AMD's Ryzen CPUs (Ryzen/TR/Epyc) & Vega/Polaris/Navi GPUs

So, Intel will have higher IPC, but at slightly higher cost. It should have a six-channel memory, if rumors are true. As such, depending on the ability of threads your programs use, and whether Ryzen server chips are unlocked, it should perform better, to a point. Intel is rumored going from 40 up to 48 qpi lanes. Some of these, in a dual processor setup, are dedicated to communication with the other chip, meaning you cannot use all lanes on a multi-cpu setup. But, since each 8-core set is said to add more lanes, this Ryzen 16C should have 48 lanes, give or take. The 24C variant is where a lot shifts toward AMD on performance, albeit due, in part, to raw performance. As such, it is the value of the chip over Intel's at a set number of cores that impresses. As I've said before, for mission critical or those that need best ipc with a lower set number of cores, Intel may be the better purchase (if money is no object). But, things really get interesting at the 24/48 chips and above. But until productivity software (some of which for design caps at 22 threads) adds better core/thread support, it does come down to needs...

But the price/performance on the server side will really eat into Intel, bringing hundreds of millions to billions to AMD...

Sent from my SM-G900P using Tapatalk

Edit: @Mr.Koala -there was an interesting article on why Zen was changed to Ryzen. It had to do with trademark. They wanted to call the final chip Zen like they did with Vega. They changed it because of trademark. As such, the architecture code name has been left behind. Also, if using code names, it would be called Naples. Overall, it isn't that important as it is all the same architecture and this is a leak, so the names, at the moment, just act as a reference placeholder. Or that is my thoughts on it...

 

BREAKING NEWS: The most critical critic of Ryzen in NBR wants to try one

IT'Z GUNNA B GUD!!

 

I forgot to mention in my above analysis (where I spoke of the 22 thread limit), the 24 core chip can likely disable SMT. That chip may have a price of $1500-1750. It should have the six channel mem. I'm really interested in how these 8 core sections are arranged on the SOC to better understand the nature of the beast....

Sent from my SM-G900P using Tapatalk

Edit: @Papusan, @tilleroftheearth

Here is my thought. If they tie the cache of the CCX together, stacking the 8 core complexes so that you have 12 cores per side with the cache tied on those 12 cores avoiding the CCX complex latency found on the Ryzen 7 series (or at least lessening it from the 140ns to range from 40-90ns), then setting affinity of the 22-thread limited application to those 12 with SMT enabled will give you huge performance. If disabled, each thread gets its own physical core. It will have more PCIe lanes than Intel. And it should contain more bandwidth, even with the memory latency hit compared to Intel. That should explain why I wasn't to know more, including of each CCX requires using the infinity fabric imposing a latency penalty, requiring software optimization. But, at that price, you'll get double the physical cores at the price of Intel's current offering and will likely still have more cores than Intel's new offering.

I'm all about making sure people are paired with the proper hardware for their needs.

Also, AMD may be doing cores disabled like on the consumer line to flesh out the server line. Because of this, finding out which chips will utilize quad, six, and eight channel memory is something we must wait to see. It is exciting to see what will come. With the increased market share, they should meet their debt obligations for 2018 and potentially have extra for more R&D. They also intend to do Zen 2 in early 2018. This means it will not possess 7nm as that will not be ready for volume risk chips from glofo until late 2018. So, 7nm will not arrive until Zen 3 at the earliest.

On video cards, Navi will be 7nm, but is set for late 2018-2019. So, the question is whether they intend to do a 7nm Vega variant for late 2018, giving more time and pushing Navi to late 2019-2020. This depends on Nvidia, which in 2018 is rumored to use TSMC's 12nm. Right now, Pascal is on 16nm and Vega will be 14nm. So that should give a timeline.

Next year, Intel may move to the 10nm server chips, eating into the expected life of skylake-E/P/X. This would be to combat the higher core count/performance of AMD server chips. So, the next couple years look fun, but getting below 7nm will be difficult. Also, in 2020 or so, we will move to either HBM3 or DDR5 and PCIe 4.0, so that with 7nm will be a couple years, most likely...

 

I guess I assumed that Ryzen would be the name for the enthusiast/consumer, and Naples for the server market

Sent from my SM-G935T using Tapatalk

 

This is a server chip as far as I can tell. I guess we'll start to see Opteron builds once again, which is awesome no matter how you look at it.

Thanks for that bit.

Am I missing something?

 

@tilleroftheearth

 

You don't say. How did that happened?

 

I'm not 100% sure, but it seems @OverTallman smell it

 

Well your English should be better than mine, if I can deduce that from the words then both of you can

 

Ignore list. I think some forum changes should be made, so things like this are visible. It hurts discussion otherwise.

 

Well, My English sucks sometimes, So don't include me

 

So, once the 16-core, quad channel was announced at about $1000, the price/performance shifted to where he would build it without waiting for Intel's answer. I still believe the 24-core may well be the sweet spot on price/cores/channels of memory, so we stated to discuss the nuances of the server chip and platform.

As to everyone, your English is fine. I can easily understand what you guys are saying without needing to ask for clarification.

Edit: so, you may well know better than I, whether or not the Opteron lines are able to be overclocked (or what time frames they were able to be overclocked).

Sent from my SM-G900P using Tapatalk

 

I am just now trying to wrap my head around having a 8C, 16T, CPU for my home system and now you guys are all teasing about a 16C, 32T, 4CCX CPU..... POW> <POW

 

Times are changing! Now software has to catch up for better utilization of multi-threading... Considering Intel revised earnings projections, citing margin losses, they do expect to cut prices as well. This means if you need higher IPC and faster speed, it will be cheaper from Intel than it used to be, whereas, if you need processing power, AMD gives it in a brute force manner through more cores. With this primarily being server chips, AMD doesn't take as much of a performance hit as NUMA node processing is common and supported by software. So, for workstations and home users, you'll have a blend of Intel and AMD users, but on servers, AMD will really make some inroads (especially with the supported PCIe lanes which on Naples flagship (which is a 32C/64T, 8CCX beast) is more than Intel's and a projected 178GB/s bandwidth for memory).

This is kicking off a start for better multi-thread utilization in software...



Sent from my SM-G900P using Tapatalk

 

So, here is Intel's response to AMD's Naples (note-both companies flagship are 32C/64T):
http://www.tweaktown.com/news/56761/intels-next-gen-32c-64t-chip-teased-100g-omnipath/index.html

Edit: What makes this a response is that Intel previously only planned a top CPU of 28C (or at least reported until now). By creating one with 32C, when previously not planned, this is a response!

Sent from my SM-G900P using Tapatalk

 

Intel's 32-core benchmark leaked. I think Intel may be scared.

http://www.guru3d.com/news-story/32-core-intel-xeon-results-surface-in-geekbench.html

Sent from my SM-G900P using Tapatalk

Edit: for estimation purposes, the 1800X can hit around 22,000 in geekbench 4 @4.1GHz with 2400mhz dual channel ram. If you half the frequency, making it 2050MHz, you'd estimate about half the score, around 11,000 (but would have faster memory bandwidth). Multiply that by 4 because of 4 8-core complexes, y you get about 44,000. Now, the 32 core base clock is 1.6 with a boost to 2.8. If all cores can hit 2.0, then it shows some lag behind Intel, but not a lot. If 1.8GHz, then you get around 40,000. Intel's score was with the CPU running 2.1 with 2.3GHz boost. So, it would be estimated to be 10-20% faster than the 32C AMD, but using a much larger footprint on size (the monster fpga socket). There are benefits and drawbacks, but I'm waiting to see the infinity fabric versus omnipath bandwidth comparisons. Also, with faster bandwidth, infinity fabric is supposed to scale a bit with ram speed, so with 8 channels, we may see higher than what I estimated. Intel gives optane x point support and TB support. So, it will be interesting...

 

With the 8C16T 1800x at 30,000 stock you can see where Intel is sweating bullets.

 

Quick note, this was on geekbench 4, not 3, which GB 3 is where the 1800X sees over 30,000 stock. See my above edit on the last post.

Sent from my SM-G900P using Tapatalk

Edit: I've seen above 24,000 at 4050, so ram does effect the test a bit with the 1800X. So my estimate in the last post could be right to low on performance, potentially.

 

Well at 4.0 GHz

 

Fair enough. But after factoring ram speed for server chips and the official supported speed of 2667, half the speed on the CPU of the 8-core, etc., it will come down to cost and space...

Edit: But, it is closer than it has been between these two in a long while...

Sent from my SM-G900P using Tapatalk

 

wonder on the same system between Linux and windows which, if either, yields a better score. That 27800 is close to that 10% loss at 24000 on Windows, could there truly be an issue with the scheduler and CCXs compared to Linux?

 

Current Linux schedulers don't have knowledge about Ryzen's topology neither. Though depending on which scheduler is being used it might be much more adaptive.

That's one of the things people could blame AMD for in this launch which many call rushed. If Windows doesn't have scheduler optimization ready one can blame Microsoft for being lazy/uncooperative. But if Linux doesn't have it in advance it's entirely AMD's fault. (Intel and ARM always do.) The earliest Zen-aware patch should arrive in Linux 4.12 which will take a while.

 

The reason I wonder is when I think of distributed computing and servers as such I think of Linux/Unix before windows. being multi CCX is it possible the kernel treats the Ryzen core as a type of distributed computing environment? Admittedly I am not too well versed here.

 

I read they were already working on the schedule for Linux. Cannot remember when they said it would be ready.

@TANWare - you have a really good point! So there may not be much, if any difference, between them at all. That would mean Intel is in huge trouble because of the higher price!

Sent from my SM-G900P using Tapatalk

 

No.

Ryzen 7 reports all the 8 cores to be within the same package. There is no additional information for the scheduler to understand the CCX*2 structure other than running benchmarks to tune itself.

 

Just did some math. If it scaled perfectly, that 27,800 score would be 11,120 @1.6GHz. Multiplied by 4 for the 32 cores, and that hits 44,480. If it runs at 1.8GHz, that would be 50,040. So I think Linux really gives the boost, meaning Intel could be in deep **** if the Naples chip, with good ram speed, can sit at 2.0GHz!

Edit: @2GHz, perfect scaling, it would be about 55,600. Obviously some factors need controlled for, but that means a 1.9 all cores comes close to matching Intel's leak. So, let's see what all core multiplier it can do...

Sent from my SM-G900P using Tapatalk

 

User Benchmark.

http://www.userbenchmark.com/UserRun/3172664

 

The article I read on scheduler optimizations mentioned that the ones released in January our February for the FX series on Linux gave Ryzen a boost, even though Ryzen's isn't ready yet...

Sent from my SM-G900P using Tapatalk

 

Here was my first run at Firestrike.

http://www.3dmark.com/3dm/18739346

 

Great physics score! Still waiting for May. My purchase decision will be much clearer then...

Sent from my SM-G900P using Tapatalk

 

2nd try before bed. Only 1 person beating me in the ryzen 1800x 1080 TI category.

http://www.3dmark.com/fs/12069596

 

You're welcome Ryzen rumors: Perception vs. reality with AMD's new CPU - Pcworld.com

Perception: Ryzen is terrible for gaming
Reality: Not true

Perception: AMD is as good as Intel in gaming today
Reality: Partially true

Perception: An 8-core chip is a better gaming CPU if you want to be the next PewDiePie
Reality: True

Perception: It’s Windows 10’s fault
Reality: Not true (but once a suspect, always a suspect)

Perception: Reviewers who wrote negative things about Ryzen are shills for Intel
Reality: Not true (mostly)

Perception: There’s a massive motherboard shortage
Reality: Mostly true

 

Live Stream started 38 minutes ago...click timeline to go to beginning - fun commentary on Ryzen / Intel / etc

Testing Live Gameplay and Streaming with AMD Ryzen 1800X


Video Editing on the Ryzen 1800X - 30 Day Ryzen Challenge

 

How do you have time to watch all the videos? And yooo are still married?

 

The key's to the joy of life, I know I put them somewhere...

 

AMD RYZEN 5 AND 3 PREVIEWED!


10 Years of Flagship AMD Video Cards BENCHMARKED!

 

The lower end chips sound like they will be a good bang for the buck though assuming they perform decent and can overclock better than the 8 core chips. Will be interesting.

 

If you watched the right videos - or read the right articles, you'll know the 8 core 16 thread performs about the same as a 4 core 8 thread Ryzen 1700 / 1700x / 1800x with 4 cores or 1 CCX disabled in many games / programs.

If AMD allowed the Ryzen 5 / 4 / 3 to OC past 4.1ghz with their 4 Core (+ 4 thread) they would outperform the Ryzen 7 in those same games / programs, for a lot less money.

It's not gonna happen

And, for the same reason, there won't be any single CCX CPU's either, with pure latency across the same CCX vs 2 CCX's extended latency - it would also outperform the Ryzen 7.

Maybe AMD Ryzen 4 core + 4 thread single CCX only CPU's could have been shipped, but that would have continued to lock in developers to only build for those many threads.

AMD needs to walk a tight line of providing just the right balance of cost + performance while giving the world 16 thread CPU's so developers are motivated to build more multi-thread aware software.

And, provide a simple single CPU dual CCX model for developers to work with that allows their software to be X-CCX aware so it will scale up through the server class CPU's with multiple sockets using multiple CCX CPU's.

It's all good stuff, enjoy the gaming now, and stay for the high thread count software

 

It doesn't matter what AMD wants, they simply can't. The existing Ryzen 7 chips are already seeing this massive voltage wall whose foot starts in the 4.0~4.2G range, and it's highly consistent even across the different official binning levels. This limit is controlled by the combination of arch and fab, neither of which can be significantly modified at this point. Unless a huge breakthrough happens at GloFo there is nothing they can do to make a single CCX chip any faster given the same TDP, save for the difference caused by not having 4 cores taking some power at idle. Back in the days when AMD cores were eating huge amount of power with no load that might be relevant, now the difference is neglectable.

You want a fast 4 core Ryzen with no NUMA penalty, just take a Ryzen 7 and disable one CCX. Beside reduced cost a single CCX die doesn't offer much/any benefit over this.

 

And, that's why AMD won't make one, it wouldn't make financial sense, everyone would buy the single CCX 4 core CPU for less, OC to 4.1ghz, and forget about the rest.

Lower cost Ryzen 3/5 at the same or better performance of Ryzen 7 can't happen.

The Ryzen 3/5 performance can be "close enough" for the price to be worth the purchase

'Simulating' AMD Ryzen 5 1600X, 1500X Gaming Performance

 

These are my thoughts on the matter too.

 

http://hexus.net/tech/news/mainboard/103786-msi-intros-a-xmp-memory-profiles-am4-motherboards/
https://wccftech.com/msi-enabling-axmp-ryzen-motherboards/
So faster mem coming soon (originally estimated for 2 months, which would have coincided with a May update). If ready for the Ryzen 5 and 3, we may see better scores soon!

And the video card side:
https://www.pcgamesn.com/amd/amd-vega-gpu-linux
https://videocardz.com/67524/amd-linux-drivers-lists-seven-vega-10-ids
http://hothardware.com/news/amd-rx-500-series-to-be-powered-by-polaris-21-20-and-12-gpus

 

Very nice. I am currently running stable at 2666 with 2 sticks of 16gb each but as of now according to the motherboard support that is as high as you can go for 32 gb. I will definitely upgrade if 3200 to 3600 is supported in the 2x16gb.

Edit: not really willing to try out some high end ram at the moment without the boards showing support. It seems some people on different forums have been killing certain boards. Good idea to stay close to stock for now I think until the bios gets worked out on all these boards.

I know only a few boards right now are supporting overclocking BLCK from the bios but I find it strange that the MSI carbon does not have an option to change it in the bios but has a option to change it in the MSI software.

 

So, I've read about the bricking boards. Certainly working out the support is worth waiting for (my hold recommendation).

As to bclk, it has to do with software manipulation of the timing versus a chip supporting the regulation of timing. Few boards have the chip, so if it doesn't support it now, it never will.

 

Makes sense, I never even gave it a try on my board but why would they even have it for use in the software then? I am assuming if the board does not have chip and you adjusted up in the software it just would not make any changes?
Edit, never mind I just reread your sentence.

 

Some that have tried this suggest that to be the case in performance before and after changing it by software on boards without the chip.

 

Just to clarify, when I got home I checked and even though it shows you can change the BLCK in the software when you click on the icon it says "this mother does not support Base clock adjusting". Here was a Cinebench R15 run. Hwinfo was enabled after the run but the temp is around 70C while running at this clock and voltage.

 

The temperatures might not be reported correctly.
AMD stated that the programs have been reporting incorrect temps... by about 20 degrees C off

 

The temps seem to be accurate with hwinfo from what I've read at least with the new beta. I think atleast who knows for sure.

 

You'd think something like temperature measurement would be an easy thing to get right for launch day! I mean, don't temperature sensors have to conform with certain standards so readings can be taken - I'd think so, but not aware of the 'standards', seems like it should be an easy fix.