Zen hasn't taped out yet

[DrMrLordX]

One other thing to keep in mind when examining potential Zen memory bandwidth requirements: AMD has been losing memory bandwidth for years due to poor IMC performance. The last best-in-class IMC they had was from K8 (it had no competition in the x86 world). Intel has been consistently outperforming AMD in the IMC department since Nehalem. They got more out of DDR3-1600 than AMD. Zen needs to turn around that situation. If it does, then that's just another factor that could contribute to 8c/16t Summit Ridge existing quite happily with only two DDR4 channels. Even something as pedestrian as DDR4-2666 should work if they can produce better results from the Zen IMC.

 

[itsmydamnation]

One other thing to keep in mind when examining potential Zen memory bandwidth requirements: AMD has been losing memory bandwidth for years due to poor IMC performance. The last best-in-class IMC they had was from K8 (it had no competition in the x86 world). Intel has been consistently outperforming AMD in the IMC department since Nehalem. They got more out of DDR3-1600 than AMD. Zen needs to turn around that situation. If it does, then that's just another factor that could contribute to 8c/16t Summit Ridge existing quite happily with only two DDR4 channels. Even something as pedestrian as DDR4-2666 should work if they can produce better results from the Zen IMC.

It doesn't not really matter. latency is the most important thing for IPC and we only have piledriver to look at from there. throughput isn't that big of a deal, just look at things like 860 vs 920, 4820 vs 4770 etc to see how little difference that memory bandwidth makes. There are cases where for CPU bandwidth does matter, they just don't really exist in standard server or client workloads. Also consider that a Zen core is 256/128 L/S and intel is 512/256. its only on those very wide workloads that you have a chance of bottlenecking on bandwidth.

To many people take intel marketing as gospel, thats why every new feature is mega AWEOMSE tm but real world performance only continues on its incremental improvements.

 

[Boze]

Then everyone got obsessed with the quad channel memory.

Because you brought it up. YOU brought up RAM cost as a consideration.

Then I showed you that it $80 to activate your quad channel on an LGA 2011-3 board.

Meaning "expensive RAM" is not the barrier to entry for maximum performance on Intel HEDT. Its not even really a consideration nowadays. The good part about that is that cheap RAM everywhere benefits AMD as much as Intel, if not more.

 

[AtenRa]

Because you brought it up. YOU brought up RAM cost as a consideration.

Then I showed you that it $80 to activate your quad channel on an LGA 2011-3 board.

Meaning "expensive RAM" is not the barrier to entry for maximum performance on Intel HEDT. Its not even really a consideration nowadays. The good part about that is that cheap RAM everywhere benefits AMD as much as Intel, if not more.

You should read my posts again then because you missed the context,

Im talking about the platform (Motherboard + 4x Memory sticks + CPUs + Heat-sinks).

1. Socket 2011-V3 motherboards are more expensive than Mainstream Socket 1151 boards.

2. In order to get the benefit of the quad memory channel you need double the ram capacity and double the price. That means 4x 4GB vs 2x 4GB since 4GB sticks are the smallest of DDR-4.
But even if you want the same 16GB DDR-4 capacity 2x 8GB sticks are $20-30 cheaper than 4x 4GB.

3. Socket 2011-V3 CPUs are 140W TDP and come without heat-sinks. So the vast majority of users will also need to spend an extra $40-50 or more for an after market heat-sink, unless they have one that is socket 2011-V3 compatible.

After that everyone got obsessed with the quad memory channel and forgot all the rest.

 

[CHADBOGA]

You might also want to have a look at your signature.
I remember shortly before P4 launched, Paul Demone called K7 a tractor.
You see people sometimes makes mistakes. The people that say something unlike all your post where you add nothing.

Your post doesn't make any sense.

 

[MrTeal]

I find it highly unlikely they first taped out GPUs on GloFo and not CPUs.

Or are you suggesting next gen AMD 2016 GPUs will be made on GloFos 14nm LPP and not on TSMC ??? Because we already know they will only made two different dies this time.

To that end we know that the small Polaris GPU that RTG previewed will be produced by GlobalFoundries on their 14nm process, meanwhile it remains to be seen how the rest of RTG’s Polaris GPUs will be split between the fabs.

Well, we now have this question answered. At least one of the new 2016 AMD GPUs will be on GloFo's 14LPP process. While we don't know if there might have also been 14LPP CPU test chips spun, the existence of GloFo's statement that they've completed silicon and AMD is doing validation work could definitely have been referring to these new GPUs.

 

[Sweepr]

Nice find MrTeal. So the taped out FinFET product could very well be a GPU.

 

[Phynaz]

http://seekingalpha.com/article/374...y-investors-conference-transcript?part=single

Jeez...

Seems they were right, at least going by what Lisa Su said at Computex.

 

[Phynaz]

The same people that remind us of the WSA that AMD has with GF will of course negate that it s in AMD interest to fab their APU/CPUs at GF like they logicaly did at this point, instead you ll had denial such as :

Lol, they never fabbed any GPU at GF, but hey, no need of past history this time...

Oops...

So far you have provided zero evidence about your (false) claim that Zen taped out months ago. Throwing 'ad hominem' doesn't help.

Right, if they never fabbed GPUs at GF it certainly means they never will. Facepalm.

:thumbsup:

The ad hominem started one page ago, and certainly not from me.

That said it would be suicidal to not get the CPUs/APUs being produced first at GF given their WSA constraints, that s why GF tape outs announcements are likely to be about CPUs/APUs exclusively.


Apparently they will stick with TSMC for GPUs in a first stage, this has the advantage to tame any supply issue that could originate from GF.

Oops #2.

 

[Phynaz]

Nice find MrTeal. So the taped out FinFET product could very well be a GPU.

Yup, turns out it was!

 

[krumme]

As Oops seems to interest you; for real Oops you migh want to start on page 1