AMD - Zen chitchat

[Corky34]

Doubt what?

And what does better even mean?

Look I'm not trying to argue with you, quiet the contrary as I've already said that IMO Ryzen offers a better all round balanced system, however not everyone looks at different things, for some people the only thing that counts is higher FPS and their willing to pay hundreds more for an extra 5%

[watercooled]

I think the point that CAT is trying to make, is it's not as simple as just 'higher FPS' as the i5 system gives slightly higher peak FPS but tanks more in the busier scenes which will impact on overall experience.

[Corky34]

I know the point he's making but like i said some people don't care about things like that, for some people best simply means higher FPS.

I'm not saying i agree with that point of view or that it's right but in the end it's their money.

EDIT: Just to clarify since the early days of 3D accelerators and 3DFX I've disagreed with using peak FPS as a metric for measuring what's "best" but soon came to realise when everyone rushed out to buy Matrox, 3D Rage and RIVA 128 that used DirectX instead of Glide that people cared more about peak FPS than quality and ever since then peak FPS as a metric seems to have stuck, probably in no small part because trying to measure "quality" is like trying to heard cats.

[CAT-THE-FIFTH]

Looks like more AMD mini-ITX motherboards have appeared:

https://www.cclonline.com/product/23...board/MBD2209/
https://www.cclonline.com/product/23...board/MBD2188/
https://www.cclonline.com/product/23...board/MBD2216/

[gupsterg]

AMD has confirmed Ryzen's got a problem under certain conditions on Linux.

From what they say it seems under certain conditions, heavily loading multiple threads along with less than optimal cooling, causes Ryzen to read/write too the wrong memory location, i guess because of a localised hot spot within the silicon.

My R7 1800X has the fault, batch code 1710, ie produced week 10 year 17. Not tested the R7 1700, but likely, that is 1709.

Interestingly enough Epyc and Threadripper don't have the same problem so it makes me wonder if they are using a different stepping.

TR is B2 IIRC, Ryzen is B1. Gotta get some bits together for WC and then will do some AIDA64 as that does show stepping info.

[watercooled]

Is TR definitely B2? IIRC that was rumoured early on but retail TR is actually still B1.

[gupsterg]

Edited above post, but will defo know soon .

[CAT-THE-FIFTH]

Looks like if NAND pricing going up was bad enough,so is GDDR so might be worth jumping onto any good graphics card deals you see now:

https://www.techpowerup.com/236201/g...cs-card-prices

HBM2 is probably not affected - good timing for AMD!

[watercooled]

I wonder if that's because demand is up, or supply is down? I don't think mining has as big of an impact on the GPU market (in terms of production) as people make out - it's too volatile to commit to ramping supply in a significant way I would've thought - exactly as AMD found out back in the 2013 craze. Production had likely slowed a bit for Polaris anyway. That's why it affects pricing so much, and stock levels.

Edit: apparently it's due to supply being down according to the source article: http://www.digitimes.com/news/a20170816PD205.html?mod=2

[CAT-THE-FIFTH]

Have people seen this:

http://www.chinatimes.com/newspapers...8000038-260202

(AMD) this year Jedi big counterattack, x86 acceleration processor (APU) and server processor market share steadily rising, the graphics processor (GPU) is also sold out of stock. (2330) embrace, 7 nm APU and GPU orders will be in the first half of next year in the TSMC production capacity of the film production and production, and to speed up the process of micro-speed, has decided to return to the foundry leader TSMC (2330) The

AMD in the 20 years before the 28-nanometer APU and GPU chips, mostly by TSMC OEM production, in 2015 decided to APU and GPU product line into the 14-nanometer fin-type field effect transistor (FinFET) process, the product The line has been handed over to GMF. Therefore, TSMC last year and this year, although actively seeking 16-nanometer orders, but from the main line of the main product line is almost zero.

Restricted Gloriafeld production capacity

AMD launched a new generation of Polaris architecture GPU last year, the market reaction is applauded but not called, the main reason is the occasion of the launch of the occasion, although the massive, but limited by the foundry 14 nanometer production capacity is limited, can not enlarge the shipments, wait until Competitors Huide (NVIDIA) TSMC 16-nanometer process using the Pascal architecture GPU introduced after the AMD GPU foundry production has been out. As the Huida Pascal architecture GPU is a newer architecture, AMD is due to limited production capacity and lost market opportunities.

The second half of fear and see out of stock

AMD has introduced a series of Ryzen processors and EPYC server processors based on the Zen architecture this year, and now appears to have strong sales momentum, the new Vega architecture GPU also received market praise, but AMD is now facing the biggest problem is still Gloriafeld can fully support the required 14 nm and 14+ nanometer production capacity. The industry believes that Glover in recent years, there is no expansion of 14 nanometer production capacity, limited capacity in the case of limited, do not rule out AMD APU and GPU in the second half will still be out of stock situation.

Taiwan after the annual order to eat?

In order to solve the problem of capacity constraints, while speeding up the process of micro-speed, AMD next year 7 nm APU and GPU OEM orders will return to TSMC. According to industry sources, AMD will be introduced next year, several Zen 2 architecture APU processor, and a new generation of Navi architecture GPU chips, etc., will use TSMC 7nm process mass production. If the process is smooth, do not rule out the new generation of 719 nanometer process APU and GPU will also be TSMC won the OEM orders.

AMD has not repeatedly commented on the foundry strategy, however, the executive director Su Zifeng (Lisa Su) has repeatedly said that TSMC and Gloria Plaza are extremely important foundry partners.

Ultra-third quarter of last year, the first six degrees modified with Grosfond's wafer supply agreement, which is a modification of the provisions of the ultra-wafer to Girona Fond other than the foundry to buy wafers, in order to win Other wafer foundries, ultra-pay 100 million US dollars in cash to Ge Luo Fang as a price. In other words, AMD can seek other foundry production capacity support, and TSMC is the first choice for ultra-micro partners.

(Business Times)

So AMD is going to use TSMC 7NM next year for its APUs and GPUs due to restricted 14NM capacity at Global Foundries.

[watercooled]

I do wonder what the new WSA says about using TSMC? Hopefully the AMD legal team did a good job of ensuring they're not penalised for being forced to use another foundry due to GloFo being unable to meet demand!

[CAT-THE-FIFTH]

I do wonder what the new WSA says about using TSMC? Hopefully the AMD legal team did a good job of ensuring they're not penalised for being forced to use another foundry due to GloFo being unable to meet demand!

The second part refers to that - AMD made a $100 milllon payment as part of the WSA(last year) so they could use more than one foundry:

http://www.anandtech.com/show/10631/...t-through-2020

[watercooled]

It's the last bullet point in this slide that looks menacing: http://images.anandtech.com/doci/10631/WSA5.png

[watercooled]

Hmm, it seems Skylake-X is bound by some very similar mesh bandwidth 'issues' as people complained about with Ryzen, and it's also tied to memory clocks. http://www.numberworld.org/y-cruncher/

The Skylake X L3 cache is so slow that it's barely faster than main memory in terms of bandwidth. So for all practical purposes, it's as good as non-existant.

It's likely a result of trying to keep power and die size reasonable with increasing core counts, and something developers will have to adapt to.

Nonetheless, the post makes for some IMO very interesting reading about Skylake-X and TR.

[kompukare]

http://www.numberworld.org/y-cruncher/
Nonetheless, the post makes for some IMO very interesting reading about Skylake-X and TR.

This bit struck me:

It has been confirmed with benchmarks that all the Skylake X desktops have full-throughput FMA. This directly contradicts Intel's pre-release information. Subsequently, all the pre-release reviewers apparently got their information from the same inaccurate source from Intel. So if you are looking to purchase a Skylake X system for the purpose of AVX512, you do not need to spend $1000 to get the Core i9 7900X for the full AVX512. Either of the 6 or 8 core models (7800X and 7820X) will do.

So, is it miscommunication or did something (else) change at the last minute? Since this is something Intel does for product segmentation and (as far as we know) all it takes is for Intel to blow a fuse, this is something Intel could change at a whim. Precisely the kind of thing they might do as a response to Zen.

They could do something similar with the PCIe restriction of the lower end Skylake-X's like the i7-7800X and i7-7820X as those 28 PCIe lane processor they share the same die 44 PCIe lane i7-7900X. Of course that would make the 7900X seem even more overpriced.

[kompukare]

Oh, going back to my earlier post about the TR press kit numbers, it seems someone on Reddit had a similar though only they managed more:
https://www.reddit.com/r/Amd/comment...yzen/?sort=new
Still, only 38 out of 250 identified so far.