AMD - Piledriver chitchat

[CAT-THE-FIFTH]

Someone on Flickr has posted some nice die DIY die shots of Polaris 10 (and lost more besides): https://www.flickr.com/photos/130561...57650403404920

Found through TR article: http://techreport.com/news/30496/thi...-our-enjoyment

WTF,he took apart a Sapphire Nitro???

Surely,a reference card would have made more sense??

[watercooled]

Perhaps it's a broken one?

[CAT-THE-FIFTH]

Apparently some Zen AOTS benchmarks:

https://www.reddit.com/r/Amd/comment...ts_benchmarks/

[watercooled]

Do the scores tell you anything? I'm not really sure how the AOTS benchmark works, scales across cores, etc.

[Xlucine]

AM3, that's the thing I meant. Something like this:

[watercooled]

Could be a couple of reasons behind that - it seems to be the AMD 'PRO' roadmap and I don't think Vishera had any SKUs in that sector.

Also, it came out some time before 2015, just it's still available now. Just like how the 130nm AMD Geode is still produced but rarely shown on roadmaps. I wouldn't read too much into it.

[CAT-THE-FIFTH]

We are in August now and there is still no desktop AM4 motherboards or Bristol Ridge??

I mean when is AMD going to actually get around to selling them??

Its not looking good for the APUs,because I doubt AMD would go to the trouble of selling desktop Bristol Ridge only for it to be replaced in under 12 months???

I suspect the desktop Zen APUs will be late next year,and that will mean they will be close to the 10NM Intel chips.

This is the same issue with Bulldozer,Llano and Kaveri - they were delayed so much,Intel launched newer chips by then.

Is it more issues at GF,or have AMD been stretched thin again??

[scaryjim]

AM3, that's the thing I meant. ...

Hmmmm, those Vishera CPUs are the x300 series. I suspect they've been dropped from the roadmap because having 4-year old processors on this year's roadmap doesn't look too clever...

We are in August now and there is still no desktop AM4 motherboards or Bristol Ridge??

... Is it more issues at GF,or have AMD been stretched thin again??

Optimistic view, they're going to have Zen ready in time to do a joint APU and CPU launch either late this year or early next year, and have decided that an interim Excavator APU would just be a distraction.

Pessimistic view, they're having trouble with the memory controller, or chipset, or some other part of the implementation, which is delaying the release of the platform as a whole.

Business view, the mobo manufacturers don't want to invest in putting out new motherboards for a minor incremental enhancement to Kaveri, and since AMD don't manufacture motherboards they can't release Bristol Ridge until their mobo partners get in line.

Segmentation view, the silicon (which AFAIK is basically just Carrizo?) has been pushed into the embedded & server market and is doing so well there that they can't supply sufficient volume to also make a desktop release viable.

Staggered release view, AMD wanted to get the Polaris releases out of he way first so they weren't at risk of one release overshadowing another.

Bad foundry view, GloFo are struggling to produce a chip that can meet AMD's design targets in terms of clock speed, voltage etc.

erm ... those are the main reasons I can think of, pick your favourite two or three

EDIT: oooh, thought of another one:

Business priorities view, AMD have had more orders than expected to their semi-custom business, and as that's a growth priority it's taking resources from the main desktop processor development.

[CAT-THE-FIFTH]

Hmmmm, those Vishera CPUs are the x300 series. I suspect they've been dropped from the roadmap because having 4-year old processors on this year's roadmap doesn't look too clever...



Optimistic view, they're going to have Zen ready in time to do a joint APU and CPU launch either late this year or early next year, and have decided that an interim Excavator APU would just be a distraction.

Pessimistic view, they're having trouble with the memory controller, or chipset, or some other part of the implementation, which is delaying the release of the platform as a whole.

Business view, the mobo manufacturers don't want to invest in putting out new motherboards for a minor incremental enhancement to Kaveri, and since AMD don't manufacture motherboards they can't release Bristol Ridge until their mobo partners get in line.

Segmentation view, the silicon (which AFAIK is basically just Carrizo?) has been pushed into the embedded & server market and is doing so well there that they can't supply sufficient volume to also make a desktop release viable.

Staggered release view, AMD wanted to get the Polaris releases out of he way first so they weren't at risk of one release overshadowing another.

Bad foundry view, GloFo are struggling to produce a chip that can meet AMD's design targets in terms of clock speed, voltage etc.

erm ... those are the main reasons I can think of, pick your favourite two or three

EDIT: oooh, thought of another one:

Business priorities view, AMD have had more orders than expected to their semi-custom business, and as that's a growth priority it's taking resources from the main desktop processor development.

But the CPU side is the major part of the company dragging them down - ATM Kaveri is over two years old and Vishera is getting close to 4 years old now. Kaveri was delayed too. Intel now has better media decode and encode abilities on their IGPs too and will further improve on that with Kaby Lake. Bristol Ridge is using a mature 28NM process and is a SOC chip too,so the delay is weird,unless they are really using FIVR on Bristol Ridge and hit a problem with it.

Personally I am getting a bad vibe on it - when AMD delays stuff it generally comes out a bit half baked. Phenom,Bulldozer,Llano and Kaveri were half baked and they were all delayed. The CPUs they got out on time like Vishera,Trinity and the Phenom II ended up being reasonably OK.

Its also a bit worrying that Zen looks to be delayed a quarter if the roadmap is true - AMD was on record saying late 2016 for release. That would make Vishera over 4 years old.

[CAT-THE-FIFTH]

Looks like AMD GPU marketshare has increased again:

http://www.hardocp.com/news/2016/08/...6#.V7f6srGxWHs

34% overall and around 30% for desktops.

[CAT-THE-FIFTH]

Zen coming in February 2017:

http://wccftech.com/amd-zen-intel-ka...february-2017/

[Noxvayl]

I'm looking to Raven Ridge as a possible upgrade from my FX CPU at the moment... Would be interesting to find out if it will crossfire with my RX480 or not.

[badass]

I know no one's particularly excited by the Bristol Ridge platform but it looks like it will be a little faster than Godaveri and of colurse being FM4 it's more future proof.

Anandtech have an excellent article comparing a desktop Bristol Ridge to older architectures and clocks them all at 3 GHz to compare IPC. http://www.anandtech.com/show/10436/...-athlon-x4-845

It looks like Carizzo was hobbled by the reduced L2 cache. In some benchmarks it's a lot faster and in some it's a bit slower.

Hopefully the DDR4 interface will help although with latency being more or less the same as DDR3 the effects on CPU performance will probably be slight.

[watercooled]

In terms of competition on the desktop, it looks like we might not see much from Intel until maybe 2019?

For 2017 we have Kaby Lake, which seems like basically a small clock bump as far as desktop performance is concerned.

For 2018 there are various rumours around Cannonlake and Coffee Lake, but from the leaked roadmaps it looks like 10nm Cannonlake (which I'm assuming given the Process-Arch-Optimisation cadence will be basically a die-shrunk Skylake on the CPU side) will at least start more on the low-power side, with the higher-power chips covered by the 14nm Coffee Lake.

Assuming it's real, I wonder what Coffee Lake is though, and how it fits into the P-A-O model? Rumours suggest it could have a 6 core variant though, so maybe that's it, and they're avoiding larger dies on 10nm to start with. But by 10nm their core sizes must be getting pretty tiny anyway...

Edit: I've just read another suggestion that Coffee Lake is more like a 14nm version of Cannonlake for cost reason, but... isn't Cannonlake a die-shrink of Skylake with maybe some minor tweaks itself? Maybe Intel are planning to combine the Tick and Tock for Cannonlake given its delay i.e. the architectural work on the originally planned post-Cannonlake might be far enough ahead to bring it forward to Cannonlake? But that would be quite a departure for Intel, risking a new arch on early 10nm, and where does that leave the P-A-O model? PA-O-O? PO-A-O?

[badass]

In terms of competition on the desktop, it looks like we might not see much from Intel until maybe 2019?

For 2017 we have Kaby Lake, which seems like basically a small clock bump as far as desktop performance is concerned.

For 2018 there are various rumours around Cannonlake and Coffee Lake, but from the leaked roadmaps it looks like 10nm Cannonlake (which I'm assuming given the Process-Arch-Optimisation cadence will be basically a die-shrunk Skylake on the CPU side) will at least start more on the low-power side, with the higher-power chips covered by the 14nm Coffee Lake.

Assuming it's real, I wonder what Coffee Lake is though, and how it fits into the P-A-O model? Rumours suggest it could have a 6 core variant though, so maybe that's it, and they're avoiding larger dies on 10nm to start with. But by 10nm their core sizes must be getting pretty tiny anyway...

Edit: I've just read another suggestion that Coffee Lake is more like a 14nm version of Cannonlake for cost reason, but... isn't Cannonlake a die-shrink of Skylake with maybe some minor tweaks itself? Maybe Intel are planning to combine the Tick and Tock for Cannonlake given its delay i.e. the architectural work on the originally planned post-Cannonlake might be far enough ahead to bring it forward to Cannonlake? But that would be quite a departure for Intel, risking a new arch on early 10nm, and where does that leave the P-A-O model? PA-O-O? PO-A-O?

IIRC Coffee Lake includes 6 core CPU's on the consumer socket on the 14nm process.
I would speculate that:
The new 14nm+ node needs to get a few years of use to justify the investment
The 10nm process is anticipated to be significantly more expensive per transistor than 14nm and Intel anticipate a slow and poor yield ramp.

I expect Coffee Lake to be the same as Kaby lake but more like additional SKU's rather than changing the process or architecture.
The initial Cannon Lake will be the smallest die mobile chips that will benefit the most from the smaller, lower power process. Their smaller die size will help offset the high initial defect rate.

Once the 10nm process is more mature, I expect Intel to release the desktop and server Cannonlake CPU's.

So I think Coffee Lake will just be the same O as Kaby Lake in the P-A-O cycle.

[DanceswithUnix]

The new 14nm+ node needs to get a few years of use to justify the investment
The 10nm process is anticipated to be significantly more expensive per transistor than 14nm and Intel anticipate a slow and poor yield ramp.

Intel are famous for moving on to newer processes and products early on the grounds it is better for them to disrupt their own products than have a competitor do it for them. I think right now things are looking more fragile for Intel than ever, their process lead over other foundries isn't buying them as much as it used to, packaging tech like HBM and 3D memory stacking is becoming more important. If they *could* roll out a more advanced process, then I'm sure they would be all over it to try and keep their lead.

Allowing for the usual "the numbers aren't comparable", reports are that Intel are being caught up (http://www.fudzilla.com/news/process...ping-over-10nm) and even things like their new flash replacement isn't working out as it should (http://semiaccurate.com/2016/09/12/i...y-much-broken/).

Intel haven't dropped the ball yet, and history of the P4 shows they can drop the ball for quite an extended period of time before it starts to really matter, but things aren't looking quite as rosy as they used to for them.