M to Glofo to acquire the rights to fabricate the products on other foundries.

About TDPs:

AMD obviously didn’t state what specific SKUs were going to launch with the Zen architecture, what clock speeds they would run at, or even what TDPs they were targeting. Instead we were left with a vague but understandable remark of “comparable TDPs to Broadwell-E”.

https://www.pcper.com/reviews/Processors/AMD-Zen-Architecture-and-Performance-Preview