AMD - Zen chitchat

[scaryjim]

What's alienware doing with a very conservatively clocked 12 core part?

Given it's an engineering sample, I'd say .... testing it?

We don't actually know what's happening with Snowy Owl yet, but since a 12 core CPU must be at least a 2 die MCM, you need to consider that clocking it higher will mean an increased TDP: e.g. the 1600 with 65W TDP is 3.2GHz - 3.6GHz. From that we can surmise that a 12 core processor at the same clocks would require a 130W TDP - quite sizeable.

At a rough guess I'd say Dell's 12 core ES probably targets a 95W TDP. Interestingly, 95W is the top TDP for AM4. Snowy Owl on AM4 after all? I wonder if there are NC pins on AM4 that can be used to plumb out more PCIe lanes and/or memory channels...? Could X399 use the same socket in a different configuration, and if so how much cross compatibility would there be*?

Oh, so many questions...!

* EDIT to add:

It occurs that the Snowy Owl SCM and MCM packages mentioned in slides share a pin-compatible BGA interface. What if AM4 is actually designed for 2 die MCMs and the Ryzen chips released so far only use a limited subset of its pins...?

[kompukare]

I wonder if there are NC pins on AM4 that can be used to plumb out more PCIe lanes and/or memory channels...? Could X399 use the same socket in a different configuration, and if so how much cross compatibility would there be*?

Memory channels is highly unlikely as the whole point of AM4 was to be a cheap unifying platform for APUs and CPUs without the cost (both in money and power) of quad channel. However, whether the CCXs would allow AMD to easily package a chip with four CCXs is another question. They'd still have to physically fit in AM4 too but a quick google says AM4 is 1600mm² (40mm x 40mm) and Ryzen 7 is ~200mm², so that shouldn't be a problem. But whether any current AM4 boards would be able to handle twice the TPU is another question.

Anyway, that Italian site has an article saying Naples and Threadripper will share almost the same socket:
http://www.bitsandchips.it/52-englis...st-same-socket

The 8 and 4 channel memory thing is strange though.

[scaryjim]

... whether any current AM4 boards would be able to handle twice the TPU is another question ...

I'm pretty sure AM4 is designed for a top TDP of 95W - which is why I was speculating that a 2.7GHz 12-core might fit into a 95W TDP... If the point is just for lots of threads then higher clocks would be less important. Interesting thought, isn't it

... that Italian site has an article saying Naples and Threadripper will share almost the same socket: ...

It's an odd one, isn't it. SP3 has more than 4000 contacts; that's got to be a pretty big. Pin compatibility isn't that unusual between different silicon, but different numbers of memory channels are. Very confusing...

[watercooled]

AMD webcast on now. Some rumours they might announce or at least give some more details on Vega. http://webcastevents.com/events/amd/...ive/player.htm

[CAT-THE-FIFTH]

AMD Threadripper and EYPC:

http://wccftech.com/amd-unveils-16-c...h-end-desktop/
http://wccftech.com/amd-unveils-epyc...medium=related

Vega unveiled:

https://videocardz.com/newz/amd-unve...ed-radeon-vega
https://videocardz.com/69585/amd-ann...ontier-edition

If they are still using 4096 shaders it means the clockspeed must be close to 1600MHZ as previously rumoured.

CPU and GPU roadmaps are revealed:

https://videocardz.com/69571/amd-rev...d-zen-roadmaps

So AMD will move to an "improved" 14NM node before moving to 7NM with Zen 2 and Navi.

[watercooled]

If I understand it correctly, EYPC is the brand name replacing Opteron? Shame as I quite like the Opteron branding.

It's nice to see they've (apparently) improved the clocks on Vega vs Polaris and the engineering samples.

It seems like Intel have started something with the xxnm+ naming though! But it probably is more than just a marketing thing - 28nm had its fair share of implementations at both GloFo and TSMC so it's not unreasonable to expect the same for 14nm. It's just, like Intel, they were all just called xxnm and never bothered with the plus before. With increasing time between nodes, it's nice to have something new to market I guess...

[CAT-THE-FIFTH]

If I understand it correctly, EYPC is the brand name replacing Opteron? Shame as I quite like the Opteron branding.

It's nice to see they've (apparently) improved the clocks on Vega vs Polaris and the engineering samples.

It seems like Intel have started something with the xxnm+ naming though! But it probably is more than just a marketing thing - 28nm had its fair share of implementations at both GloFo and TSMC so it's not unreasonable to expect the same for 14nm. It's just, like Intel, they were all just called xxnm and never bothered with the plus before. With increasing time between nodes, it's nice to have something new to market I guess...

Someone spotted this picture of the MCM:

http://i.imgur.com/Z8geiuW.jpg


Edit!!

http://i.imgur.com/HsQM8bb.png


https://image.prntscr.com/image/9bab...0659c8259f.png



Second Edit!!

Ryzen mobile:

http://techreport.com/news/31918/ryz...aptop-near-you

[kompukare]

Don't think these two slides about IF have been posted before:


Will be interesting to see both for the APUs and Epyc (RIP Opteron) if the multi-die CCX approach has any major issues. If it doesn't, this could give a cash-strapped company like AMD great flexibility.

[CAT-THE-FIFTH]

For the APUs it won't be a limitation due to the fact they are probably only a single CCX with the second one replaced by an IGP.

[watercooled]

Off-topic a bit (but when has that ever stopped is? ) - they've confirmed a version of VEGA, though maybe just the professional one, uses 16GB of HBM2. Even Anandtech are scratching their head over that one! I wonder if they have a separate 4-die version for the professional market?

Also, unless I missed it there didn't seem to be much on the Radeon Vega i.e. the gaming one? Maybe they just didn't want to steal the glory from its own announcement, but that 'late June' seemed to refer to the professional card.

[CAT-THE-FIFTH]

Off-topic a bit (but when has that ever stopped is? ) - they've confirmed a version of VEGA, though maybe just the professional one, uses 16GB of HBM2. Even Anandtech are scratching their head over that one! I wonder if they have a separate 4-die version for the professional market?

Also, unless I missed it there didn't seem to be much on the Radeon Vega i.e. the gaming one? Maybe they just didn't want to steal the glory from its own announcement, but that 'late June' seemed to refer to the professional card.

Maybe,but it could be like with the GTX1080 when people couldn't find the GDDR5X modules used listed on the website of the company which made them. It is quite possible there might be 8GB chips which AMD can use.

[Kumagoro]

So is that EPYC as in epic do you reckon?

Well after reading other tech news it would appear it is. A bit cheesy but it could be worse.

[watercooled]

Yeah I was thinking the same to be honest, especially given its target market...

[badass]

Off-topic a bit (but when has that ever stopped is? ) - they've confirmed a version of VEGA, though maybe just the professional one, uses 16GB of HBM2. Even Anandtech are scratching their head over that one! I wonder if they have a separate 4-die version for the professional market?

Also, unless I missed it there didn't seem to be much on the Radeon Vega i.e. the gaming one? Maybe they just didn't want to steal the glory from its own announcement, but that 'late June' seemed to refer to the professional card.

With rumours of only 16,000 cards available, if true, AMD has a choice:

1. Sell those 16,000 cards for megabucks for professional users who don't mind paying 3 times the price. All cards will sell out.
2. Sell some or all of those cards for 1/3 of the price to Gamers.

I know which I'd choose!

[CAT-THE-FIFTH]

First Ryzen based laptop teased:

https://www.overclock3d.net/news/sys...red_notebook/1

[CAT-THE-FIFTH]

Interesting:

https://videocardz.com/69679/officia...se-at-computex

It looks AMD might be dual sourcing HBM2!!