AMD - Piledriver chitchat

[CAT-THE-FIFTH]

Supposedly,Zen is not fully taped out yet:

http://seekingalpha.com/article/3741...pt?part=single

But then it contradicts this and what their CEO said too:

http://globalfoundries.com/newsroom/...n-amd-products

AMD really needs to convey a singular message when it talks about new products and so does its partners. It really comes across as very disjointed.

Edit!!

The first link says for servers - hopefully it means consumer chips launch earlier.

[watercooled]

I wouldn't read too much into it at the moment, the wording seems very unclear and could be interpreted a couple of ways. He could be on about talking about the tape-out over the next few months. Because considering they're still claiming 2016 as the release year, months into 2016 seems very unlikely as a tape-out time-frame.

At one point he also says:

We haven’t given the specifics but you know they taped out in the last few months...

I've not figured out what 'they' is though.

Edit: Again it's possibly down to a misunderstanding at some point but he seems to contradict himself a couple of times with regards to time-frame, e.g. he talks about revenue ramp in 2016 - that statement could not be true for a product releasing in 2017...

[watercooled]

It looks like AMD might be soldering the die to the IHS again with Godavari. http://semiaccurate.com/forums/showthread.php?t=8709

I wonder why that is?

[Kumagoro]

Well I assume they are trying to make them as good as they can with what little they have left. I wonder how much per chip it costs to do?

Did they originally just follow Intel's lead by switching to gloop.

[watercooled]

At least it seems to show that Godavari is actually a different part to Kaveri.

WRT that conference, I listened to it myself an I'm honestly none the wiser - it's really not clear to me whether he's talking about the announcement of the tape-out, or the tape-out itself at about the 25m mark (and of course there are multiple tape-outs for a product so he could just as easily be talking about predicted final silicon tape-out, which time-wise would still work). He then goes on to talk about Zen in 2016 at about 29m30s.

Also in the podcast, from the scribbled notes I made last night and forgot to post about, a couple of other interesting points were raised. At around 10m20s he talks about their relationship with both TSMC and GloFo for the console semi-custom business. I mean I've heard rumours about GloFo console chips but considering the costs involved in multi-sourcing and the margins of the business, it still didn't seem all that likely. But I still wouldn't take it as confirmation either way as, again, the whole discussion seemed very confused.

One other thing is I think (there's a pattern here!) he said the new H2-2016 semi-custom parts are both non-ARM, so presumably x86? That was around the 15m30s mark.

But I agree with CAT - I really don't think it helps AMD that they tend to just leave contradictory information floating about. And lets not forget the big GloFo announcement about AMD FinFET tape-outs at their foundry. Now of course it could be GPUs, but TSMC still seems more likely there. And IIRC when asked about whether they'd be using GloFo more, he responded saying they're continuing to use both foundries going forward, so unless they've completely flipped around and are planning for CPUs at TSMC, what does that leave for TSMC? But then again, AMD are still being very secretive about which foundries are used for which products, so it's anyone's guess!

[watercooled]

I wonder if AMD has plans continue their cat-core processors? It's something that hasn't received much attention lately since the rumoured 20nm Nolan was scrapped along with 20nm parts in general. A cat core would be really tiny on FinFET but I wonder if developing yet another core in parallel is justified considering AMD has no interest in phones and in theory a core like Zen could scale right down to tablet form factors on 14nm?

[DanceswithUnix]

I wonder if AMD has plans continue their cat-core processors? It's something that hasn't received much attention lately since the rumoured 20nm Nolan was scrapped along with 20nm parts in general. A cat core would be really tiny on FinFET but I wonder if developing yet another core in parallel is justified considering AMD has no interest in phones and in theory a core like Zen could scale right down to tablet form factors on 14nm?

That is an interesting one. Existing jaguar apus would be ideal for things like Chromebooks, and high end tablets, but I for one have yet to see one in anything that I would be tempted to buy.

The latest range, where one pcb layout can take Carizzo or Carizzo-L chips was a good move if that allows one pcb/chassis to take a range of chips and cut down engineering cost. Still doesn't seem to have got the shops overflowing with kit using the parts though.

I guess the next xbox and ps5 consoles would want to use Zen.

[CAT-THE-FIFTH]

That is an interesting one. Existing jaguar apus would be ideal for things like Chromebooks, and high end tablets, but I for one have yet to see one in anything that I would be tempted to buy.

The latest range, where one pcb layout can take Carizzo or Carizzo-L chips was a good move if that allows one pcb/chassis to take a range of chips and cut down engineering cost. Still doesn't seem to have got the shops overflowing with kit using the parts though.

I guess the next xbox and ps5 consoles would want to use Zen.

Nintendo NX or that new handheld Steam console with the unannounced,under NDA, AMD chip??

[DanceswithUnix]

Nintendo NX or that new handheld Steam console with the unannounced,under NDA, AMD chip??

I thought the steam console sounded like a fairly standard part.

The NX though, interesting...

[watercooled]

Yeah there are a couple of AMD products which seem quite promising and are definitely hard launches but are extremely hard to find in products, let alone decent ones. I.e. it's not like they're just talking it up and MFRs refused it because of performance etc as we have hard numbers on the performance, like Beema/Mullins.

I wonder if it's down to some marketing reason i.e. they couldn't secure many design wins, or if capacity was severely limited for some reason? E.g the A4 micro 6400T, I saw a 14" Streambook with it in Tesco and was actually close to getting one for train rides etc, but they very quickly disappeared from shelves despite the smaller Atom versions being readily available long after?

Edit: I've noticed PC-Engines are currently beta-testing a new version of their APU board with an embedded Kabini SoC: http://www.pcengines.ch/apu2b4.htm

Meanwhile, Soekris have abandoned development of the net6801 with the Avaton/Rangely SoCs for cost reasons...

[CAT-THE-FIFTH]

I thought the steam console sounded like a fairly standard part.

The NX though, interesting...

No,its a handheld one and AFAIK it seems the actual production one will have a part the company cannot mention as it is under NDA!

[watercooled]

I've just looked and remembered how poor AMD's naming system is for their mobile APUs.

A8-7410
A8-7100

Have a wild guess which is the higher-performance Kaveri APU, and which one is the ultramobile Carrizo-L.

If you applied logic, you're wrong. xD

Not that Intel is much better lately after having the Atom processors adopt the same Celeron/Pentium branding as the big cores. :sigh:

[watercooled]

Question for the day - I wonder how AMD will brand their implementation of SMT? HyperThreading is an Intel trademark after all. I guess they could just stick with SMT and not correct people when they refer to it as HyperThreading as that seems to have become almost a generic trademark in the client market at least.

[DanceswithUnix]

It isn't an Intel invention, the Terra MTA was I think the first to commercially implement threading (with 1024 threads per core) followed by IBM with one of their Power CPUs and DEC had threading on an Alpha which sadly never saw production.

Given Intel were late to the game with initially a pretty lousy implementation, I always though the grand name of "Hyperthreading" was a bit of a joke.

IBM call it SMT. MIPS call it MIPS-MT.

I suspect they will just brush it under the carpet with a simple bullet point like "SMT". I wouldn't expect to buy a car with "twin cam" written on it or "Injection" as it is old tech that everyone is doing. In reality, no-one cares about SMT what they want is the brute force performance you can sometimes get on single threaded workloads with a really wide issue CPU and any threaded performance is just a bonus and gets you back up to what Piledriver can already do on video encoding benchmarks.

What I would really like to see is a BIOS option to turn the threads up to 4 per core. Go-on AMD, some server & virtualisation loads would love it.

[watercooled]

Yeah I just mean the term 'HyperThreading' itself is an Intel thing, and that's what a lot of websites/consumers use to refer to SMT in general. Those who know of earlier implementations know otherwise of course.

WRT encoding benchmarks, I've found SMT actually makes very little difference with x264, while the Bulldozer architectures with their actual integer cores per thread scaled really well. In terms of performance alone, Piledriver still performs admirably against Skylake.

[CAT-THE-FIFTH]

Hmm,somebody on OcUK forums noticed this:

https://twitter.com/blueisviolet/sta...19647372296193

Zen APU in 2016?? I wonder if AMD getting onboard with Samsung might be the reason??