Thread Ripper 2, and the effect on Intel pricing

[AndreiD]

@Panos If you reinforce the TDP limits, Skylake-X can be pretty efficient: https://techreport.com/review/32607/intel-core-i9-7980xe-and-core-i9-7960x-cpus-reviewed/13
A lot of reviewers tested with X299 motherboards which don't enforce TDP, and of course when overclocking perf/W is thrown out the window.

@Madpete TSMC's 7nm should enter HVM earlier than Intel's 10nm, so there's going to be a period in which AMD will have 7/10nm class products out while Intel doesn't. It's going to be interesting to see how much of an impact that will have on performance. Gamers Nexus has a video with David Kanter on Intel's 10nm, their delays and how it would compare with other foundry 7nm https://www.youtube.com/watch?v=dtiBEHH7mEA it's an interesting watch.

 

[sideways14a]

I cannot see intels next 12months being too much fun for them, doubt the top end HEDT chips will come close to the threadripper line in price or performance and its only going to get worse from there when the 7 nm 48core Trs appear.

Nope HEDT is a bit of a problem for intel for the foreseeable.

 

[Panos]

I cannot see intels next 12months being too much fun for them, doubt the top end HEDT chips will come close to the threadripper line in price or performance and its only going to get worse from there when the 7 nm 48core Trs appear.

Nope HEDT is a bit of a problem for intel for the foreseeable.

Actually is 18+ months to see HEDT at 10nm from Intel. Yesterdays announcement Intel said new Xeons at 10nm in 2020. Next year full 14nm lineup.

While AMD will be selling 7nm TR next summer.

 

[sideways14a]

Well to be fair i said the next 12months will not be fun for intel, i decided to be nice and not mention the 6 -12 months after that, that will be utterly devastating to there HEDT parts.....
See i thought i would be nice and fluffy, didnt want to get any more intel loverboys reaching for the 9mm or box of pills

 

[AndreiD]

Actually is 18+ months to see HEDT at 10nm from Intel. Yesterdays announcement Intel said new Xeons at 10nm in 2020. Next year full 14nm lineup.

While AMD will be selling 7nm TR next summer.

That's rumor, they already said in their Q2 earnings call that 10nm will be in mass production with "systems on shelves" in H2 2019: https://www.anandtech.com/show/12693/intel-delays-mass-production-of-10-nm-cpus-to-2019
Apparently they're jumping over Cooper Lake and going straight to Ice Lake after Cascade Lake, which puts Ice Lake (10nm Xeons) in 2019 (this is not confirmed by Intel though): https://www.anandtech.com/show/1263...lake-xeon-details-lga4189-and-8channel-memory

On Cascade Lake, one interesting thing is that they're doing Multi-Chip Packages using their EMIB technology, most likely it's going to be Xeon chips with FPGAs, but they could go the AMD route (though unlikely).

 

[humbug]

On Cascade Lake, one interesting thing is that they're doing Multi-Chip Packages using their EMIB technology, most likely it's going to be Xeon chips with FPGAs, but they could go the AMD route (though unlikely).

That's rumour.

I heard a rumor of a Xeon Advanced Processor which was an MCM. I thought it was ICL-SP based but could it be Cascade Lake based?

— Ashraf Eassa (@TMFChipFool) May 19, 2018

https://wccftech.com/intel-cascade-lake-ap-sp-xeon-server-family-mcm-rumor/

 

[smilingcrow]

If Intel was able to make 32core 14nm monolithic chip, they could have done it last year trying to hold off losing market share in the server market.

They couldn’t have done it last year as it takes too long a time to get a chip to market from scratch. Plus 32C on 14nm+ is probably too large a die.
I don’t think the core count will be an issue for Intel until AMD move beyond 8C CCXs, so it’s next year when they will really feel the pain on that level.
Prior to that EPYC is doing well on I/O, pricing etc.

 

[smilingcrow]

You need to do your research, it is unlikely that the 28-Core HEDT part from Computex based on an entirely different architecture, as referred to in another post $10,000 Xeon part, than the the X299 Skylake-X line up. The whole way the demo was done and the subsequent furore about the cooling used, showed that it was done to steal the headlines from the Threadripper 2 launch. There is a Skylake-X refresh in the roadmap for late 2018 but I will wager that the 28-core won’t be a part of the lineup.

They only said they are using the XCC dies which are not a new architecture, just the largest of the 3 die sizes.
See this link which shows the difference in die sizes amongst other things:
https://www.anandtech.com/show/1155...-core-i9-7900x-i7-7820x-and-i7-7800x-tested/6

 

[AndreiD]

Oh, seems they aren't skipping Cooper Lake: https://www.anandtech.com/show/1319...oadmap-cooper-lakesp-and-ice-lakesp-confirmed
So consumer 10nm in H2 2019, Ice Lake Xeons 2020 then.

Seems they're implementing Spectre and Meltdown mitigation with Cascade Lake.

 

[drunkenmaster]

They said consumer 10nm for holiday season 2019, people are mistakenly taking that as summer holidays/school holidays which is never what they mean by that. Holiday season as used by Intel for decades and american companies means thanksgiving to new year period meaning very end of 2019. Intel has also for both 14nm and their first attempt at 10nm gone with the smallest dual core + gpu die possible as their first part with desktop following some 4-6 months later.

Either way this is also Intel making a claim far enough out that frankly they can't know for sure. Intel was promising 10nm launches 3 months before they delayed it by another 18 months. Very end of 2019 is at this stage, not even close to confirmed.

When you see Intel roadmaps in which cooperlake is listed as in planning only very recently. It's seemingly going to be due well into 2019, like probably q3-q4 and to me implies they've felt the need to insert another generation on 14nm because they aren't even close to confident of 10nm being ready by end of next year.

Also should be noted that Charlie is implying that Icelake has a fairly major performance problem such that it could be a completely pointless architecture and everyone will want to wait till Tigerlake anyway. Right now I wouldn't remotely bank on Intel getting any real volume of 10nm before a ways into 2020 and it also looks questionable if Icelake is worth having. We could end up in another Broadwell release where Icelake is delayed till it's launched to meet the statements they made to the FCC previously only just before Tigerlake is launched to lessen the blow of it sucking.

 

[smilingcrow]

Either way this is also Intel making a claim far enough out that frankly they can't know for sure.

Exactly, it seems as if speculation is all we and they have at this point.

 

[drunkenmaster]

What I more specifically meant is most nodes go from early low volume to high volume production after a 6-8 month ramp up in volume. For them to be at high volume manufacturing end of 2019 as they are implying means they need to hit that early point where yields are at an acceptable starting point that they think with a little tweaking we can get yields to a good place for high volume production. 18 months away means they are give or take a year away from where they are hoping for good results from early production after whatever fixes they implement... if they aren't already purposefully lying about the end of 2019 date.

As in, if they currently fully believe they'll have the fixed process in place for say ~May next year and hope to get high enough yields to say this is working and begin the ramp up to full production, they are still counting on the results of early production next May to look good. But as they've failed over and over it's literally far too early for them or anyone else to say they can hit high volume production end of 2019. If they are still saying end of 2019 by the middle of next year that date gains a lot more credibility.

 

[Silent_Scone]

What's going to be interesting is how much they charge for a 28 core HEDT chip seeing as the one they demoed was a £10k chip.

I can tell you one thing, it won't be cheap. Despite the competition. It's been a few months now since Intel asked vendors to look at boards for these, so I'd expect Q4 release.

 

[beany_bot]

Xeon will be bought by huge companies with more money than sense and staff that don't even know amd exist. So hardly a worry for Intel. Said companies will. Just be sold Xeon systems by IT providers who make more profit on those than Epyc or TT systems.

 

[Silent_Scone]

You need to do your research, it is unlikely that the 28-Core HEDT part from Computex based on an entirely different architecture, as referred to in another post $10,000 Xeon part, than the the X299 Skylake-X line up. The whole way the demo was done and the subsequent furore about the cooling used, showed that it was done to steal the headlines from the Threadripper 2 launch. There is a Skylake-X refresh in the roadmap for late 2018 but I will wager that the 28-core won’t be a part of the lineup.

Wouldn't be so sure. I know for a fact Intel asked a certain vendor to design boards for exactly that many cores, and that was around the time of Computex. I'm fairly confident we'll see it by Christmas.

EDIT: Sorry, that's baloney - it was around CES at the start of the year.

 

[hominid]

Intel are in trouble with respect to servers in the near term, people can see their desperate games(1500W industrial chiller attached to one CPU). I don't mean they are going to go to 90% server share any time soon, AMD at the moment are in mid single digit %. But they can easily get to ~20% in the next ~18 months which will be worth billions in very high profit server sales.

I just hope that they can keep up the momentum past 2020 or so after Intel are forced to use glued MCM's .

 

[AndreiD]

Epyc's adoption thus far has been pretty poor though, think the MCM/NUMA latency issues are keeping some away, possibly the lower per core performance (mainly AVX) too since a lot of software is licensed based on # of cores.

 

[MissChief]

Xeon will be bought by huge companies with more money than sense and staff that don't even know amd exist. So hardly a worry for Intel. Said companies will. Just be sold Xeon systems by IT providers who make more profit on those than Epyc or TT systems.

As the saying goes, no one ever got fired for choosing IBM. And no one will get fired for choosing Intel either.

 

[Hotwired]

Epyc's adoption thus far has been pretty poor though, think the MCM/NUMA latency issues are keeping some away, possibly the lower per core performance (mainly AVX) too since a lot of software is licensed based on # of cores.

Poor according to what. I recall a figure from AMD regarding what they were aiming for and it was nothing like the "20%" that was actually and ironically said by Intels CEO... not long before he became the ex-CEO.

Dammed if I can find it now but I recall AMD looking for single figure % gains. Which is what they're getting? Sounds like a realistic target is being hit.

 

[Hades]

I'm coming at this from a very uninformed position so please excuse me if I'm talking out of my rear. But from my standpoint Ryzen feels very much like their Athlon 64 felt like. Suddenly AMD could not only compete but were out in front.