[AMD Processors] The future CPUs

[ShintaiDK]

14 nm would also mean higher clocks. See e.g.

3770K@22nm: 3.5/3.9 Ghz
6700K@14nm: 4.0/4.2 Ghz

So the net performance will be higher of Zen too @ 14 nm. You'd end up somewhere between SB and Skylake, i.e. IB/Haswell.

You really do the extreme selective compares when it suits your purpose.

What about the 4.0/4.4Ghz 22nm Haswell part?

40% IPC increase, if true, will mean a heavy clock regression. We can see on Excavator at 35W how much it decreased compared to Kaveri.

But lets see what actually gets delivered. Remember how much you are after people even if they use "may". Yet you talk in absolutes already. But again, you also made a claim that it needs to live up to

 

[Abwx]

Follow my logic here. I'll put it all in one place, to make it easier:

Using MT tests will bias the analysis, you should use ST tests to make a valuable comparisons and take account of the design peculiarities.

For one 40% is obviously an average, it will be less in Integer and likely more in FP, for instance we can be sure that the FPU will be reused since its throughput is as good as the one of a Haswell core.

Thus FP IPC can be increased by much more than those 40%, as it can be almost doubled if a single Zen core manage to extract close to its full throughput like a module does but for two threads.

 

[AtenRa]

You really do the extreme selective compares when it suits your purpose.

What about the 4.0/4.4Ghz 22nm Haswell part?

40% IPC increase, if true, will mean a heavy clock regression. We can see on Excavator at 35W how much it decreased compared to Kaveri.

But lets see what actually gets delivered. Remember how much you are after people even if they use "may". Yet you talk in absolutes already. But again, you also made a claim that it needs to live up to

Carizzo clocks were decreased vs Kaveri because of the process, 28nm HDL, not because of the IPC. If you port both Kaveri and Carizzo to 14nm FF they will get the same clocks.

 

[ShintaiDK]

Carizzo clocks were decreased vs Kaveri because of the process, 28nm HDL, not because of the IPC. If you port both Kaveri and Carizzo to 14nm FF they will get the same clocks.

I assume you are about to show the documentation for that. Because AMD says its not an issue if we are to believe their slides.

 

[Abwx]

40% IPC increase, if true, will mean a heavy clock regression. We can see on Excavator at 35W how much it decreased compared to Kaveri.

You should really stop with your fud, at 35W Excavator clock higher than Kaveri, now if AMD products do not interest you there s enough threads about the competing solutions where you can spill your usual junk analysis, and even found some believers...

I assume you are about to show the documentation for that. Because AMD says its not an issue if we are to believe their slides.

Because we should better trust your fuzzy explanations you think..?.

Well, at least you have some sense of humour, willingly or not...

 

[AtenRa]

I assume you are about to show the documentation for that. Because AMD says its not an issue if we are to believe their slides.

You mean you dont know that 28nm HDL is working better at lower power than High clocks ???

 

[ShintaiDK]

You mean you dont know that 28nm HDL is working better at lower power than High clocks ???

AMD writes that the crossover point is around 20-25W per core pair. Thats well below 35W for the package. Thats for HDL+new uarch.

 

[Abwx]

You completely missed the point in your fury.

IPC increase doesnt come for free.

That s right but they specified in which limits they have a power comsumption penalty.

That s why they stated 40% lower power comsumption at 4% higher IPC at ISSC, that was a subtle technical trap since it allowed them to publish the perf/watt improvement without disclosing the real IPC improvement, from wich we can deduct that the remaining 5-9% where compensated anyway by the process better efficency that is about 60%.

AMD writes that the crossover point is around 20-25W per core pair. Thats well below 35W for the package. Thats for HDL+new uarch.

Per core pair, that s 40W for two pairs, isnt it, plus the uncore that want 10W and we are at 50W even before accounting for a GPU that can use 20-25W...

 

[ShintaiDK]

Per core pair, that s 40W for two pairs, isnt it, plus the uncore that want 10W and we are at 50W even before accounting for a GPU that can use 20-25W...

You only prove me right.

 

[frozentundra123456]

If 4c/8t Zen is only 40% faster than Excavator...

... then it will not compete with Skylake, much less Kaby Lake or Cannonlake. AMD will need more than 40% over their current uarch to compete with the same number of cores/threads. Because of this, I'm hoping the 40% figure is incorrect.

Here we're seeing a Skylake i5 as being more than twice as fast as Kaveri. 40% more is simply not enough.

When did AMD ever underestimate projected performance increases? My feeling is that 40 percent is...... optimistic, not to mention nobody knows what kind of clockspeed they will be able to achieve. I also an with Shintai in that I will be very surprised if AMD can bring out a totally new architecture on a totally new process without delays, considering their minimal R and D budget. But maybe everything will come off perfectly and they will hit a home run. Maybe.

 

[Fjodor2001]

On the other hand, we saw a decline of max clockspeed (when overclocked) from SB to IB to HW to BW.

We did? Care to explain that?

If you're comparing 4790K to 5775C that is not relevant, since 5775C is a specialized SKU with big iGPU and relatively low CPU clocks.

Unfortunately there is no direct Broadwell comparison to 4790K, so you have to compare against Skylake 6700K.

 

[Thanatosis]

40% is an incredible jump but they have had a long time to do it.

 

[mrmt]

Damn nice. You should earn a Ph.d in moving the goal posts.

You were talking about delays -- and Intel has had more of them in recent history. So get off the soapbox and deal with it.

BTW, it's not like Intel is abandoning the Motherboard market or anything.... oh wait..... EPIC FAIL on your part.
http://www.zdnet.com/article/intel-to-leave-desktop-motherboard-business-by-2016/

Just more babble and fanboism's.

Of course Intel has more delays, they have more products than AMD. You have to develop a product to be able to delay it.

 

[Yuriman]

We did? Care to explain that?

If you're comparing 4790K to 5775C that is not relevant, since 5775C is a specialized SKU with big iGPU and relatively low CPU clocks.

Unfortunately there is no direct Broadwell comparison to 4790K, so you have to compare against Skylake 6700K.

Even with the iGPU inactive, most reviewers have not been able to get much more out of the CPU.

Example:

http://www.anandtech.com/show/9482/intel-broadwell-pt2-overclocking-ipc/11

There’s nothing much to be gained with overclocking either. Our i7-5775C CPU made 4.2 GHz, in line with Intel’s expectations for these processors. If we compare that to an overclocked 4.6 GHz i7-4790K, the 4790K is still the winner. Overclocking on these Broadwell CPUs still requires care, due to the arrangement of the CPU under the heatspreader with the added DRAM. We suggest the line method of thermal paste application rather than the large-pea method as a result.

 

[Fjodor2001]

Even with the iGPU inactive, most reviewers have not been able to get much more out of the CPU.

Example:

http://www.anandtech.com/show/9482/intel-broadwell-pt2-overclocking-ipc/11

So you were comparing to 5775C as expected, which is not relevant since it's a specialized SKU with big iGPU not designed for high CPU clocks.

You should be comparing to 6700K instead. That's the closest successor to 4770/4790K.

 

[FearfulSPARTAN]

Of course Intel has more delays, they have more products than AMD. You have to develop a product to be able to delay it.

Funny reading this after reading your back and forth argument:|.......

 

[Xpage]

I do think a majority of AMD's delays were in part to the incompetence of GF not being able to get their process ready in time or on target. Thus the tweaking the need to do for Samsung's process should hopefully be relatively simple given that it currently is working and Samsung can but not necessarily will hold GF's hand through the process of getting the Samsung process initially working. Then GF engineers can tinker to improve the process.

They do have about 9-12 months to work on this for AMD to get a 2H release of Zen (i'd say December timeframe but hope for August-Sept). Worst case AMD has to settle for a SOC process so performance will be lower. But guess that would mean a bigger bump for Zen+, assuming AMD survives and GF can improve the process for the refresh

 

[mrmt]

Funny reading this after reading your back and forth argument:|.......

Sorry, I was trying to be facetious.

But of course Intel has been having trouble as of lately, 14nm ramp up was bad enough and it seems that 10nm won't be any easier, so a lot of Intel roadmap is being reshuffled, I never denied that. On the Skylake thread I was pointing out how bad the 14nm situation was, and truth to be said I didn't think Intel was forthcoming in the way they displayed the information to investors (normalizing the BDW and IVB yields curve but at different points of the ramp up), and that's not good at all.

But there's a huge gulf between Intel reshuffling its roadmaps and AMD having entire business imploding and roadmaps being reshuffled every 6 months. AMD got a precipitous 70%+ drop in their consumer and server revenue in the last three years, that's a collapse by any standards, you cannot put Intel and AMD in the same basket at all.

 

[mrmt]

I do think a majority of AMD's delays were in part to the incompetence of GF not being able to get their process ready in time or on target. Thus the tweaking the need to do for Samsung's process should hopefully be relatively simple given that it currently is working and Samsung can but not necessarily will hold GF's hand through the process of getting the Samsung process initially working. Then GF engineers can tinker to improve the process.

I think people are having a lot of hope for a process that is built targeting SoCs, and hoping that Samsung will somehow speed up GLF adoption of their process, move that is basically against their own interests.

 

[VirtualLarry]

But there's a huge gulf between Intel reshuffling its roadmaps and AMD having entire business imploding and roadmaps being reshuffled every 6 months. AMD got a precipitous 70%+ drop in their consumer and server revenue in the last three years, that's a collapse by any standards, you cannot put Intel and AMD in the same basket at all.

What if AMD is just a "canary" for the entire PC market? Intel's been having some slow quarters too, lately. What if desktop performance no longer matters to the majority of the PC market, and what if mobile is not just "a" thing, but "the" thing, and Intel is going to waste their war-chest trying (unsuccessfully) to break into that market?

 

[mrmt]

What if AMD is just a "canary" for the entire PC market? Intel's been having some slow quarters too, lately. What if desktop performance no longer matters to the majority of the PC market, and what if mobile is not just "a" thing, but "the" thing, and Intel is going to waste their war-chest trying (unsuccessfully) to break into that market?

Intel doesn't have much choice except break into mobile or die trying, because this is where computing is going from here, mobile and later IoT. And while I think part of the issues with 14nm and 10nm nodes is Intel trying to generate a competitive enough advantage on mobile, another part of the issue is Intel trying to bring down costs as much as possible, even at the expense of performance gains, in order to shield their consumer PC business from the stagnation/decline of the next few years. I think they have a future at least as a foundry partner for the mobile companies.

AMD OTOH doesn't seem to have a bright future. They tried and failed to break into markets outside PCs, so whatever chance they have depends on taking on Intel.

 

[dark zero]

If Intel can't win the mobile market, they can RUIN it. Just look what they did to the Netbook market.

 

[DrMrLordX]

Kaveri A10-7770K and A10-7850K were announced in January 2014 and were available in retail stores immediately.

More recently, The A10-7870k launched on-time.

Are we then expecting Zen to have far, far higher clocks? I thought we were talking about quad vs quad.

Probably not, but we should expect its base clocks on 95W parts (such as the 8c/16t part) to be higher than the base clocks of Carrizo when constrainted to a 15-25W thermal envelope which is what we're seeing from a lot of notebook OEMs.

The bottom line is that you can't look at Carrizo in a gimped laptop where it's been hobbled, look at its absolute performance numbers, and then conclude that Zen will be a mere %40 faster than that.

Right now there aren't many comparisons between Carrizo and Kaveri, but hopefully that will change soon.

If Intel can't win the mobile market, they can RUIN it. Just look what they did to the Netbook market.

Uh.

Intel created the netbook market with Atom. The popularity of tablets paired with bluetooth keyboards ruined the netbook market.

 

[CHADBOGA]

If Intel can't win the mobile market, they can RUIN it. Just look what they did to the Netbook market.

We can only hope. :awe:

 

[Yuriman]

Probably not, but we should expect its base clocks on 95W parts (such as the 8c/16t part) to be higher than the base clocks of Carrizo when constrainted to a 15-25W thermal envelope which is what we're seeing from a lot of notebook OEMs.

The bottom line is that you can't look at Carrizo in a gimped laptop where it's been hobbled, look at its absolute performance numbers, and then conclude that Zen will be a mere %40 faster than that.

Right now there aren't many comparisons between Carrizo and Kaveri, but hopefully that will change soon.

Hence my choice to use AMD's statement about Excavator having 4-15% better IPC than Kaveri as one of the factors in my reasoning, rather than trying to figure out the performance differences myself.