16-core Steamroller may be coming to servers after all

[JDG1980]

AMD apparently posted developer documentation for an upcoming "Family 15h Models 30-3Fh" server processor. 8M/16C on a single die:
http://www.techpowerup.com/196952/amd-readies-16-core-processors-with-full-uncore.html

"Family 15h Models 30-3Fh" corresponds to Steamroller in AMD's other documentation. The fact that they are designing a new die, and trying to fit in 8 modules (which wasn't done on 32nm - they had two dice on-package connected with MCM) seems to hint that this will indeed be a 28nm part using Steamroller cores. Server parts already usually run at lower clock speeds than client parts, so the clock speed deficit of 28nm may not be as big a deal on the server side.

 

[inf64]

There is no such thing on any public roadmap so I don't think we will see such a part. Maybe in 2016 we get a new socket with a new core (post Excavator) that will power 8C+ parts. Until then it's Warsaw(PD based).

 

[Phynaz]

Reading the design, looks to be (2) four module CPU's in an MCM package using HT to link them up. That's why it has three HT links instead of four.

 

[ShintaiDK]

The article is based on another one with a roadmap from early 2011.

Its not gonna happen.

 

[inf64]

They are reading this from the latest SOG... It was released just a few days ago when Kaveri launched.
Having said that there is no such part on any roadmap so I doubt this will become a product, ever.

 

[Jimzz]

Yea I bet they produced, or at least designed, some but saw the gains were not worth the cost. Because of this no AM3+ FX cpus as well.

Remember a new CPU has to go through a lot of testing before many shops will allow them to run in their server. The time to design, test, build, and support may not be worth it if the sales are not there. So keep selling ones that are already validated and build up to the next design.

 

[FlanK3r]

its document from January 2014....

 

[ZGR]

I wouldn't be surprised if its really 16 Compute Units where its 2 modules for 8 threads total + 8 CU's from the GPU or something like that.

 

[JDG1980]

I think it's a mistake to take roadmaps too seriously. They are marketing documents first and foremost. If something is put on a roadmap and then delayed or not released, everyone will be disappointed, and the company might even lose sales from people who decide not to buy the existing product because they're waiting for its successor. If something is released without being put on a roadmap, then it's an added bonus; users are happy and competitors might be taken by surprise.

Remember that early roadmaps indicated that Intel's Broadwell would be BGA-only, until this was walked back. So these things are not set in stone.

Steamroller actually makes more sense on the server than on the desktop. Kaveri was rather disappointing because the clock-speed deficit cancelled out the IPC gains, and floating-point performance did not improve, so some applications saw a regression. But server parts are usually clocked considerably lower to begin with - AMD's current 8M/16C Opterons top out at 2.8 GHz, for instance. And server workloads tend to be integer-focused and heavily multithreaded. Steamroller improves integer IPC by a bit more than 10%, and drastically reduces the CMT penalty for multithreading. Combined with the power saving of 28nm, this could make 8M/16C Steamroller much more attractive on the server than its Piledriver counterpart. Of course it still won't match Intel on the high end, but margins should still be better than with desktop parts.

 

[Exophase]

There were other configurations mentioned in the software optimization guide that never happened. I'm pretty sure I remember the old cancelled 5-module layout in there. I don't think having some capability in hardware for scaling really means anything. It's pretty much a given that AMD didn't start Steamroller with no intention of making big chips around it, but times change.

 

[Erenhardt]

I wouldn't be surprised if its really 16 Compute Units where its 2 modules for 8 threads total + 8 CU's from the GPU or something like that.

That would be around 150% die size of kaveri (50/50 4xCPU/8xIGP)

 

[toshevopc]

https://www.change.org/petitions/ad...-processor-with-steamroller-microarchitecture check this out maybe we can change something