AMD launches Zen+ 12nm Ryzen and X470 motherboards

[Atari2600]

Sorry what is GPP?

Some Nvidia partner thing where they screw over everyone else.

 

[Topweasel]

Sorry what is GPP?

Depends on how you look at it. Really its Nvidia's attempt to own and control sub brands that other companies have developed.

Things like early access to hardware and engineer support have been normally given. Now it's behind a paywall called GPP. The cost is giving up your "gaming brand" and maybe even gaming title. Companies only offering Nvidia hardware will eat this up in a heartbeat (EVGA), so this causes the rest of the companies to want to join in for fear of being left behind and possibly even lose allocation of chips.

So when you sign up for GPP, you basically agree that your gaming hardware can only be "Geforce" and not Radeon. Their idea is they want people to know that they are getting the "Geforce Experience" when buying hardware with brands associated with the GPP. That means that Gigabyte and MSI who have demonstratively joined the GPP can no longer sell AMD products labeled with their popular brands like Aorus and MSI with their "Gaming" line. Their is some debate on whether even the term or Gaming is also disallowed on AMD based products. Both these companies recently dumped their current brandings of AMD cards on their sites to either without a sub brand or a brand originally branded for lost cost offerings.

As for Asus we have to wait and see. They were an early company we thought had signed with them, but they have not done any re-branding of RoG products on their site and personally if I was them I would be pissed. They worked tirelessly for over a decade building the RoG brand and now Nvidia is basically telling them how and what they can offer through it.

 

[Ancalagon44]

Thanks guys.

Kinda disgusting and I hope that AMD sues about it. Anyway, wrong thread for this.

 

[The Stilt]

How much faster is Skylake per clock cycle?

According to my latest test, Skylake-X has 14.6 - 28% higher IPC than Ryzen (post Spectre & Meltdown fixes).

 

[dnavas]

According to my latest test, Skylake-X has 14.6 - 28% higher IPC than Ryzen (post Spectre & Meltdown fixes).

Is that due to vector? That's ... a lot more than I would have guessed.

 

[The Stilt]

Is that due to vector? That's ... a lot more than I would have guessed.

14.634% excluding all >= 256-bit workloads, 23.632% excluding single extremities (low / high), 28.002% including all workloads in the suite (30 workloads).

 

[rainy]

According to my latest test, Skylake-X has 14.6 - 28% higher IPC than Ryzen (post Spectre & Meltdown fixes).

With the same level of software optimization?

 

[Agent-47]

The cost is giving up your "gaming brand" and maybe even gaming title. Companies only offering Nvidia hardware will eat this up in a heartbeat (EVGA), so this causes the rest of the companies to want to join in for fear of being left behind and possibly even lose allocation of chips.
...
these companies recently dumped their current brandings of AMD cards on their sites to either without a sub brand or a brand originally branded for lost cost offerings.

a major problem of this is the loss of the best cooling tech. e.g. Twin Forzer from MSI is no longer available for AMD cards as Gaming X branding is now reserved for nv: https://www.msi.com/Graphics-cards/#?tag=Radeon^TM-RX-Vega⁶⁴

 

[The Stilt]

With the same level of software optimization?

The same generic binary for all of the tested CPUs.
If the workload supported instruction runtime detection (e.g. Blender, libAV, libGMP, X264, X265) then obviously the feature was used.

 

[CatMerc]

The same generic binary for all of the tested CPUs.
If the workload supported instruction runtime detection (e.g. Blender, libAV, libGMP, X264, X265) then obviously the feature was used.

Delta quite a bit more than I expected. AMD has their work cut out for them with Zen 2.

 

[The Stilt]

Delta quite a bit more than I expected. AMD has their work cut out for them with Zen 2.

The biggest issue on Summit Ridge was the limited frequency, not the IPC itself.
We'll see how much Pinnacle Ridge manages to remedy that, but obviously the clocks will still be behind Intel regardless.

At this point it is rather futile to speculate what the 7nm node will bring, but frankly I don't expect that AMD can match Intel (in consumer segment) as long as GlobalFoundries is used as a process provider.
It's been more than half a decade since GlobalFoundries was able to provide even a remotely competitive manufacturing process (the matured 32nm SHP). AMD definitely knows how to swim, but swimming becomes immensely
hard when you have a cinder block tied to your limb, dragging you down.

I'd wish the next Zen iteration would be a wider design, since the >= 256-bit workloads is where current Zen designs get annihilated by recent Intel designs.
AVX512 appears to be rather useful so being able to execute those instructions at 1/2 rate compared to 1/4 rate would become handy, especially when Intel consumer parts are also going to support AVX512

 

[CatMerc]

The biggest issue on Summit Ridge was the limited frequency, not the IPC itself.
We'll see how much Pinnacle Ridge manages to remedy that, but obviously the clocks will still be behind Intel regardless.

At this point it is rather futile to speculate what the 7nm node will bring, but frankly I don't expect that AMD can match Intel (in consumer segment) as long as GlobalFoundries is used as a process provider.
It's been more than half a decade since GlobalFoundries was able to provide even a remotely competitive manufacturing process (the matured 32nm SHP). AMD definitely knows how to swim, but swimming becomes immensely
hard when you have a cinder block tied to your limb, dragging you down.

I'd wish the next Zen iteration would be a wider design, since the >= 256-bit workloads is where current Zen designs get annihilated by recent Intel designs.
AVX512 appears to be rather useful so being able to execute those instructions at 1/2 rate compared to 1/4 rate would become handy, especially when Intel consumer parts are also going to support AVX512

Gotta work with your constraints. If you don't believe your partner can deliver the process performance you're after, then make your design wide and extract every ounce of performance you can get per clock.

 

[dnavas]

The biggest issue on Summit Ridge was the limited frequency, not the IPC itself.

True, though I'm surprised to see 14%, I was expecting closer to 7%.

I'd wish the next Zen iteration would be a wider design, since the >= 256-bit workloads is where current Zen designs get annihilated ...

Yes, I'm surprised AMD wasn't more aggressive about 256bit support. If Zen2 doesn't get at least half-rate AVX512, that's going to make it hard to (self) justify a TR purchase.

 

[raghu78]

The biggest issue on Summit Ridge was the limited frequency, not the IPC itself.
We'll see how much Pinnacle Ridge manages to remedy that, but obviously the clocks will still be behind Intel regardless.

At this point it is rather futile to speculate what the 7nm node will bring, but frankly I don't expect that AMD can match Intel (in consumer segment) as long as GlobalFoundries is used as a process provider.
It's been more than half a decade since GlobalFoundries was able to provide even a remotely competitive manufacturing process (the matured 32nm SHP). AMD definitely knows how to swim, but swimming becomes immensely
hard when you have a cinder block tied to your limb, dragging you down.

I'd wish the next Zen iteration would be a wider design, since the >= 256-bit workloads is where current Zen designs get annihilated by recent Intel designs.
AVX512 appears to be rather useful so being able to execute those instructions at 1/2 rate compared to 1/4 rate would become handy, especially when Intel consumer parts are also going to support AVX512

I think Globalfoundries has turned a corner after the IBM acquisition. I think 12LP and Zen+ will be the first evidence of that. GF 7LP is extremely competitive with Intel 10nm and TSMC 7nm. The way I see it Intel's manufacturing lead is definitely gone. Moreover predicting the future based on the past is not a good idea. Intel executed flawlessly for so many years and now stumbled so badly at 10nm that TSMC is going to be first to the next gen process node in 2018. btw i am talking about real products in very high volume and not some PR exercise which Intel did with CNL 2+0 in late 2017. The volume of TSMC N7 in 2018 will embarass Intel 10nm volume in 2018. I will also bet that Zen 2 for servers will launch before ICL-SP given that AMD has wisely moved to multi die at 14nm to avoid yield issues with massive dies at 7nm and beyond. Intel is unlikely to get to such a solution before Sapphire Rapids in 2020. Anyway we will know in a little over 3 weeks if GF has turned a corner.

 

[french toast]

14.634% excluding all >= 256-bit workloads, 23.632% excluding single extremities (low / high), 28.002% including all workloads in the suite (30 workloads).

The biggest issue on Summit Ridge was the limited frequency, not the IPC itself.
We'll see how much Pinnacle Ridge manages to remedy that, but obviously the clocks will still be behind Intel regardless.

At this point it is rather futile to speculate what the 7nm node will bring, but frankly I don't expect that AMD can match Intel (in consumer segment) as long as GlobalFoundries is used as a process provider.
It's been more than half a decade since GlobalFoundries was able to provide even a remotely competitive manufacturing process (the matured 32nm SHP). AMD definitely knows how to swim, but swimming becomes immensely
hard when you have a cinder block tied to your limb, dragging you down.

I'd wish the next Zen iteration would be a wider design, since the >= 256-bit workloads is where current Zen designs get annihilated by recent Intel designs.
AVX512 appears to be rather useful so being able to execute those instructions at 1/2 rate compared to 1/4 rate would become handy, especially when Intel consumer parts are also going to support AVX512

Wow that still quite a large difference, you do some good work with you reviews. However if we look at the execution resources available to both architecture's it is understandable, skylake is a little wider is it not, with more AGUs and a larger OoO engine (196 Vs 224??) Not including the Simd units. if anything I expected AVX instructions to be even more pronounced than that...still a big gap to make up.

Regards to global foundries, up untill now they have been crap, but 7nm is a joint effort with IBM IP/expertise. It looks competitive.
I agree with you zen 2 needs to be significantly wider as frequency scaling is going to hit a wall soon, probably around 5ghz on stock consumer tdps.
---Zen 2 I predict 6 core CCX, 4x SMT. extra execution units, probably bring the AGUs up to 3, increase the OoO engine, buffers, branch prediction, change the cache and probably increase the L2.
IPC integer 15-20%, FP 50%, SMT yield upto ~50%.
Clocks probably stay similar to they are now, maybe an extra 5% over pinnacle ridge...12/48 3.8/4.5ghz ...105-120w...all offset by 7nm HPC.
Edit; Maybe disable SMT 4 for mainstream desktop and push up frequencies instead, enable it on threadripper and Epyc..makes sense actually.

I think Globalfoundries has turned a corner after the IBM acquisition. I think 12LP and Zen+ will be the first evidence of that. GF 7LP is extremely competitive with Intel 10nm and TSMC 7nm. The way I see it Intel's manufacturing lead is definitely gone. Moreover predicting the future based on the past is not a good idea. Intel executed flawlessly for so many years and now stumbled so badly at 10nm that TSMC is going to be first to the next gen process node in 2018. btw i am talking about real products in very high volume and not some PR exercise which Intel did with CNL 2+0 in late 2017. The volume of TSMC N7 in 2018 will embarass Intel 10nm volume in 2018. I will also bet that Zen 2 for servers will launch before ICL-SP given that AMD has wisely moved to multi die at 14nm to avoid yield issues with massive dies at 7nm and beyond. Intel is unlikely to get to such a solution before Sapphire Rapids in 2020. Anyway we will know in a little over 3 weeks if GF has turned a corner.

I agree 7nm looks very good, 12nm very average imo.

 

[TreyJ93]

Yay... for the ~0.0001% of people that overclock.


Meanwhile, for the other ~99.9999% of us; the max turbo frequencies are approximately identical.

How much faster is Skylake per clock cycle?

It would actually benefit a lot of people especially Ryzen users to spend some time tinkering. Mindblanktech on YouTube got his 1700x to keep up and in some cases surpass a 5GHz 7700k once tuned. That's a hell of a lot of performance being left on the table.

I aprreciated not everyone has the time and some can't be bothered. Not everyone is a hard core enthusiast.

 

[IRobot23]

Sorry what is GPP?

Geforce Partner Program = Nothing good.

Well, I like that Precision boost 2.0. Most of users will only need to apply XMP on DRAM.

 

[IRobot23]

The biggest issue on Summit Ridge was the limited frequency, not the IPC itself.
We'll see how much Pinnacle Ridge manages to remedy that, but obviously the clocks will still be behind Intel regardless.

At this point it is rather futile to speculate what the 7nm node will bring, but frankly I don't expect that AMD can match Intel (in consumer segment) as long as GlobalFoundries is used as a process provider.
It's been more than half a decade since GlobalFoundries was able to provide even a remotely competitive manufacturing process (the matured 32nm SHP). AMD definitely knows how to swim, but swimming becomes immensely
hard when you have a cinder block tied to your limb, dragging you down.

I'd wish the next Zen iteration would be a wider design, since the >= 256-bit workloads is where current Zen designs get annihilated by recent Intel designs.
AVX512 appears to be rather useful so being able to execute those instructions at 1/2 rate compared to 1/4 rate would become handy, especially when Intel consumer parts are also going to support AVX512

Isn't AVX512 very inefficient?
Maybe ZEN 2 will be wider, but I don't think soo.

 

[Tuna-Fish]

GF 7LP is extremely competitive with Intel 10nm and TSMC 7nm.

The numbers published by GF marketing are very competitive with Intel 10nm and TSMC 7nm. The actual performancce of the node remains to be seen.

 

[raghu78]

The numbers published by GF marketing are very competitive with Intel 10nm and TSMC 7nm. The actual performancce of the node remains to be seen.

All we have are marketing numbers for all next gen nodes. Actual performance of Intel 10nm and GF 7LP will only be known when products built on them launch in 2019.

 

[iBoMbY]

I'd wish the next Zen iteration would be a wider design, since the >= 256-bit workloads is where current Zen designs get annihilated by recent Intel designs.
AVX512 appears to be rather useful so being able to execute those instructions at 1/2 rate compared to 1/4 rate would become handy, especially when Intel consumer parts are also going to support AVX512

I have seen AVX512 mentioned for Zen3, which should be for 2020. But with the recent rumors about a 16c/32t die for Zen2 (which was previously more than once seen as a 12c/24t die), I begin to wonder if they may release two different 7nm dies in 2019, more or less close together, with the bigger one already having AVX512?

 

[IRobot23]

All we have are marketing numbers for all next gen nodes. Actual performance of Intel 10nm and GF 7LP will only be known when products built on them launch in 2019.

This time GF has IBM.

 

[CatMerc]

It would actually benefit a lot of people especially Ryzen users to spend some time tinkering. Mindblanktech on YouTube got his 1700x to keep up and in some cases surpass a 5GHz 7700k once tuned. That's a hell of a lot of performance being left on the table.

I aprreciated not everyone has the time and some can't be bothered. Not everyone is a hard core enthusiast.

He does get some weird results all the time. I don't trust his benchmarks.

 

[IRobot23]

He does get some weird results all the time. I don't trust his benchmarks.

No actually he gets good results, because with LLC Ryzen is really fast. Some say IMC should get higher speeds, but it achieves 50GB/s easily.

With decent timings and speed you can actually achieve 55GB/s on dual channel... hopefully ryzen 2000 will be able to push near 60GB/s mark.

I think most server tasks are bandwidth limited, so maybe AMD could hurry up and deliver CPU with HBM. HBM could improve ST performance in most tasks... For gamers, AMD need to improve their design and improve DRAM latency, I fear that this InfinityFabric design is not design for that. IPC is good, but they should improve it or at least reach higher clocks.

 

[moinmoin]

According to my latest test, Skylake-X has 14.6 - 28% higher IPC than Ryzen (post Spectre & Meltdown fixes).

Meh, back when Skylake-X was detailed I called the design decision of heavily investing into wider AVX misguided and completely unsuitable for consumer processors, and I still stand by it. I'd rather AMD revives its heterogeneous computing initiative and find a way to transparently uses GPGPU for wider AVX SIMD instructions.

For comparisons, we are talking about 28% higher IPC at best while its cores (incl. L3 cache) are over 64% bigger, both at 14nm where Intel still has a significant advantage.