*** AMD ThreadRipper ***

[drunkenmaster]

That is a good point, I hadn't considered if it was physically possible for intel to compete on PCIE lanes. Is one of the reasons for this because AMD implementation of PCIE is more efficient than Intel?

More efficient, no not really, it's simply a case of die size more than anything else. AMD have a larger effective die by splitting it 4 ways and they can afford to go bigger because it's smaller dies. AMD literally couldn't make a chip with the same transistor count and same features if it was one die. GloFo has a current 600mm^2 reticule size, meaning it physically can't make chips with a die larger than that size, yet with 4 chips EPYC has something akin to a 770mm^2 of effective die size. Or another way to think of that, if AMD made a single monolithic die they'd have to cut out 170mm^2 of transistors and the things that would get cut to bring the size down would be a couple mem channels, a bunch of cores and a bunch of pci-e lanes.

 

[AllBodies]

Supposedly Intel's 10nm should be going into volume production very soon with the first products towards the end of the year and desktop parts around the end of the year which on paper would give Intel around a year advantage on 10nm. AFAIK performance wise there is generally parity between the two nodes with IBM's 7nm being a bit easier to design products for.

A year headstart is stretching it, as with Broadwell they dropped the desktop version of their first 10nm chips due to yields, temp/heat/power issues so desktop looks like it won't come till mid/late 2018 and the places AMD is competing the most right now and really the higher margin and most important segments for AMD are higher end mainstream desktop, ie 8+ cores on mainstream and HEDT and server, AMD looks like they'll have 7nm chips around the same time or even before Intel has 10nm chips in the same area. How long is the lag between Intel putting normal desktop chips on a process then the server/HEDT versions launching? If we're talking August-December 2018 for hexcore 10nm apus from Intel then we're probably looking at mid 2019 for Skylake-X/SP replacements which AMD should have 7nm chips out by then.

Then unlike Broadwell and early skylake which had no competition from AMD at all except 28nm 5 year old architecture APUs from AMD, AMD will still have 14nm dramatically more competitive and likely far far faster GPU side chips available as Intel launches the new desktop and mobile 10nm chips. So while when Broadwell mobile launched it was far far beyond the closest chips from AMD, Cannonlake won't have that same advantage and will in fact likely be significantly behind many AMD APU benchmarks.

Currently Intel is also saying that the first iteration of 10nm isn't even as good as 14nm ++.
http://techreport.com/review/31660/intel-defends-its-process-technology-leadership-at-14nm-and-10nm

Graph a ways down with Intel saying 14nm ++ having straight up better transistor performance than 10nm, so 10nm like 14nm at first should be offering a reduction in clock speeds for probably the first 18 months compared to 14nm, this is another reason for Intel delaying desktop, without a significant architecture jump forwards then 5-10% lower clock speeds will mean 10nm chips that use less power but aren't actually any faster than Kaby/Coffeelake chips.

Yup, it looks like the earliest we'll see desktop 7700K-class processors from Intel on 10nm is late 2018. Around the same time as AMD start bringing 7nmLP products to shelves.

Intel is using 10nm for Laptop first. And even if they give coffeelake a short lifetime (say 9 months), it isn't even out yet!

Then to really compete with Zen/Zen2 they need to launch a 10nm HEDT platform, not just standard desktop. So that's going to be 2019 sometime.

It looks very plausible Zen2-Threadripper2 on 7nmLP will be out before (or at the same time) as Intel's mature-10nm HEDT product.

And how much different is Ice Lake anyway? I've seen nothing so far touting big performance gains, and it's supposedly as close as 12 months away. It's certainly not MCM, so at best Intel will win 6-8 cores and below with it, and lose in 10+ cores due to their clocks falling off with the monolithic design.

 

[Rroff]

Graph a ways down with Intel saying 14nm ++ having straight up better transistor performance than 10nm, so 10nm like 14nm at first should be offering a reduction in clock speeds for probably the first 18 months compared to 14nm, this is another reason for Intel delaying desktop, without a significant architecture jump forwards then 5-10% lower clock speeds will mean 10nm chips that use less power but aren't actually any faster than Kaby/Coffeelake chips.

It is a trade off for density though where 10nm will pack a lot more into the same space.

Yup, it looks like the earliest we'll see desktop 7700K-class processors from Intel on 10nm is late 2018. Around the same time as AMD start bringing 7nmLP products to shelves.

Intel are claiming they are going to production so if that is true they'll be sitting around on a shelf somewhere for months if they aren't out until late 2018. They also claim to have taped out 2nd generation 10nm products which usually means 10-14 months from that to hitting the shelves.

 

[drunkenmaster]

It is a trade off for density though where 10nm will pack a lot more into the same space.

But with the same basic architecture there isn't much more to shove in that space, just margins to increase. There is a reason that even with new nodes performance is primarily increasing with clock speed while die size on each node reduces rather than staying the same and offering way more cores or performance by changing the architecture more significantly. So the density mostly effects Intel's bottom line rather than the end user and considering the first gen of 10nm will offer lower clock speeds at least the first range of chips aren't looking like being at all exciting products for the end users. There is a reason they feel forced to go to hexcore on APU, because they have spare die size in abundance and massive margins to eat into but architectural gains are getting smaller and smaller.

But the main point is by the time Intel get the bigger chips out, HEDT/Server, AMD will also have 7nm chips available and current process to next process, AMD is going to get a dramatically larger increase in density and clock speed compared to what Intel will gain and well, AMD look almost certain to use that extra die space for actual performance rather than just increasing their own margins. We already know 7nm EPYC will be 48 cores, which heavily implies we'll be seeing 6 cores per CCX, which would mean rather than higher margin 16 core HEDT, 8 core mainstream and 4 core APU we're looking to be seeing 24 core HEDT, 12 core mainstream and 6 core APU and with maybe up to 25% clock speed increases.

 

[AllBodies]

Intel are claiming they are going to production so if that is true they'll be sitting around in a shelf somewhere for months if they aren't out until late 2018.

The most recent thing I can find is articles from June saying Ice Lake is "taped-in", which is before "tape-out" and tape-out normally means 9-12 months from shelves at the earliest.

Coffee Lake is 14nm desktop and is also still Skylake anyway.

Cannonlake is just Coffee Lake shrunk to 10nm, and Tiger Lake is a refinement of Cannonlake.

So Ice Lake is the first thing that isn't a revision/shrink/iteration of Skylake, and it hasn't taped-out yet.

Also since by Intel's own figures their 10nm is over 2x the density of their 14nm, I expect Ice Lake desktop to be 8-core like Ryzen. If it isn't, they're just taking the mick even more. And it looks like 7nm Zen2 is going to be 6-cores per CCX, meaning Ryzen2 will be 12-core. So they need to be careful if they attempt to continue holding back core-count.

 

[Rroff]

Tape in isn't really a thing anyhow even if it was its usually a 2-6 week period before tape out - maybe a bit longer for 10nm design I dunno.

 

[drunkenmaster]

Tape in isn't really a thing anyhow even if it was its usually a 2-6 week period before tape out - maybe a bit longer for 10nm design I dunno.

Tape outs do get significantly longer as the nodes shrink, more complexity in process lengthens the process but ultimately they aren't producing Icelake any time soon, Cannonlake has gone into production but it's strictly dual core lower power mobile stuff as with Broadwell being out for nearly a year before the desktop Broadwell hit.

yes it's not a small market and laptops are important, but AMD don't even have their Zen 14nm stuff out today. 14nm Zen APUs, both dual and quad core will be dramatically more competitive even against 10nm Cannonlake than 28nm AMD APUs are right now against 14nm Kabylake mobile APUs. So currently the gap between Intel/AMD in mobile is far larger than it will be early next year when we have 14nm Zen APUs vs a mixture of quad core 14nm Kabylake APU and dual core Cannonlake, then a year later it will be quad/dual core Intel 10nm apus against quad/dual core AMD 7nm APUs, well, knowing AMD, maybe hex core APUs on 7nm.

So ultimately without Intel going for desktop 10nm much if at all before AMD do and AMD might be ahead when it comes to server/hedt, then being first to 10nm mobile isn't going to have much of an effect on AMD at all.

The stuff they are talking about taping out is their first real desktop stuff and that kind of lead in time puts it towards late 2018. In general it's usually more like 18 months for tape out of x86 stuff so I'm expecting Icelake to be very late 2018 or early 2019 which is about the same time frame I'm expecting 7nm Zen to be making an appearance.

 

[AllBodies]

Tape outs do get significantly longer as the nodes shrink, more complexity in process lengthens the process but ultimately they aren't producing Icelake any time soon, Cannonlake has gone into production but it's strictly dual core lower power mobile stuff as with Broadwell being out for nearly a year before the desktop Broadwell hit.

yes it's not a small market and laptops are important, but AMD don't even have their Zen 14nm stuff out today. 14nm Zen APUs, both dual and quad core will be dramatically more competitive even against 10nm Cannonlake than 28nm AMD APUs are right now against 14nm Kabylake mobile APUs. So currently the gap between Intel/AMD in mobile is far larger than it will be early next year when we have 14nm Zen APUs vs a mixture of quad core 14nm Kabylake APU and dual core Cannonlake, then a year later it will be quad/dual core Intel 10nm apus against quad/dual core AMD 7nm APUs, well, knowing AMD, maybe hex core APUs on 7nm.

So ultimately without Intel going for desktop 10nm much if at all before AMD do and AMD might be ahead when it comes to server/hedt, then being first to 10nm mobile isn't going to have much of an effect on AMD at all.

The stuff they are talking about taping out is their first real desktop stuff and that kind of lead in time puts it towards late 2018. In general it's usually more like 18 months for tape out of x86 stuff so I'm expecting Icelake to be very late 2018 or early 2019 which is about the same time frame I'm expecting 7nm Zen to be making an appearance.

It'll likely be 6 core + around 1000 GPU cores, up from 4 core + 700 GPU cores for AMD's 7nm APUs.

Zen2 looks very likely it's 6 cores per CCX, due to 'Starship' the EPYC successor being 48 cores, and the top APU is 1 CCX + GPU cores. Also 7nmLP gives a 60% power reduction like-for-like, so they'll probably use that dramatic power reduction to go for similar power draw but far more compute power. I can't see much reason to reduce the APU power draw further, the Zen + Vega APUs really come out at 50% less power draw than the Excavator APUs (which are already fine for battery life).

And also bear in mind we're seeing the full R7 1700 get into laptops on 14nm.

 

[awaybreaktoday]

From the charts in the above looks like Boost is 4.0 on all cores xfr 4.1.....

Quote
Published on Jul 19, 2017
Exclusive first look at AMD's Ryzen 1950X 16 core / 32 thread Cinebench R15 run against Intel Xeon E5-2679v4 20core / 40thread.

I'm proud to bring you this first look at the rendering power of the soon to be mass released Threadripper 1950X 16core professional CPU from AMD.

Special thanks to Filip at Dell for early access to Threadripper hardware.

AMD 1950X system specs:
Threadripper 1950X 16c/32t (8 MB L2 + 32MB L3)
x399 motherboard
64GB 3200mhz RAM c15
850W Platinum PSU
SSD M.2 + 8TB HDD

Xeon system specs:
Intel Xeon E5-2679v4 20c/40c (50 MB L3)
x99v3 Supermicro motherboard
256GB 2400m

 

[AllBodies]

From the charts in the above looks like Boost is 4.0 on all cores xfr 4.1.....

Quote
Published on Jul 19, 2017
Exclusive first look at AMD's Ryzen 1950X 16 core / 32 thread Cinebench R15 run against Intel Xeon E5-2679v4 20core / 40thread.

I'm proud to bring you this first look at the rendering power of the soon to be mass released Threadripper 1950X 16core professional CPU from AMD.

Special thanks to Filip at Dell for early access to Threadripper hardware.

AMD 1950X system specs:
Threadripper 1950X 16c/32t (8 MB L2 + 32MB L3)
x399 motherboard
64GB 3200mhz RAM c15
850W Platinum PSU
SSD M.2 + 8TB HDD

Xeon system specs:
Intel Xeon E5-2679v4 20c/40c (50 MB L3)
x99v3 Supermicro motherboard
256GB 2400m

I'm not sure that's right unless they overclocked it. It's meant to have a 3.5 or 3.6 GHz all-core boost at stock.

Also that adds up with the 1800X, it scores about 1620 at stock all-core boost of 3.7 GHz, and about 1780 overclocked to 4.1 GHz all-core.

So unless there's a scaling problem, the 1950X must be running about 3.5-3.6 GHz.

I think they're maybe confused because XFR has 2 different max speeds. One for 1-2 cores and one for all-cores.

 

[Rroff]

Tape outs do get significantly longer as the nodes shrink, more complexity in process lengthens the process but ultimately they aren't producing Icelake any time soon, Cannonlake has gone into production but it's strictly dual core lower power mobile stuff as with Broadwell being out for nearly a year before the desktop Broadwell hit.

yes it's not a small market and laptops are important, but AMD don't even have their Zen 14nm stuff out today. 14nm Zen APUs, both dual and quad core will be dramatically more competitive even against 10nm Cannonlake than 28nm AMD APUs are right now against 14nm Kabylake mobile APUs. So currently the gap between Intel/AMD in mobile is far larger than it will be early next year when we have 14nm Zen APUs vs a mixture of quad core 14nm Kabylake APU and dual core Cannonlake, then a year later it will be quad/dual core Intel 10nm apus against quad/dual core AMD 7nm APUs, well, knowing AMD, maybe hex core APUs on 7nm.

So ultimately without Intel going for desktop 10nm much if at all before AMD do and AMD might be ahead when it comes to server/hedt, then being first to 10nm mobile isn't going to have much of an effect on AMD at all.

The stuff they are talking about taping out is their first real desktop stuff and that kind of lead in time puts it towards late 2018. In general it's usually more like 18 months for tape out of x86 stuff so I'm expecting Icelake to be very late 2018 or early 2019 which is about the same time frame I'm expecting 7nm Zen to be making an appearance.

You are thinking that Intel will continue on their trajectory as per before Ryzen and not double down in response. When I posted about it a bit over a month ago I basically posted what you are saying but things have now changed.

 

[Panos]

From the charts in the above looks like Boost is 4.0 on all cores xfr 4.1.....

Quote
Published on Jul 19, 2017
Exclusive first look at AMD's Ryzen 1950X 16 core / 32 thread Cinebench R15 run against Intel Xeon E5-2679v4 20core / 40thread.

I'm proud to bring you this first look at the rendering power of the soon to be mass released Threadripper 1950X 16core professional CPU from AMD.

Special thanks to Filip at Dell for early access to Threadripper hardware.

AMD 1950X system specs:
Threadripper 1950X 16c/32t (8 MB L2 + 32MB L3)
x399 motherboard
64GB 3200mhz RAM c15
850W Platinum PSU
SSD M.2 + 8TB HDD

Xeon system specs:
Intel Xeon E5-2679v4 20c/40c (50 MB L3)
x99v3 Supermicro motherboard
256GB 2400m

Automated XFR on Ryzen 9 works on 4 cores. 2 per die. And there is no indication from the video that is running on all cores as you say.
And 3200Mhz C15 is slow ram

Funny enough, I was looking at the EPYC pricing. The 16 core one costs $750, so it should be pitted against the 2679v4

 

[drunkenmaster]

You are thinking that Intel will continue on their trajectory as per before Ryzen and not double down in response. When I posted about it a bit over a month ago I basically posted what you are saying but things have now changed.

What do you think they've changed in the past month, you can't move a process forward that quickly, 10nm + isn't going to come sooner at this stage. Icelake won't come quicker at this stage, Cannonlake isn't going to be more than dual core, Coffeelake can't have more than 6 cores. The nearest thing whose schedule can really change is the successor to Icelake because they can decide to start working on the tapeout process much sooner. Even then they'll be causing a much smaller gap than usual between releases, say they could bring in the successor 4-6 months, that also means 4-6 months less to work on improvements and it means 4-6 months less process tweaks when it's ready.

Intel haven't changed much of anything, there are stories about them working on a game changing architecture, but that architecture isn't going to be icelake or the icelake successor, we're talking likely 2.5-3.5years away.

The easiest/main thing Intel can do is bake in a infinity fabric type interconnect and move their servers into MCM style packages to match what AMD are doing with EPYC, and that could also come down to HEDT for Intel. But again at this stage I would be surprised to see that before the successor to Icelake-SP and being a couple of years away.

CHanges made now by Intel in the past 6 months won't actually cause a change to the end user for a couple of years and as before, even when those changes start to appear, unlike now, AMD will be on a comparable process rather than one a decent step behind Intel, so for any major leap Intel can make, AMD will make similarly large gains simply from process competitiveness. Arguably AMD will get far more IPC performance than intel over the next couple of generations simply because Zen is a new design and will find real world bottlenecks to be fixed while Intel's core architecture is near the end of it's life and has been optimised to within an inch of it's life.

 

[Rroff]

I don't expect to see any difference to Coffee Lake - AMD hasn't left them room to manoeuvre on that any how.

 

[haszek]

I really love the look of the Asus board
http://wccftech.com/asus-rog-zenith-extreme-x399-amd-ryzen-threadripper-cpu-motherboard/

unknown price at this point but expectations are close to $500

 

[N19h7m4r3]

I really love the look of the Asus board
http://wccftech.com/asus-rog-zenith-extreme-x399-amd-ryzen-threadripper-cpu-motherboard/

unknown price at this point but expectations are close to $500

I bloody well hope not! A decent x299 board like the TUF Mark 1 is significantly cheaper than that.
I hope AMD's partners don't go silly on the motherboard prices; as that shoots the platform cost in the leg.

 

[haszek]

I bloody well hope not! A decent x299 board like the TUF Mark 1 is significantly cheaper than that.
I hope AMD's partners don't go silly on the motherboard prices; as that shoots the platform cost in the leg.

But this one is based on rampage not tuf

 

[Curlyriff]

Yeah you can pick up a reasonble X299 board for £220 so honestly these need to be no more than the £300 ish mark for the ultra high end with them down to £200 for the lower end to really be viable.

 

[gavinh87]

I bloody well hope not! A decent x299 board like the TUF Mark 1 is significantly cheaper than that.
I hope AMD's partners don't go silly on the motherboard prices; as that shoots the platform cost in the leg.

There will certainly be boards that don't cost $500.
I expect £250-£300 for a decent one.

 

[Deleted member 66701]

Got my eye on the Taichi - wish it had a bit more white like the x370 board though as I've just got white ram!