Discussion AMD Ryzen MegaThread! FAQ and Resources

[juanrga]

I also find it very hard to believe that the actual TDP for the 8 core 16 thread R7 1800X is only 95W. To be honest 125W wouldn't really surprise me, but if AMD comes in at your estimated 105W - 117W that would still be amazing considering the i7 6900K is has a TDP listed at 140W. If AMD actually pulls off 95W though that would make for a "cool running" 8 core 16 thread processor and leave a little more headroom for overclocking, in theory.

My former 105W estimation was confirmed by CPCHardware

https://mobile.twitter.com/CPCHardware/status/828666744432652290/actions

but it is without turbo. The 117W is my estimation with turbo enabled. Reviews will show which is the real dissipation, but the comments from the computerbase editor seem to confirm it is more close to 120W than to 95W.

Note also that the 140W on Broadwell-E include the power hungry 2x256bit SIMD units lacking in RyZen.

 

[juanrga]

About all the above discussion about foundries, recall that

Globalfoundries "14nm" = true 20nm

Globalfoundries "7nm" = true 14nm

https://semiaccurate.com/2016/09/26/globalfoundries-7nm-process-isnt-even-close-name/

 

[eidolon171]

Don't get me wrong, I'm not saying they aren't doing great work these days (they are), but saying that they will be ahead of Intel(I don't mean nm, I mean quality of production) seems like a bit of a stretch to me. Definite salt on that one. In my humble opinion, their best case is GF 7nm = INT 10nm, or somewhere thereabouts. Also, if you don't mind my asking, how did you come about the internship? Was it part of school?

Yeah, through my school. The internship was a nanoscale science and technology initiative. It paired undergraduate engineers and physicists together to develop new DFMs (Design For Manufacturing) processes for industry, in exchange for hands-on training with their multi-million dollar equipment (and you know, modest compensation). My team set up a testing station for parametric fault isolation that was supposed to be used for the 10nm samples they had at the Fab10 in Fishkill. I assume it got re-purposed for a different process node though, if it's still even in use.

I understand why you're pessimistic about GF, but they've come a long way since 2012, not least of that credit derived from the acquisition of IBM's manufacturing division (like Fab10), which was a unit arguably capable of equal or possibly even higher quality than Intel's manufacturing division.

Kind of on that point, the whole nanometer naming scheme has lost a lot of its meaning since the mid-noughties. IEEE (Society of Electrical Engineers) pushed for the establishment of new standards, which ties the "nm" label to a design ratio, rather than a design dimension. Similar to the changes made in RF (Radio Frequency) parameters for 3g, 4g, 5g, etc. My understanding is that Intel's 45nm Core/Nehalem was by most accounts a literal 45nm design dimension, but as to whether or not the same is true of subsequent die-shrinks beyond that, it's merely speculation to those of us not on Intel's payroll. Intel has a better product out there today, and that surely influences our belief in the superiority of some of their engineering practices, but in an ever-changing technological landscape what's true today may not be true tomorrow.

I think there is a high probability GF's 14LPP is going to be successful. As far as 7nm Zen+, AMD's Gray Hawk, and all that goes, I'm just cautiously optimistic.

 

[french toast]

About all the above discussion about foundries, recall that

Globalfoundries "14nm" = true 20nm

Globalfoundries "7nm" = true 14nm

There]https://semiaccurate.com/2016/09/26/globalfoundries-7nm-process-isnt-even-close-name/

There is alot of speculation there and no hard facts, you could rightly argue im doing the same thing as i cant post links ATM, but a quick and easy google search pulls up a recent ee times article which shows tsmc and others have smaller sram sizes and fin pitch at 7nm vs intels 10nm, you can check your self, intel is ahead today with 14nm and 10nm Q4 2017, but as soon as the big three hit 7nm in 2018 that advantage is gone and MAYBE reversed.

People forget about all the trouble with yeilds for 14nm, all evidence points to 10nm being far worse, intel have gone on record saying 10nm will be used to lower power consumption not boost frequency, that might change with zen or it might indicate they have hit their frequency limit with skylake uarch and can only cut power.

Times are changing.

 

[french toast]

Don't get me wrong, I'm not saying they aren't doing great work these days (they are), but saying that they will be ahead of Intel(I don't mean nm, I mean quality of production) seems like a bit of a stretch to me. Definite salt on that one. In my humble opinion, their best case is GF 7nm = INT 10nm, or somewhere thereabouts. Also, if you don't mind my asking, how did you come about the internship? Was it part of school?

Yeah, through my school. The internship was a nanoscale science and technology initiative. It paired undergraduate engineers and physicists together to develop new DFMs (Design For Manufacturing) processes for industry, in exchange for hands-on training with their multi-million dollar equipment (and you know, modest compensation). My team set up a testing station for parametric fault isolation that was supposed to be used for the 10nm samples they had at the Fab10 in Fishkill. I assume it got re-purposed for a different process node though, if it's still even in use.

I understand why you're pessimistic about GF, but they've come a long way since 2012, not least of that credit derived from the acquisition of IBM's manufacturing division (like Fab10), which was a unit arguably capable of equal or possibly even higher quality than Intel's manufacturing division.

Kind of on that point, the whole nanometer naming scheme has lost a lot of its meaning since the mid-noughties. IEEE (Society of Electrical Engineers) pushed for the establishment of new standards, which ties the "nm" label to a design ratio, rather than a design dimension. Similar to the changes made in RF (Radio Frequency) parameters for 3g, 4g, 5g, etc. My understanding is that Intel's 45nm Core/Nehalem was by most accounts a literal 45nm design dimension, but as to whether or not the same is true of subsequent die-shrinks beyond that, it's merely speculation to those of us not on Intel's payroll. Intel has a better product out there today, and that surely influences our belief in the superiority of some of their engineering practices, but in an ever-changing technological landscape what's true today may not be true tomorrow.

I think there is a high probability GF's 14LPP is going to be successful. As far as 7nm Zen+, AMD's Gray Hawk, and all that goes, I'm just cautiously optimistic.

Your comment is appreciated, people dont realise just how good IBM are/were with manufacturing technology, globalfoundries were a disaster for a few years but they have got back on track, then you have tsmc and Samsung who are being funded by Apple, Qualcomm, nvidia and samsung themselfs, they have made huge strides lately and if their data they have provided is true then by 2018 they could overtake intel, possibly all three.

 

[french toast]

I also find it very hard to believe that the actual TDP for the 8 core 16 thread R7 1800X is only 95W. To be honest 125W wouldn't really surprise me, but if AMD comes in at your estimated 105W - 117W that would still be amazing considering the i7 6900K is has a TDP listed at 140W. If AMD actually pulls off 95W though that would make for a "cool running" 8 core 16 thread processor and leave a little more headroom for overclocking, in theory.

My former 105W estimation was confirmed by CPCHardware

https://mobile.twitter.com/CPCHardware/status/828666744432652290/actions

but it is without turbo. The 117W is my estimation with turbo enabled. Reviews will show which is the real dissipation, but the comments from the computerbase editor seem to confirm it is more close to 120W than to 95W.

Note also that the 140W on Broadwell-E include the power hungry 2x256bit SIMD units lacking in RyZen.

At the recent Amd avent, they did a perf per watt measurement at the wall, with the amd platform consuming 30 watts less, this was intel loved cinebench so no avx256 or even dual channel which would have made intel look worse, the tdst was not advertised but it is in the footnotes.
Also we have a chinese leak putting a 1700x vs 6800k 8/16 vs 6/12, both locked at 3.4ghz.
Through a number of tests it showed 1700x consuming between 25-39% less energy, the 6800k had 4 less threass and wasn't even running quad channel which have bumped up the watts for intel by at least 5w.

Those are more legit that canard QS sample on non final firmware and bios.
Both companies use different criteria to judge tdp so its possible to push both past tdp in various circumstances, except intel has that 256bit tdp headroom.

Check forbes for the last story and AT forum for the last.

Your right lets wait for reviews.

 

[french toast]

@French Toast
Why do you think it will be >= Intel's 10nm?
Given past history, between all major fabs that produce CPUs (SS, Intel, IBM, GF) Intel has the best node and is closest to the marketed number. GF/SS 14 is not a true 14nm and nothing (to me) points to true 7nm, or it equalling intels 10nm that will be slightly mature by that time.

I disagree, I think Global Foundries is doing great work these days. I interned at the facility in Malta, NY. The lab was renovated after they bought the IBM IP with multi-billions.

http://saratogatodaynewspaper.com/index.php/today-in-saratoga/business/item/203-globalfoundries-ups-malta-investment-by-another-$2-billion

And worked with the University at Albany to start this massive R&D project,

http://www.globalfoundries.com/newsroom/press-releases/2016/02/09/suny-poly-and-globalfoundries-announce-new-$500m-r-d-program-in-albany-to-accelerate-next-generation-chip-technology

They even made it to the table first with a 5G RFSOI,

http://www.globalfoundries.com/newsroom/press-releases/2017/02/17/globalfoundries-announces-availability-of-45nm-rf-soi-to-advance-5g-mobile-communications

GF is top-tier these days. I helped set up the 10nm fab line (Which they shut down a while back, and skipped ahead to fitting the 7nm fab line), and it produced really high quality samples. If they say they can manufacture a viable 7nm AMD CPU, I believe them.

First hand experience is appreciated, cheers

 

[juanrga]

The first RyZen review?

https://www.guru3d.com/news-story/first-ryzen-7-1700x-review-finds-its-way-onto-the-web.html

 

[juanrga]

I also find it very hard to believe that the actual TDP for the 8 core 16 thread R7 1800X is only 95W. To be honest 125W wouldn't really surprise me, but if AMD comes in at your estimated 105W - 117W that would still be amazing considering the i7 6900K is has a TDP listed at 140W. If AMD actually pulls off 95W though that would make for a "cool running" 8 core 16 thread processor and leave a little more headroom for overclocking, in theory.

My former 105W estimation was confirmed by CPCHardware

https://mobile.twitter.com/CPCHardware/status/828666744432652290/actions

but it is without turbo. The 117W is my estimation with turbo enabled. Reviews will show which is the real dissipation, but the comments from the computerbase editor seem to confirm it is more close to 120W than to 95W.

Note also that the 140W on Broadwell-E include the power hungry 2x256bit SIMD units lacking in RyZen.

At the recent Amd avent, they did a perf per watt measurement at the wall

"at the wall"... but I am not discussing total platform power. I am discussing only CPU dissipation.

 

[salgado18]

The first RyZen review?

https://www.guru3d.com/news-story/first-ryzen-7-1700x-review-finds-its-way-onto-the-web.html

1700X loses in all game benchmarks, although almost all are on old tech, and rely more on single-thread + frequency combo, which is a lot higher on the Intels. But then it wins on threaded benches, and loses to the 6950x because of two less cores. Also it was tested with a very bad 2133 RAM, while the others were using up to 3200.

Even if correct and valid, better wait for more in-depth reviews, with comparable platforms and relevant benches. But, in any case, it's very nice to see Ryzen in action (and keeping up in CB15 ST 😀 )


Edit: Ryzen 7 1700 price just went up in my country since anouncement! Get ready to pay early adopters premium

 

[James Mason]

The first RyZen review?

https://www.guru3d.com/news-story/first-ryzen-7-1700x-review-finds-its-way-onto-the-web.html

1700X loses in all game benchmarks, although almost all are on old tech, and rely more on single-thread + frequency combo, which is a lot higher on the Intels. But then it wins on threaded benches, and loses to the 6950x because of two less cores. Also it was tested with a very bad 2133 RAM, while the others were using up to 3200.

Even if correct and valid, better wait for more in-depth reviews, with comparable platforms and relevant benches. But, in any case, it's very nice to see Ryzen in action (and keeping up in CB15 ST 😀 )

Well, they said because of BIOS issues they couldn't get the RAM to OC at all, even to XMP levels. Kinda why early reviews are bad idea. The actual review samples may have special access to those things.

 

[8350rocks]

About all the above discussion about foundries, recall that

Globalfoundries "14nm" = true 20nm

Globalfoundries "7nm" = true 14nm

https://semiaccurate.com/2016/09/26/globalfoundries-7nm-process-isnt-even-close-name/

The old 14nm XM was a 20nm BEOL with 14nm FEOL which would have been something like 16nm overall.

The 14nm Samsung process is smaller than GF's 14nm node.

Also, the 7nm node is IBM's 7nm node, and it is of no relation to Samsung's current node, or GF's old 14nm node. 7nm @ GF using IBM tech is expected to be equally dense, or more dense, than intel's potential 10nm.

EDIT: Charlie is usually reliable about this sort of stuff...but the numbers for the process for Ryzen are already out...and the gaps between it and intel are quite small in terms of density and transistors. Intel's 14nm node is not even 14nm as it is defined scientifically...and the Samsung 14nm node is much closer to Intel's 14nm than anyone else.

 

[juanrga]

The first RyZen review?

https://www.guru3d.com/news-story/first-ryzen-7-1700x-review-finds-its-way-onto-the-web.html

1700X loses in all game benchmarks, although almost all are on old tech, and rely more on single-thread + frequency combo, which is a lot higher on the Intels. But then it wins on threaded benches, and loses to the 6950x because of two less cores. Also it was tested with a very bad 2133 RAM, while the others were using up to 3200.

Even if correct and valid, better wait for more in-depth reviews, with comparable platforms and relevant benches. But, in any case, it's very nice to see Ryzen in action (and keeping up in CB15 ST 😀 )


Edit: Ryzen 7 1700 price just went up in my country since anouncement! Get ready to pay early adopters premium

The memory used was in reality a 3200 kit

but it seems the mobo downclocked it to default due a problem with the XMP profile. Of course, we must wait for proper reviews from Toms, Anand, and others, but good to have something to read whereas waiting.

 

[jaymc]

They must be working on a bios update.. otherwise they would want the reviews out before their product is available to the public, wouldn't they ?

Edit:

Looks like there starting to sell out 1800's gone off Amazon and Newegg.

 

[juanrga]

About all the above discussion about foundries, recall that

Globalfoundries "14nm" = true 20nm

Globalfoundries "7nm" = true 14nm

https://semiaccurate.com/2016/09/26/globalfoundries-7nm-process-isnt-even-close-name/

The old 14nm XM was a 20nm BEOL with 14nm FEOL which would have been something like 16nm overall.

The 14nm Samsung process is smaller than GF's 14nm node.

Also, the 7nm node is IBM's 7nm node, and it is of no relation to Samsung's current node, or GF's old 14nm node. 7nm @ GF using IBM tech is expected to be equally dense, or more dense, than intel's potential 10nm.

EDIT: Charlie is usually reliable about this sort of stuff...but the numbers for the process for Ryzen are already out...and the gaps between it and intel are quite small in terms of density and transistors. Intel's 14nm node is not even 14nm as it is defined scientifically...and the Samsung 14nm node is much closer to Intel's 14nm than anyone else.

I am not talking about the old 14XM which was canceled before seeing the light, but about Globalfoundries 14LPP which is used by AMD on RyZen. 14LPP has a 20nm BEOL and a 14nm FEOL, but the size is not anything close to 16nm, but ~19nm. There is no "19nm" node per ITRS rules, aka 14LPP is a true 20nm node.

Charlie claim that Glofo "7nm" is in reality a 14nm node is based in information given by Globalfoundries. Check the link.

 

[gamerk316]

The first RyZen review?

https://www.guru3d.com/news-story/first-ryzen-7-1700x-review-finds-its-way-onto-the-web.html

IPC looks a touch below Haswell, which is about what we all figured. Wins synthetics due to having more cores, again, about what we expected.

I'll run some quick math, but I don't see anything too surprising. I'd keep an eye on those memory numbers though; potentially a BIOS issue, but then again, remember the Bulldozer launch?

EDIT

I'm not going to crunch the gaming numbers, because they look bad due to lack of core loading. I'll use Cinebench to calculate IPC, as it should scale.

Single Thread 1700x:
148 = IPC * 3.4
IPC = 43.53

Single Thread 4770k:
154 = IPC * 3.5
IPC = 44

Virtually identical in single thread performance, implying Ryzen falls *slightly* behind Haswell.

Multi-thread 1700x:
1507 = IPC * 3.4 * 16
IPC = 27.70

Multi-thread 4770k:
795 = IPC * 3.5 * 8
IPC = 28.39

Same result, Haswell *slightly* ahead in IPC.

Now I will calculate SMT gains, since I have both Single and Multi numbers for Ryzen:

Single-Thread performance = 148 * 16 = 2368
Actual Multi Thread performance = 1507

Performance Difference: 44.44%

Compared to HTT:

Single-Thread performance = 154 * 8 = 1232
Actual Multi Thread performance = 795

Performance Difference: 43.12%

So HTT and SMT are putting up similar improvement numbers (as measured against perfect single-thread scaling).

Nothing too shocking, though I honestly expected better SMT scaling then what AMD is getting right now.

 

[french toast]

Note the iranian review was on a ES sample, had crap memory and cooler, no turbo and far from final bios, it still looked awesome

oh r1700 gets xfr at presumably 50mhz, the 'x' suffix models get double xfr on x370 mobos.

 

[french toast]

The first RyZen review?

https://www.guru3d.com/news-story/first-ryzen-7-1700x-review-finds-its-way-onto-the-web.html

IPC looks a touch below Haswell, which is about what we all figured. Wins synthetics due to having more cores, again, about what we expected.

I'll run some quick math, but I don't see anything too surprising. I'd keep an eye on those memory numbers though; potentially a BIOS issue, but then again, remember the Bulldozer launch?

what planet are you on? Its 2% below broadwell and 6.5% bellow kabylake.

Edit looks like the same ES sample as canard, its running no turbo or xfr, and is memory starved with non final firmware

 

[french toast]

About all the above discussion about foundries, recall that

Globalfoundries "14nm" = true 20nm

Globalfoundries "7nm" = true 14nm

https://semiaccurate.com/2016/09/26/globalfoundries-7nm-process-isnt-even-close-name/

The old 14nm XM was a 20nm BEOL with 14nm FEOL which would have been something like 16nm overall.

The 14nm Samsung process is smaller than GF's 14nm node.

Also, the 7nm node is IBM's 7nm node, and it is of no relation to Samsung's current node, or GF's old 14nm node. 7nm @ GF using IBM tech is expected to be equally dense, or more dense, than intel's potential 10nm.

EDIT: Charlie is usually reliable about this sort of stuff...but the numbers for the process for Ryzen are already out...and the gaps between it and intel are quite small in terms of density and transistors. Intel's 14nm node is not even 14nm as it is defined scientifically...and the Samsung 14nm node is much closer to Intel's 14nm than anyone else.

I am not talking about the old 14XM which was canceled before seeing the light, but about Globalfoundries 14LPP which is used by AMD on RyZen. 14LPP has a 20nm BEOL and a 14nm FEOL, but the size is not anything close to 16nm, but ~19nm. There is no "19nm" node per ITRS rules, aka 14LPP is a true 20nm node.

Charlie claim that Glofo "7nm" is in reality a 14nm node is based in information given by Globalfoundries. Check the link.

He gives no hard data juan, check the projected sram cell sizes and fin pitch for both intel and glofo.
Glofo 7nm is smaller intels 10nm
Speaking of which more reports today of more delays for 10nm.

 

[juanrga]

Note the iranian review was on a ES sample, had crap memory and cooler, no turbo and far from final bios, it still looked awesome

oh r1700 gets xfr at presumably 50mhz, the 'x' suffix models get double xfr on x370 mobos.

The first RyZen review?

https://www.guru3d.com/news-story/first-ryzen-7-1700x-review-finds-its-way-onto-the-web.html

IPC looks a touch below Haswell, which is about what we all figured. Wins synthetics due to having more cores, again, about what we expected.

I'll run some quick math, but I don't see anything too surprising. I'd keep an eye on those memory numbers though; potentially a BIOS issue, but then again, remember the Bulldozer launch?

what planet are you on? Its 2% below broadwell and 6.5% bellow kabylake.

Edit looks like the same ES sample as canard, its running no turbo or xfr, and is memory starved with non final firmware

It is not an ES. The Ryzen ES start with a "1" or a "2" depending if it is a first gen or second gen engineering sample. This chip start with a "Z". It is the QS for the 1700X. It is technically identical to the retail chip. The QS are the chips sent to mobo makers, integrators, etc. just to ensure that their hardware/software will work with the retail chips. The differences between a QS and the retail chip are "cosmetic". The QS doesn't come in a fancy box, and identifies itself with a encoded string instead with the commercial name.

 

[juanrga]

No sure if this is a joke or real, but CanardPC is reporting the existence of a Special BIOS for Reviews

 

[-Fran-]

That is a ROG board. I will say that is Asus exclusive.

I had a couple of ROG boards and they do have *very* specific things for OC and breaking records.

If such setting exist across all review MoBos (which in all Youtube videos seem to be different), then there might be a conspiracy theory to follow. For now, tons of salt there, Juan.

Cheers!

 

[juanrga]

That is a ROG board. I will say that is Asus exclusive.

I had a couple of ROG boards and they do have *very* specific things for OC and breaking records.

If such setting exist across all review MoBos (which in all Youtube videos seem to be different), then there might be a conspiracy theory to follow. For now, tons of salt there, Juan.

Cheers!

They refer to this like the "BIOS Presse Asus" a "Media Review BIOS"

 

[french toast]

Like i keep saying, non final firmware makes all the difference, that firmware enables 3800mhz ddr4 by the way

 

[-Fran-]

That is a ROG board. I will say that is Asus exclusive.

I had a couple of ROG boards and they do have *very* specific things for OC and breaking records.

If such setting exist across all review MoBos (which in all Youtube videos seem to be different), then there might be a conspiracy theory to follow. For now, tons of salt there, Juan.

Cheers!

They refer to this like the "BIOS Presse Asus" a "Media Review BIOS"

How they word it is important in this case, I'd say. Do you have a link for what CanardPC said specifically?

If Ryzen CPUs perform better in *certain* boards in reviews that will affect MoBo sells. Maybe there is merit for a conspiracy, but the plot always thickens 😀

Sounds like I will need more popcorn for the 2nd, haha.

Cheers!

EDIT: Added word.