Haswell vs Kaveri

And AFAIK SOI has no inherent density problems and it is contrary supposed to help in this matter.

We may thus only blame only AMD's implementation of the process.
 
AMD states in their slides that main reason for better density is process tailored to APU's where they sacrificing top CPU clock speed at hight TDP for more GPU friendly and denser transistors and metal stack.
 
Plus, GloFo's 28nm SHP process was supposed to have been canceled (no sign of it on their website) in which case Kaveri would be made on 28nm HPP and the "SHP" on AMD's slide would be a typo.
 
Alexko said:
Plus, GloFo's 28nm SHP process was supposed to have been canceled (no sign of it on their website) in which case Kaveri would be made on 28nm HPP and the "SHP" on AMD's slide would be a typo.
Click to expand...
Wasn't reported that AMD payed GF quite a bit of money for the developement of 28SHP and AMD only canceled the development agreement for future custom processes? Starting with the 20nm node AMD will probably use the same processes every GF customer has access to.

In short, 28SHP isn't a typo, it's basically a custom process for AMD.
 
Gipsel said:
Wasn't reported that AMD payed GF quite a bit of money for the developement of 28SHP and AMD only canceled the development agreement for future custom processes? Starting with the 20nm node AMD will probably use the same processes every GF customer has access to.

In short, 28SHP isn't a typo, it's basically a custom process for AMD.
Click to expand...
In that case it could be bulk or SOI, judging by GPU friendly process, I would have to assume bulk because thats what AMD has been using for GPUs since forever.
 
Alexko said:
I would assume that drivers fall back to some kind of basic, poorly optimized support.
Click to expand...
Nah, it either supports it or it doesn't with a new architecture especially something like this with big changes to the memory setup. If it was a different SKU of the same architecture or something fine, but this is a new and unique thing.

That's not to say they are completely "optimized" or anything, but there's no way you'd be running 3D stuff on any "generic" driver.
 
This is the slide about APU optimized process:
NYtZrWl.jpg


Googling around it seems that either 28nm SHP process is just a typo and or exchangable term with 28nm HPP or it really is SOI or FD-SOI process exclusive for AMD products.
According to this article from March 2013:
http://www.pcper.com/reviews/Editorial/GLOBALFOUNDRIES-2013-and-Beyond/Two-Technologies-Watch

FD-SOI was planned for risk production end of '13 and full production Q1'14 but I'm not convinced GloFo and AMG would fail to mention it ..
 
Last edited by a moderator:
My opinion: 28SHP is a PD-SOI process specifically for AMD.

And 28SHP is definitely a separate process from 28HPP:

file1yjk7.jpg


A 28nm FD-SOI process is also offered through ST, but that is again a different one.

st_fdsoi.jpg
 
Last edited by a moderator:
Did anyone just watch the CES presentation by Lisa Su of AMD? There were some specific benchmarks of Kaveri against i5 4670. Was there anything new here or did they gel with anticipated performance?

Also, Beema and Mullins seem to slap down Bay Trail easily
 
fellix said:
Any info on the L2 size? It looks like there is 1MB per module from this shot. :???:
Click to expand...

They are differently oriented from the Bulldozer and Piledriver caches too. (I've also wondered why they mirror flipped the cores in Haswell relative to SB and IVB.) Anyway all sources seem indicate 2mb per module which are now dynamically sized to save power. Seems like steamroller is the first BD derived design with just two levels of cache, and maybe there are die savings from the simpler traces resulting from that decision. L3 design never was as tight as Intel's ring bus design and L2 was relatively huge for each module compared to a Nehalem or Sandy bridge HT core. It was probably better for AMD to stick to just a well module coupled L2 since good shared, modular L3 is hard to do.
 
I am having trouble finding the slides from yesterday. Some of the slides she used yesterday were posted earlier in the day from other AMD events and posted earlier in this thread. My take away was generally positive but I couldn't tell if any Kaveri/Beema/Mullins "news" was made..i.e. the performance is better than leaked slides floating around or as anticipated.

The replay isn't up on http://www.youtube.com/amd yet...I expect it to be updated in a few hours.
 
You must log in or register to reply here.
Back
Top