AMD - Piledriver chitchat

[DanceswithUnix]

Some Zen chitchat:

http://www.fudzilla.com/news/noteboo...n-has-16-cores

So 16 cores? Oh come on, that can't be right. 16 threads was expected, assuming the usual dull 2 threads per core and the previously predicted 8 cores.

Integrated graphics with HBM memory? Once again, this starts to sound laptop oriented. I have little use for integrated graphics besides using it to order a new graphics card if the main one dies and the last time that happened I was using a 7900GT so a pretty rare event. So what is in the extra cost of HBM memory on board for the likes of me?

4x DDR4 channels sounds nice, but again if it is one dimm per channel then it kind of has to be.

Not sure what to make of that rumour.

[watercooled]

That collection of rumours sounds a bit far-fetched TBH. For starters, what use would 16 cores be for current client use?

In terms of die space, 16 cores plus a rumoured huge GCN block would mean a fairly huge die. Maybe both are true for their own parts i.e. 16 core server part, and a client part with large IGP.

A while back, AMD showed their concept for a HBM-DDR hybrid, so a HBM-equipped APU might only have a small-ish amount of onboard memory, which should help massively with the currently bandwidth-starved graphics performance.

Another possibility, which isn't IMO one of the most likely scenarios, is >2 threads per core SMT. But again, I'm not sure how useful that would be for current client workloads.

I don't doubt that AMD or Intel could release 16 core client CPUs if they felt the need.

=================================================================

In other news, I've just been looking at the pricing of Intel's Atom-core Celeron/Pentium processors, and they're weirdly more expensive than their Haswell/Broadwell counterparts! How does that work?
E.g:
http://www.cpu-world.com/news_2015/2...rocessors.html
http://www.cpu-world.com/news_2015/2..._released.html

The release prices are also listed on the Wikipedia pages. BT-D is $94, but you can get a Haswell Pentium for $64. Even the AE is only $72!!

[scaryjim]

According to the fudzilla story on AMD's recently revealed five year plan:

AMD plans to introduce what it described as a High Performance Computing APU or HPC for short. This APU will carry a sizable TDP between 200 and 300 watts.

So, huge die, massive TDP, lots of execution cores, stacked memory, targeting the HPC market, presumably with a specific focus on HSA enabled applications. That'd make the 16 core rumour sound less ridiculous: yes it's a Zen part, no it's not a client APU.

Also mentioned in that article is the possibility that the ARM-based K12 will support > 2 simultaneous threads per core.

Give me a second, I just need to open a new box of Maldon....

[shaithis]

They need a gaming chip to win back enthusiasts. I really cannot see how they can do that with a 16-core chip. If it is 16 core it says to me it will fall over for the exact same reason their current FX line is snubbed by many gamers.

Obviously, by the time we see it in the wild DX12 may be here and help but there will be lots of non-DX12 games in benchmarks.........hopefully they are working on a successor to the AM3+ platform as well.

[watercooled]

Some more details about the new cores would be nice. Maybe we'll find out in May, when there's some conference IIRC?

AMD are being really quiet and the Internet and 'news' articles are full of random rumours, a lot of which seem to have their roots in forum speculation.

[watercooled]

Chipworks speculate that some of the 14nm Exynos wafers might be from GloFo: http://www.chipworks.com/en/technica...ung-galaxy-s6/

Interesting if true, as it would mean they're fairly close in their 14nm ramp.

[smeders]

Would be great if AMD caught up to Intel on the CPU side of things. Would go for an AMD build in a heartbeat if it matched Intel on single thread!

[smeders]

If AMD can start to match Intel's single core performance especially in gaming (without insane OC like the FX8350) then they would defo start to get back in the market and if they keep prices low go blow for blow compared to Intel.

[DanceswithUnix]

Some more interesting data points on Fudzilla today.

I think this shows why we keep seeing stuff on 28nm...



(from http://www.fudzilla.com/news/process...e-implications)

This also ties in with the comment in http://www.fudzilla.com/news/process...-10nm-schedule that tsmc 16nm has cost per gate problems.

So new processes are well on the way, but you have to be prepared to pay for them.

[watercooled]

That graphic was released ages ago by ARM, they've just stolen it and removed the ARM branding.



AFAIK though they were talking about 2015 prices, so that's to be expected with 16/14 being very new. According to that, 20nm matches 28, which could be because it's had time to ramp since late 2014? It's normal for new nodes to be expensive early on in the ramp.

Intel are still pretty confident that normalised cost/transistor is decreasing:

From http://www.anandtech.com/show/8367/i...logy-in-detail

[DanceswithUnix]

Fudzilla did say where they got the slide from, but their assertion was that the graph was still correct.

Intel have to say that cost per transistor is decreasing, as that is kind of what their business is based on

[watercooled]

You just have to be sure to compare apples to apples, and I think that ARM graphic is used out of context in a few places. I can't find anything to suggest that ARM intended it to mean anything other than cost at a given point in time, vs cost at the same point on the ramp.

[DanceswithUnix]

You just have to be sure to compare apples to apples, and I think that ARM graphic is used out of context in a few places. I can't find anything to suggest that ARM intended it to mean anything other than cost at a given point in time, vs cost at the same point on the ramp.

Modern processes are expensive, and have only been economically viable thanks to things like increasing the wafer diameter at the same time.

It looks like Intel are making their new budget smartphone chips on 28nm: http://www.fudzilla.com/news/process...g-this-quarter

All first-generation SoFIA parts will be manufactured in 28nm, which is actually a step back from the 22nm Bay Trail generation. However, the 28nm node should be good enough for such products, and more importantly, it is significantly cheaper than 22nm, 20nm or 14nm nodes.

Edit to add: Hang on, are those RockChip manufactured parts? That does change things quite a bit.

[shaithis]

Odd seeing intel using Mali graphics rather then their own (heavily pushed) HD series.

[watercooled]

The reason I read for SoFIA on TSMC 28nm is because of the integrated modem.

It's completely normal for bleeding edge nodes to be more expensive to start with, but costs drop dramatically over time. We've been on 300mm wafers for some time for large dies and 450mm is in development. However IIRC a few companies were hoping to introduce it alongside EUV which wasn't ready in time for 14nm.

WRT Mali, Intel have used PowerVR graphics in their current and past mobile (phone) Atom processors, including their newest Merrifield/Moorefield processors.

Ctrl+F for Mali: http://www.anandtech.com/show/9131/i...-devices-group

[watercooled]

Not sure if legit: http://videocardz.com/55358/xfx-rade...les-for-camera