Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
AMD Carrizo and Carrizo-L to share the same platform
http://www.tomshardware.com/news/amd-carrizo-carrizo-l-notebook-apu,28345.html
A lot of that is just marketing junk that doesn't interest me. But this does:
"Our scalable approach, which hasn't been used in the PC industry for many years now, addresses the complete stack from top to bottom with a single platform, meaning that our OEM partners will be able to use "Carrizo" and "Carrizo-L" to serve a wide range of the market – spanning form factors and price points from all-in-ones and mainstream notebooks, to ultra-thin and light performance and gaming notebooks – without needing to design, develop, and validate multiple platforms."
As a quick reminder, Carrizo and Carrizo-L are the successors to Kaveri and Beema, respectively: AMD's relatively higher and lower performance x86 CPU architectures.
As you presumably know, if you buy an LGA 1150 motherboard, there are a lot of different Haswell processors you can put into it. But you can't plug in an Atom processor. Likewise, if you buy a Socket AM3+ motherboard, there are a lot of different Vishera processors that you can plug into it. But you can't put in a Kaveri or Beema processor. Different platforms take different sockets, and hence different motherboards.
What AMD is doing is to say, all of the next generation AMD architectures--both the high end and low end--will use the same platform. This will allow a laptop vendor to design a laptop once, and then offer both a cheap Carrizo-L laptop based on it, and also a more expensive, higher performance Carrizo laptop based on the same thing. Hopefully this will lead to better uptake of Carrizo in laptops than Kaveri has seen.
I previously speculated that the reason why Carrizo-L was the successor to Beema, while Mullins had no successor this year in spite of Beema and Mullins being the same chip, was so that Carrizo-L could have two memory channels rather than one. Intel offers a dual-channel version of Bay Trail Atom, and Beema could certainly use the extra bandwidth more than any Atom chip ever made. So while this doesn't confirm my speculation, it does at least point in the right direction. It would be very strange and likely unprecedented to have Carrizo and Carrizo-L share a platform without even using the same number of memory channels.
AMD is also promising a next generation graphics architecture, and presumably the same architecture that is used in what will probably be the Radeon R* 300 series. Rumors have said that AMD was going to produce those GPU chips on Global Foundries' 28 nm SHP process--which was originally designed for Kaveri, to allow high clock speeds. Designing the architecture and implementing it on only one process node, both for their high end and low end integrated graphics and also for discrete cards, certainly simplifies things somewhat for AMD. And it also helps AMD fill their requirement to buy a bunch of wafers from Global Foundries (formerly AMD's in-house fabs), since Global Foundries continues to develop process nodes specifically for AMD as part of the terms of the sale.
Also from the linked article:
"we're bringing a brand new graphics architecture integrated into Carrizo. This will be our biggest leap ever from an energy efficiency perspective."
We'll see about that, of course. Nvidia just got quite an energy efficiency jump going from Kepler to Maxwell. And Kepler wasn't a bad architecture to begin with, either, so AMD might well need their next generation graphics to be their biggest energy efficiency leap ever just to stay competitive.
Comments
I guess this really comes as no surprise, as AMD pretty much indicated that Vishera/FX-series was the end of the road for a dedicated desktop chip, and they were going to shift everything over to APUs for consumer computing (combined desktop/laptop). I don't really blame them for that.
It will be interesting to see how these Excavator cores stack up (and scale up).
AMD's server lines have been leaning toward "ambidextrous" cores - CPUs with both x86 and ARM cores, and those were leaning towards the Jaguar/Puma line, not the Bulldozer/Excavator higher performance line. But then again, in servers, more cores tends to trump faster cores, and energy use is hugely important in high density applications, so that makes a bit of sense -- most server applications aren't clamoring for a faster general purpose core, they are just throwing more cores at the problem.
So I guess it will be interesting to see how Carrizo stacks up against Vishera then, getting back to something gaming-oriented.
I also found this read:
http://www.extremetech.com/computing/178752-amds-next-gen-carrizo-apu-features-leaked-shows-greater-focus-on-power-efficiency
It's a bit dated, so some of it may be obsolete by now, but there are some interesting parts to this.
This article also indicates that FX is end of the line for dedicated desktop chips. Not a huge surprise there.
More interesting, the PCI configuration - Carrizo is supposedly dropping down to just 16 lanes for the on-die controller: enough for a single PCI 3.0 x8, or 2 slots at x4. It's not even a single x16, which most enthusiasts/gamers look for, but I doubt that will severely cripple a discrete GPU. But it does basically eliminate Crossfire support on discrete GPUs, or force motherboard vendors who want to have multiple PCI interfaces to implement it themselves in their own external chip (which has been done in the past, but adds significant expense to the motherboard).
But then again, as the author of the article explains: most people buying an APU are not really in the Performance Crossfire market anyway, at least past using it to aid with the on-die APU graphics.
That they site Bright Side of News as their source is reason for a healthy skepticism of the claims.
What Carrizo could plausibly do is what Intel has essentially done since Sandy Bridge: have a PCI Express x16 controller on die that communicates with the chipset, then the chipset splits that into more PCI Express lanes. The idea is that if you have PCI Express connections to a video card, some extra USB and SATA ports, and various other things, they probably don't add up to more bandwidth than would overload the bandwidth from the CPU to the chipset, so it works fine.
And there's also the possibility of multiple dies, with an SoC version for laptops and a slightly different die for a socketed version in a desktop. Intel has a bunch of different Haswell dies, for example.
http://techreport.com/news/27652/amd-carrizo-will-face-off-with-broadwell-u-wont-land-in-desktops
AMD has apparently now confirmed that there isn't going to be a socketed desktop version of Carrizo. That doesn't mean that it absolutely can't go in a desktop at all, but it does mean that the ability to add a discrete video card would be very restricted, and you also wouldn't get very many USB or SATA ports.
My take on this is that it means that AMD knew that Carrizo wouldn't be a very good desktop chip, anyway. If it's not going to be any better than Kaveri, why bother making it? Intel is basically facing the same situation on Broadwell in a desktop. Making it an SoC really is better for laptops, and if you can bring down power consumption at 3 GHz at the cost of giving up the ability to clock as high as 4 GHz, that's worth it in a laptop.
Hopefully AMD will have something good for desktops with their Zen architecture due next year. That won't be yet another Bulldozer derivative, but it will be a chance to fix everything that ails AMD's CPU architectures. I'd expect massive improvements, if only because it's much easier to get big improvements over your last architecture when your last architecture was bad.