New Zen microarchitecture details

[Tup3x]

Yea im expecting the 6c 12T at close to $299.

In my dreams: 1 core/50 €. So... 2 core ~100 €, 4 core ~200 €, 6 core ~300 €, 8 core ~400 €. Über mega WC 8c edition >500 €.

 

[Atari2600]

Here's a thought. Maybe the F3 stepping and the F4 stepping are identical in terms of bug fixed, the only difference being clocks.

Very possible.

- Or the F3 stepping contains bugs in certain parts which can be fused off and retained as 6C12T or 4C8T parts.
- Or the F3 stepping contains bugs in SMT control which can be fused off for 8C8T/6C6T/4C4T parts.
- Or the F3 stepping contains bugs that mean XFR doesn't work correctly, leading to these being basement parts without full turbo.
- Or something else.

 

[CatMerc]

https://imgur.com/a/X4r4Y

Previously unseen slides. Put them all into one gallery for ease of access.

 

[inf64]

Very possible.

- Or the F3 stepping contains bugs in certain parts which can be fused off and retained as 6C12T or 4C8T parts.
- Or the F3 stepping contains bugs in SMT control which can be fused off for 8C8T/6C6T/4C4T parts.
- Or the F3 stepping contains bugs that mean XFR doesn't work correctly, leading to these being basement parts without full turbo.
- Or something else.

I think that latest stepping performs the same as CanardPC's ES (which had the workarounds for uop cache and SMT bugs deactivated). CanardPC measured around 35% advantage for 1C/2T Zen vs same clock 1M/2T PD which roughly translates to ~50-55% better ST IPC vs XV core. That is now reaffirmed via sweclockers leak that latest Zen ES are having 55% better IPC vs XV, so AMD beat the figure they publicly stated by around 1.55/1.4=1.1 or 10%. Just about enough to be around ~5% behind Skylake since majorD already showed us his results where Skylake is around 61% faster than XV at the same clock. This also means that Broadwell core is also likely a bit faster then Zen, but not by much on average.

 

[krumme]

https://imgur.com/a/X4r4Y

Previously unseen slides. Put them all into one gallery for ease of access.

Very interesting stuff from my perspective.
I am actually surpriced it have taken so many years to get such a flexible and scalable solution like this stretching multiple dies and even sockets. Its kind of a basic technology platform you need to use all your building blocks to handle the complexity and speed up ttm. Really what was needed to bring your cpu gpu and memory technology platforms into play.
This is actually the first time we see something that make the ATI acquisition make sense. Bolting a cpu and gpu together on the same die inventing hsa and calling it synergy when complexity and dev cost exploded was a bad excuse and old school thinking. This is more it. Great stuff.

 

[coercitiv]

so AMD beat the figure they publicly stated by around 1.55/1.4=1.1 or 10%.

Mhmmm, not quite, that's either 0.55/0.4=37.5% in relative terms or simply 1.55-1.4=15% in absolute terms.

 

[inf64]

Mhmmm, not quite, that's either 0.55/0.4=37.5% in relative terms or simply 1.55-1.4=15% in absolute terms.

Nope but I don't want to get into math discussion and derail this thread for no reason . In short, XV is 100, New Zen is 155 , old Zen is 140. New Zen is 155/140 faster or 10%. The end .

 

[coercitiv]

Nope but I don't want to get into math discussion and derail this thread for no reason

Then don't answer and/or don't make "mic drop" statements. The fact that "New Zen" is 10% faster than "Old Zen" does not equate with AMD beating their claim of +40% IPC by 10%.

 

[raghu78]

Mhmmm, not quite, that's either 0.55/0.4=37.5% in relative terms or simply 1.55-1.4=15% in absolute terms.

Then don't answer and/or don't make "mic drop" statements. The fact that "New Zen" is 10% faster than "Old Zen" does not equate with AMD beating their claim of +40% IPC by 10%.

Dude you need to get your basic math right.

Excavator = 100
AMD Zen (inital estimates) = 1.4
AMD Zen (Final silicon) = 1.55

AMD beat out their original estimates for Zen by (1.55 - 1.4 )/1.4 * 100 = 10.7% . If you cannot understand it atleast do not argue on a tech forum about it.

 

[coercitiv]

If you cannot understand it atleast do not argue on a tech forum about it.

A CPU maker estimate their new under development CPU to be 10% faster than their old model. At launch time they announce they beat their own estimate by 100%. How fast is the new CPU relative to the old model? Good thing we're on a tech forum where people have common interests and can calmly discuss numbers.

 

[Atari2600]

Lies, damn lies and statistics.

Your all right, from different points of view. [/kenobi]

 

[leoneazzurro]

It's quite simpe: if you are referring to the absolute Zen performance, it went up around 10% with respect to the expectations.
If you are refferring to the performance delta increase to the old model, it went up 37.5% more than the expectations.
You are simply looking at different points of view.

 

[inf64]

A CPU maker estimate their new under development CPU to be 10% faster than their old model. At launch time they announce they beat their own estimate by 100%. How fast is the new CPU relative to the old model? Good thing we're on a tech forum where people have common interests and can calmly discuss numbers.

New CPU is 21% faster than the old model and 1.1 or 10% faster than the estimated/announced model. Dunno what that has to do with Zen since AMD never stated by how much they beat their estimate. We now have a figure from sweclockers and a number from CanardPC and we know exactly by how much: 1.55/1.4=1.107. Zen on the shelves (if 55% number holds) will be ~10% faster than what AMD promised, simple as that.

 

[Gideon]

A CPU maker estimate their new under development CPU to be 10% faster than their old model. At launch time they announce they beat their own estimate by 100%. How fast is the new CPU relative to the old model? Good thing we're on a tech forum where people have common interests and can calmly discuss numbers.

Dude, percentage calculation is slightly trickier than it sounds. It all depends on what you take as baseline, but more often than not, the baseline is the last value.

It all depends on the wording as well. If "AMD beats their own estimate by 10%", then people pretty much always take the estimate (140 in this case) as the base value.

 

[Doom2pro]

https://imgur.com/a/X4r4Y

Previously unseen slides. Put them all into one gallery for ease of access.

Looks like AMD took all their IP and created a sort of API (Rather ASIC Programming Library)/Library around it so they can easily combine different parts together to make a product. This is going to be most useful in their semi-custom business I suspect.

 

[CatMerc]

Holy hell you guys, you just entered possibly the most pointless discussion in this entire thread. None of your math is wrong, you're just looking at different things.

 

[zinfamous]

Holy hell you guys, you just entered possibly the most pointless discussion in this entire thread. None of your math is wrong, you're just looking at different things.

I wanted to post this. I'm glad someone did. We need new leaks because you guys are running around your heads arguing about the most nonsensical and irrelevent crap.

 

[ecogen]

I wanted to post this. I'm glad someone did. We need new leaks because you guys are running around your heads arguing about the most nonsensical and irrelevent crap.

Don't know what else we can get from leaks at this point. We probably know final clocks, ballpark performance, and ballpark launch date.

Edit: I guess OC numbers?

 

[inf64]

Don't know what else we can get from leaks at this point. We probably know final clocks, ballpark performance, and ballpark launch date.

Edit: I guess OC numbers?

We don't know the prices of SKUs and prices of motherboards. But I bet they will be way more affordable than competition.

 

[ecogen]

We don't know the prices of SKUs and prices of motherboards. But I bet they will be way more affordable than competition.

If Zen delivers and the 6c is at mainstream i7 price, I'll be a very happy man.

 

[Gideon]

We don't know the prices of SKUs and prices of motherboards. But I bet they will be way more affordable than competition.

We also don't know much about the XFR, except it scales with temperature. I personally hope it can also dynamically adjust voltage. This would provide a really flexible auto-overclocking when combined with the Steamroller Adaptive Clocking feature.

Basically something similar to GPU boost.

In my dream scenario each processor would have:

• A list of not-to-exceed voltages per temperature bin. The lower the temperature, the higher the voltage (No2 might allow more than water, water more than air, etc ...), on the other hand, if temps go crazy the CPU starts to undervolt.
• If there is load, clocks will go as high as possible (max might be a never-reachable, 6Ghz) while being constantly throttled by the Adaptive Clocking feature to avoid crashes.
• Obviously there needs to be a configurable max Wattage limit (either socket, mb or PSU related), but if temps are the main regulator you don't need much else.

Overall, the idea is, rather than having "max turbo frequency", the CPUs would have guaranteed all-core base clock, and max allowed voltage (that is still covered by warranty and is set by AMD).
In reality both would be dependant on the aforementioned voltage/temperature lookup table.

Adaptive Clocking part is probably the hardest to pull off. Detecting the failure clockspeed in vastly different workloads, (ST vs MT, heavy floating point (AVX) workloads vs Integer, etc), but if that can be done, the rest shouldn't really be a problem.

 

[ultimatebob]

I wanted to post this. I'm glad someone did. We need new leaks because you guys are running around your heads arguing about the most nonsensical and irrelevent crap.

Yeah, some actual production product benchmarks would be nice right about now. This rampant speculation is killing me.

 

[raghu78]

Are you sure its not 32 PCIe and quadchannel per socket?

And the fabled server APU was nothing but vague slides 2 years ago.

http://wccftech.com/amd-zen-naples-server-cpu-vega-gpu-platform/

What was being told about Naples now seems to be true. 32C/64T, 8 channel DDR4 memory and 128 PCI-E 3.0 lanes. AMD really went big on this server platform. Intel Purley is running into a monster AMD server platform.

 

[CentroX]

Has AMD done the impossible? To be able to leapfrog Intel with Ryzen despite the fact that Intel has 10x the R&D and manpower? Or is yet another hypetrain that will crash at launch?

Lisa Su seems extremely confident.

http://finance.yahoo.com/news/amd-ceo-why-its-good-to-be-the-smaller-guy-195512478.html

 

[Phynaz]

Has AMD done the impossible?

No. By definition you can't do the impossible

It's the CEO's job to paint the company in the best possible light.