New Zen microarchitecture details

[Fanatical Meat]

Enough of the yields estimates. When will we see more benchmarks dammit

 

[jpiniero]

I did the picture comparison of Skylake and Broadwell-E and this is what I came up with:

Skylake 4+2+8MB L3 (123 mm2)
Cores: 42.58
L3: 6.3
DC Memory Controller: 6.3
System Agent: 20.66
IGP: 42
Empty Space: 2.14

Broadwell-E: 10+0+25 MB L3 (246 mm2)
Cores: 61.11
L3: 44.18
QC Memory Controller & "Stuff": 44.14
Uncore: 63.23
Empty Space: 25.80

May not be totally accurate... but maybe this is a good comparison? But in any case, 18-20 mm2 for the L3 on Summit Ridge would not be bad by any means.

 

[krumme]

With infinity fabric i really dont understand what we need apu for?

Arm73/xx and the likes can take low end and emulate windows. X86 is for high perf.
Right now amd can make a 25w tdp 4c/8t laptop no probs as there is 180 32 server cpu on the way. Package it with vega in different sizes and hbm and you have a blast of a laptop solution without shelling out billions for developing different apu configurations. Its console budget and takes tens of millions pcx sold to pay for those insane cost.
Then Intel can take the lower midrange. If a 4c is 95mm2 its small and darn cheap to produce and with a small vega can enter a lot of machines in the 50w tdp ballpark for the lowest and slowest part and easily scale to monster 300w solutions for 8c big vega solutions.

Le-go.

 

[jpiniero]

With infinity fabric i really dont understand what we need apu for?

The APU die would be smaller than separate dies, plus the interconnect costs. And it's going to be at least two years before HBM gets cheap enough to do what you are suggesting.

 

[Glo.]

With infinity fabric i really dont understand what we need apu for?

Arm73/xx and the likes can take low end and emulate windows. X86 is for high perf.
Right now amd can make a 25w tdp 4c/8t laptop no probs as there is 180 32 server cpu on the way. Package it with vega in different sizes and hbm and you have a blast of a laptop solution without shelling out billions for developing different apu configurations. Its console budget and takes tens of millions pcx sold to pay for those insane cost.
Then Intel can take the lower midrange. If a 4c is 95mm2 its small and darn cheap to produce and with a small vega can enter a lot of machines in the 50w tdp ballpark for the lowest and slowest part and easily scale to monster 300w solutions for 8c big vega solutions.

Le-go.

Have you considered that they can be seperate? CPU and GPU seperate dies, with 2 stacks of HBM2 on interposer, and connected with Infinity Fabric.

 

[cytg111]

Have you considered that they can be seperate? CPU and GPU seperate dies, with 2 stacks of HBM2 on interposer, and connected with Infinity Fabric.

APU, HSA, GCN. I am guessing latency would become an issue with sep. dies?

 

[Glo.]

APU, HSA, GCN. I am guessing latency would become an issue with sep. dies?

WHOLE package would be connected through Infinity Fabric and Interposer.

 

[Doom2pro]

Yields will be better with a MCM type configuration on interposer with IF connecting separate GPUs and CPUs and possibly HBM2... There is even word that organic packaging is getting good enough to possibly use HBM without an interposer in the future.

 

[jpiniero]

Yields will be better with a MCM type configuration on interposer with IF connecting separate GPUs and CPUs and possibly HBM2... There is even word that organic packaging is getting good enough to possibly use HBM without an interposer in the future.

It's definitely an interesting concept; but like I said it's something for the future and not the present or near future. Keeping it ondie saves power too.

 

[CatMerc]

An error of a few pixels might quickly add up by repeating it. I'd say, there is at least a 5% error in measurement. It was actually just done in a discussion related to adding L3.

Yup, it's important to remember these are estimates. There's margin of error of I'd say 10% in either direction.

 

[CatMerc]

http://www.bitsandchips.it/english/54-english-review/7926-being-lisa-su?showall=1

Rory Read, as CEO, has taken up Sanders’ idea, creating a team of senior "pissed-off" engineers (Zen/K12 team), and calling back some never forget AMDers: Raja “The Oracle” Kojuri, Terry “Catalystmaker” Makedon, and so on. Do you remember Mr. Wolf in Pulp Fiction? "I'm Rory Read. I solve problems".

The mental image of a bunch of angry engineers working out of pure spite for Intel is incredibly entertaining.

 

[krumme]

Have you considered that they can be seperate? CPU and GPU seperate dies, with 2 stacks of HBM2 on interposer, and connected with Infinity Fabric.

Well i thought i wrote that ?

 

[krumme]

APU, HSA, GCN. I am guessing latency would become an issue with sep. dies?

Yep but where is this hsa going in the real world?

 

[CHADBOGA]

http://www.bitsandchips.it/english/54-english-review/7926-being-lisa-su?showall=1



The mental image of a bunch of angry engineers working out of pure spite for Intel is incredibly entertaining.

He is clearly a very big fan of Jack Trout.

 

[krumme]

It's definitely an interesting concept; but like I said it's something for the future and not the present or near future. Keeping it ondie saves power too.

Yes. But imo this needs to be seen from a cash, risk and differentiate perspective also.
What is the startup cost for a new apu die even if you have all the ip in more or less synthesized form? 1.2B? 0.2B?

 

[Glo.]

Yep but where is this hsa going in the real world?

The development of it has stalled because there was no viable hardware. Nobody was interested apart from abusing it.

Right now, AMD may have finally brought to us extremely viable hardware. Everything lies in the hands of the software developers right now.

If software will follow, HSA is the next big thing in computing.

 

[krumme]

The development of it has stalled because there was no viable hardware. Nobody was interested apart from abusing it.

Right now, AMD may have finally brought to us extremely viable hardware. Everything lies in the hands of the software developers right now.

If software will follow, HSA is the next big thing in computing.

What about the status on arm and android?

 

[Glo.]

What about the status on arm and android?

One simple word, to which everything comes to end: software.


Do you have any? Is it meaningful?

 

[krumme]

No i guess not except perhaps gaming and vr here?
Where is the benefit on desktop? If i need to use gpu for rendering i tell the program that?

 

[Toettoetdaan]

Yep but where is this hsa going in the real world?

Well, where is AVX2 and AVX512 going?
If someone is willing to optimize for it and find some usefull usecases, it gains traction.

 

[krumme]

Well, where is AVX2 and AVX512 going?
If someone is willing to optimize for it and find some usefull usecases, it gains traction.

Well its going nowhere?
Intel have this brilliant segmentation--greedy practice for tech that they kill their own stuff themselves but it also goes to show how difficult it is with new extensions even when you are the dominant player.
One can also argue the value of avx2 and avx512. Like using a family car as tractor. So its an uphill battle against Lee.

 

[Phynaz]

One simple word, to which everything comes to end: Money.


Do you have any? Is it meaningful?

FTFY

 

[frozentundra123456]

Yep but where is this hsa going in the real world?

"Fusion is the future"---AMD 2008. Still is the future apparently.

 

[Sven_eng]

"Fusion is the future"---AMD 2008. Still is the future apparently.

Except that 99% of Intel CPUs are "fusion" of CPU and GPU and have been since 2011?

 

[Arachnotronic]

Except that 99% of Intel CPUs are "fusion" of CPU and GPU and have been since 2011?

99% is too high an estimate. The bulk of Intel's client CPUs have GPUs on board, sure, but I doubt many of their server CPUs have GPUs on them