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

AMD Ryzen 5 1400 gaming performance leaks - analysed on YT - Guru3d.com

 

https://m.facebook.com/story.php?story_fbid=1415631835175966&id=165710583501437

Sent from my SM-G900P using Tapatalk

 

Much higher res / quality direct on youtube...


Related, some games show about 10fps improvement from 2133mhz to 3200mhz, while most show no difference - CCX related slowdown only affects some games - while others must need other optimizations. Interesting stuff

https://www.techpowerup.com/231585/amd-ryzen-infinity-fabric-ticks-at-memory-speed
https://www.techpowerup.com/reviews/AMD/Ryzen_Memory_Analysis/

 

Really, finding how excited Oxide is about the potential for scaling is what got me more enthused. In other quotes from interviews, they also discussed how thousands of hours have been spent optimizing for Intel over the past 5 years, but barely any for AMD. Also, this was only the start and they expect even more from future optimizations. I'm hoping the companies optimizing will shame other companies to then work more with AMD and allow scaling to n-cores (true multi-thread optimization). Time will tell. If it does work well, the HEDT I want to build will last me for a very long time!!!

Edit: To those that don't know, I don't really game much, but having a workstation that I can use for decent gaming would be nice....

 

https://www.hardocp.com/news/2017/03/29/amd_threadripper_16core_32_thread_monster_rumor

To build on this, the rumors put the 16C/32T at 3.6 base/4.0 Boost 180TDP (a hot and high powered chip). Considering that both the 6950X and the 1800X require LN2 to reach that, even the jump with skylake-E/X IPC and higher frequency would not be able to compete at 12-cores (if it is true this chip hits 2500 in cinebench).

 

Please keep this on topic, I have deleted some posts that were after where I tried to bring this back on topic.......

 

AGESA Update and mentioning next month for Ryzen:
https://www.extremetech.com/computing/246304-amd-fix-coming-fused-multiply-add-fma3-ryzen-bug
https://www.bit-tech.net/news/hardware/2017/03/31/amd-agesa-update-april/1
How to calculate infinity fiber transfer rate:
https://linustechtips.com/main/topi...-a-noticeable-impact-in-performance-on-ryzen/
https://www.techpowerup.com/231585/amd-ryzen-infinity-fabric-ticks-at-memory-speed



Now, why this is important. The 256 bit bi-direction crossbar is multiplied by the memory controller clockspeed, divided by 8 to switch from bits to bytes. We will also look at what quad-channel can provide, estimating a 70% efficiency in scaling. The AGESA update will enable higher multipliers from the current cap of 3200, so we will start there and examine up to 4000MHz.

256 bits*3200 MHz/2*1 byte/8 bits = 51200 MB/s (51GB/s)
256 bits*3600 MHz/2*1 byte/8 bits = 57600 MB/s (57.6GB/s)
256 bits*4000 MHz/2*1 byte/8 bits = 64000 MB/s (64GB/s)

Now, that is what current Ryzen owners can expect after the AGESA update, depending on the speeds supported and available (some have suggested as fast as 4500MHz by multiplier may be supported, suggesting a drastic speed up of bandwidth for the data fiber). To look at what may be available on ThreadRipper, let's use a 70% efficiency for scaling:

51200*1.7= 87040 MB/s (87GB/s)
57600*1.7= 97920 MB/s (97.9GB/s)
64000*1.7= 108800 MB/s (108.8GB/s)

Now, Intel plans on using its 100 omnipath with Skylake-E/X with hexa channel memory, which, although faster, does not provide higher data fiber speeds. Further, the IMC on Ryzen, even though running slower speeds, has been getting as high as 90% of theoretical max out of the ram for bandwidth on single rank DIMMs, much higher than Intel's which is high 60s or 70s. As such, we may see better scaling using quad channel, making the 100GB/s threshold potentially achievable at lower speeds. That we will need to see.

Overall, what this means is Ryzen will get a kick in the pants with higher speeds increasing the infinity fiber transfer speed and potentially alleviating, in part, some latency (which is even more important for ThreadRipper as it uses MCM to tie the dies together using infinity fabric). So, unless Intel releases a 14-core or 16-core, AMD will likely take the crown for HEDT, if rumors are true...

 

So, assuming your math is correct (I calculate 51200 MB/s as 50GB/s, for example), AMD will achieve parity with Intel sometime this year if a lot of if's go according to rumors? And given that the best RAM is used in the AMD setup.

Can I ask what makes you use a 70% scaling efficiency? A guess, or, is there anything to back this up for Ryzen based platforms?

 

Actually, that is using numbers from Intel quad scaling for bandwidth, which can range from 68-80% depending on ram, speed, and IMC variance. So I chose the 70% to hit the low-end, just so that it was a fairer analysis for what to expect. Considering AMD actually has more bandwidth than Intel, it isn't unreasonable to expect more, but AMD has more latency than Intel, something that cannot be forgotten.

Also, just a reminder, this is only parity on data fabric. But, it is suggested that ThreadRipper will get 100GB/s while it was either 176 or 178GB/sec from infinity fiber on 8-channel. How they got to the 178GB/s on Naples boggles me as a 100% scaling would be a speed of 2781.25MHz for the ram speed... But time will tell on that...

 

Thanks, I am learning, even if it doesn't seem to be on point.

But all this what/if and if/when guessing... it is really tiring. Is it not? Not tiring as in having a good conversation... but tiring in the sense that it is not upping the actual 'facts' one iota.

From a practical view; while there are always ways to improve any process; I don't see to many limitations with Intel's top platforms either.

Will the latency be an non-issue for AMD (I don't think so). And it has yet to be proven if getting parity or surpassing Intel is a real key to better performance. Yeah; theoretically it's better period. But are the trade offs to get there now worth it?

As you say; only time will tell.

 

Yes, but the larger point is to not write off AMD at this point, especially considering Intel and AMD platforms will be released around the same time. This means that data on real life performance will be available at the same time. To show that the theory puts them literally head to head in some ways allows for an informed decision on whether to hold off or if these releases will have any effect on potential workloads. It is more a gauge, in that manner, I suppose...

 

I picked up a Asrock Taichi, 1700X and some Gskill Trident 3200 14-14-14-35 to update my other system. Will be interesting to see how well I can OC the ram once I get it and put together the system. Seen on other forums people making it to 2933 or somewhere in that range but with more board support maybe higher.

 

You should check out *chew's thread at xtremesystems.org he is a memory master.

Sent from my SM-G935T using Tapatalk

 

Next month is when the real memory fun comes with the AGESA update!

I actually wonder, with the scaling that has been seen from 2400-3200 (or 3466, but with the PCIe kicking to half bandwidth 2.0x16 because of over 5% bclk OC), I almost wonder if the ram update was purposely held back to prevent getting a better feel for what ThreadRipper and Napples will do. If you look at the memory update release, it does not coincide with the Ryzen 3 or 5 releases. They didn't have it ready for release, but gave a two month period. What happens in late May? Computex! Considering the latency becomes less noticeable due to the transfer rate being increased (along with more tweakable settings in ram), I could see this being done by them to give an even more impressive reveal. If you look at the Ryzen 8-core, it was to drum up interest for workstation type tasks with the promise that Naples would blow that away. I can't prove it, but if performance jumps drastically after the update, I will always believe it was gamesmanship! (still will get one though)

 

That is a strange way to do business. The one thing that kept me from buying Ryzen is the memory issues, less so gaming. They lost sales to Intel and if they did this on purpose, someone needs to be terminated.

 

Well, they mentioned the ram issue with speed at release and it is way better than early silicon. This is probably why the MB companies had no interest (low transfer rate). Now, considering they said it would be fixed, gave a timeline, it was whether to trust them or not. The MB choices are the only reason I didn't (ram was second). But, as I said, 6 of 1, half dozen of the other. The real money maker is in the server market (like $16-17B versus $2B for this market). If you have a card you are trying to hide while leaving too little time for Intel to respond with tweaks or design a slightly modified chip, which we saw with the release of the 8-core chips and pricing, is it so inconceivable you sacrifice the small fish for the big market? They pulled the die-hard AMD fans PLUS some others already. In about 2 weeks, we get to see quarterly earnings report from them. Now, if this fixes the issue with ram and opens up even faster data transfers, do you think it will matter? (Please also consider that Skylake-X is built on 14nm+, NOT 14nm++, which means Kaby lake process, not coffee lake process). Also, it will be made up to you with the release of an HEDT platform (which if you are like me, you would prefer anyways, especially if the 16C chip can OC all cores to 4.0-4.1)! Am I wrong here?

Edit: A little more on the server market -
http://www.idc.com/getdoc.jsp?containerId=prUS41419716
They have it at closer to just under $14B.

This is the entire gaming market:
http://www.pcgamer.com/pc-gaming-hardware-market-breaches-30-billion-for-the-first-time/
http://wccftech.com/pc-hardware-gaming-market-30-billion-2018-jpr/
Remember, this includes more than just CPUs (Think largest capture by GPUs . . .)!

 

Ryzen Revisit: RAM OCing, Windows Updates, & EFI Updates

 

G.Skill Flare X 3200 MHz (AMD Ryzen) memory review - Guru3d.com

"Getting the Most out of your AMD Ryzen build. Tweaking the CPU and using faster memory"

 

So, I will be posting 2 articles that say the same thing as a source. This has to do with Intel adoption of HBM2 on CPU and I would love it if the people that were part of the discussion on HBM2 without GPU (so outside of AMD's APU line) would jump in. ( @Mr.Koala , @Papusan , @triturbo , @tilleroftheearth , @hmscott , @TANWare , @Rage Set )

http://www.digitaltrends.com/computing/kaby-lake-g-amd-gpu/
http://wccftech.com/intel-kaby-lake-g-hbm2-gpu-multi-die/


"The two models feature 4 cores so we are looking at a quad core design. The CPU die comes with a GT2 graphics chip and the two SKUs have TDPs of 100W and 65W. The chip doesn’t contain any on-package cache so it would be extremely weird for a normal quad core chip with GT2 graphics to pack such a high TDP. The reason it is so high is because the chips are said to pack HBM2 memory and a discrete GPU on a separate die.

The separate GPU will be connected to the main CPU die through a PCIe x8 (Gen 3.0) interface and the HBM2 memory will serve as the main on-board VRAM for the discrete level GPU. There’s no telling if the on-board GPU is an in-house Intel design or part of the alleged licensing deal between Intel and AMD that will allow Intel to use AMD’s graphics IP for their processors."

So, awhile back we discussed the use of HBM2 as either a memory that hot stacks get swapped from DDR4 to HBM2 or using HBM2 as a cache. ( @Mr.Koala I believe suggested this use). We also discussed the use of HBM2 as a cache instead. Well, one reason it was written off was because software companies would not optimize for this if just AMD did it. Well, if these chips do not contain on-package cache, it sounds like HBM2 may take that place. If more widely adopted, eventually, this could find itself in the fray with AMD chips. As such, I'd like to revisit what this could mean for AMD and potential uses of this to replace cache. Specifically, with the transfer rate tied to the IMC speeds, how, even with getting rid of the cache, the HBM2 with infinity fabric would be able to perform (taking into account the discussion above on transfer rates with infinity fabric).

Edit: Also, remember the High bandwidth cache controller known to be on Vega in the future.

Edit 2: This is even though the no cache likely refers solely to the graphics portion, not the CPU logic cores. Still, if you could entertain the notion.

 

@ajc9988

It would be really, really wierd if that sentence means what it means literally, which is the data can not rest anywhere between the register and the (V)RAM (HBM2) interface. A low level cache has a latency of a few processor circles, a RAM interface goes to 3 digits. And because there is no cache for the data from multiple fetches to sit it would be that many cycles for every operation, with the processors waiting most of the time. HBM is designed to be wide but slow (hence the increase in power efficiency), it has more latency than the alternatives, making the situation worse, and the way you can mask that latency is to cache the data.

It could be argued that the Fiji chip in the AMD Fury line doesn't benefit from HBM as much as it could because it doesn't have enough caching, especially compared to the Maxwell counterpart.

 

Here is where I wonder the impact of the High Bandwidth Cache Controller introduced on Vega. I'd also like to point out, if memory serves, that AMD is adding an L2 cache to Vega, which further gives credit to your statement. Now, if the HBCC acts to help with the latency issues and mapping, it may help behavior to be closer to cache, like an L3, but as you said, no cache would be an oddity. Now, if an HBCC type controller was engineered for use, potentially being different from AMD's solution, how could that play a role?

Sent from my SM-G900P using Tapatalk

 

On-die L2 caches have existed on GPUs for a while. The L2 on Vega is probably improved, but the existence of L2 is not new. GPU APIs (OGL/DX/OCL/CUDA) have provided explicit control of L2 behaviour since as far as I could remember.

Whatever AMD or from-AMD desgin on that Intel package is likely a auxiliary component. It's hard to imagine any cache or memory controller from AMD having direct impact on the CPU side. Intel's own 3-layer cache hierarchy plus DDR4 controller should be there as usual.

 

https://www.hwinfo.com/txt/History3264.txt

============================== Version 5.50 ==================================
Released: Apr-4-2017

- Added NVIDIA GeForce GTX 1080 Ti.
- Fixed reporting of AMD Ryzen memory timings if first memory controller is not active.
- Fixed occasional erratic sensor readouts on ASUS CROSSHAIR VI HERO.
- Added reporting of SMU Firmware Revision for AMD Ryzen.
- Enhanced sensor monitoring on ASUS SABERTOOTH 990FX R3.0.
- Fixed reporting of CPB status on AMD Ryzen.
- Simplified adjusting of sensor item settings via right-click context menu.
- Added board-level monitoring of CPU current and power on ASUS CROSSHAIR VI and PRIME X370.
- Enhanced sensor monitoring on GIGABYTE X370 series mainboards.
- Enhanced sensor monitoring on some ASRock X370 series mainboards.
- Added support of some later Corsair H110i coolers.
- Added reporting of AMD AGESA version.
- Added a new telemetry method to monitor Ryzen CPU+SoC voltage, current and power.
- Improved support of Intel Skylake-SP.
- Fixed monitoring of water sensor temperatures on ASUS MAXIMUS IX and CROSSHAIR VI series.
- Added AMD Radeon RX 570 and RX 580.
- Improved support of AMD ThreadRipper.

 

I was going to point out that the HBM2 was graphics dedicated, but I saw your edit 2 clarification. I also agree with @Mr.Koala that the HBM2 latency would cause more harm than good serving as a cache. My opinion on HBM is split. GDDR5 is near its maximum scalability and efficiency and HBM solves that, but cost plays an important role. Which is why I don't see it moving away from graphics in the near term.

 

AMD Ryzen 7/5/3 PRO confirmed
Published: 18 hours ago |
https://videocardz.com/newz/amd-ryzen-7-5-3-pro-confirmed

"AMD officially requested its Summit Ridge AM4 SOCs and Ryzen 7/5/3 series to be certified to comply with USB 3.1 Gen1 standard. However, there is one thing that was not officially discussed before, and that’s Ryzen PRO series.

This is, in fact, the first and only proof of Ryzen 7 PRO 1700, Ryzen 5 PRO 1600, Ryzen 5 1400 and Ryzen 3 PRO 1200 existence. Little is known at this point, but we can imagine that those are business oriented variants of consumer Ryzen processors. Maybe they lack overclocking support? Maybe they are cheaper? Maybe they ship with lower clocks and at a lower price? I guess we will hear more about Ryzen PRO soon."

 

From the competitor... Intel Cannon Lake Mobile Processors This Year? - Guru3.com

"So Coffee Lake is the next 14nm CPU design after Kaby Lake bringing six-core / 12-thread processors into high-end laptops and standard desktops for the first time. Then late 2017 Intel would shrink Kaby Lake down to a 10nm node, these are Cannonlake processors."

 

Still working on stability of the ram but I believe the best I can get is 3200 @ 18, 18, 18, 36 - 1.35.

http://www.userbenchmark.com/UserRun/3319810

 

how are the memory speeds effecting the Cinebench scores?

 

Ryzen Harder: Windows 10 VS Windows 7 Testing & Overclocking

 

Here is a test with the ram running at default.

 

It is good to see there is at least some boost. I guess since Cinebench is well multithread optimized there are little if any CCX latency issues.

 

@TANWare
The latency issue comes partly from OS thread scheduler's excessive load balancing behavior (which could actually be bad even without the multi-CCX design), so rendering is not immune to it. Besides, unless Cinebench is smart enough to prepare multiple copies of the same scene for different threads to read from, the 2 CCXes still need to access each other for the data.

 

I was wrong about the ram being stable at 3200. I am stable at 4.0 @ 1.375 which will be my everyday from here on out but not stable with the ram overclocked to 3200. I will have to continue tweeking to see if I can get the ram stable at 3200. Will need to do more research on additional setting for the ram in the asrock bios.

 

http://m.hexus.net/tech/news/cpu/104314-latest-aida64-boasts-radeon-rx-570580-16c32t-ryzen-support/

Sent from my SM-G900P using Tapatalk

 

I was kind of curious if Aida64 was working correctly with the new Ryzen as I was getting crashes with Prime95 blend test that I was not seeing with Aida64. Maybe the newest update will have better support.

 

Some sweet leaks in their original changelog

"The release notes of this freshly updated app detailed some interesting new hardware support additions. Unfortunately, the page has since been edited to remove the premature hardware references, but luckily the full bullet pointed list has been preserved by sites like Tweaktown. Below you will find the unedited release notes bullet point list from before FinalWire's hasty edits:

• System Stability Test / new page for unified readings
• System Stability Test / warning message when TdrDelay is too short
• improved Windows product type detection
• improved chipset information for Intel Skylake-E/EN/EP/EX
• preliminary SPD SMBus support for Intel Skylake-E/EN/EP/EX
• improved chipset information caching
• CPU TDP limit detection for AMD Ryzen 5
• CPU package temperature measurement for Intel Skylake-E/EN/EP/EX
• CPU VDD and CPU VDDNB voltage, current and power measurement for AMD Ryzen
• improved support for AMD Zen server CPUs
• preliminary support for AMD Zen 12-core and 16-core HEDT CPUs
• GPU information for AMD Radeon RX 570 (Polaris 10)
• GPU information for AMD Radeon RX 580 (Polaris 10)
• sensor support for Dell SMI of OptiPlex 5050
• fixed: motherboard specific sensor info for EVGA E75x, E76x, E77x
• fixed: CPU diode temperature measurement for AMD Ryzen (-20 Celsius offset on 1600X, 1700X, 1800X)"

Non-mobile link:

Latest AIDA64 boasts Radeon RX 570/580 & 16C/32T Ryzen support
http://hexus.net/tech/news/cpu/104314-latest-aida64-boasts-radeon-rx-570580-16c32t-ryzen-support/

 

Have you tested latest AIDA64 BETA ?
Latest Prime95 v29.1 with Ryzen support.

 

Yes there is always some inefficiency but where an app sets affinity for the threads to a core it usually is much better suited for the multiple CCX environment.

Edit; With my old C2Q, with 2 CCX, there was no difference between W7 and the enhanced scheduled of W8 in Cinebench. I found this was the case most of the time where affinity for threads was set programmatically and not handled by the Windows internal scheduler.

 

Latest prime but just downloaded the latest Aida

 

ebay sellers already shipping?

AMD Ryzen 5 1600 Processor 19MB Cache 3.6 GHz Precision Boost with Wraith Cooler
Authorized AMD Dealer - Brand New - Full Warranty
http://www.ebay.com/itm/AMD-Ryzen-5...cision-Boost-with-Wraith-Cooler-/292070787421

AMD Ryzen 5 1400 Processor with Wraith Stealth Cooler Processors
Authorized AMD Dealer - Brand New - Full Warranty
http://www.ebay.com/itm/AMD-Ryzen-5...Cooler-Processors-YD1400BBAEBOX-/292070769679

 

I refer to my previous post. Intel should get the finger out of.... More AMD to the people with nice prices... Faster we can see something from Intel The way competition is meant to be. For us

 

Agreed. I built a desktop and updated my wife's desktop for a few reasons. Seemed like a good deal, decent configuration and wanted to support somebody other than intel for a change in hopes that AMD does decent and continues to put out something that rivals intel for a while and maybe we can see some quick advancement and better prices. Little competition would be nice again.

 

AMD Announces Ryzen-Optimized Windows Power Plan, Updated Ryzen Master Tool And Total War: Wahammer Performance Patch

"AMD agrees, and since the minute Ryzen launched, it's been hard at work making sure the platform as a whole is running at its best. At launch, it was found that Ryzen didn't perform favorably in 1080p gaming compared to Intel's processors, and while part of that could be attributed to Ryzen's slightly weaker single-threaded performance, the biggest issue likely has to do with Windows itself - or at least its power profile."

 

This may be changed in the creator's update. But, either way, these issues will be mostly ironed out by the time of the HEDT release!

Sent from my SM-G900P using Tapatalk

 

Good to see, but I'm thinking this isn't an improvement over the High Performance Power Plan Profile in Windows - as in it won't offer increased performance over that one, that's my assumption. But, good to see that it seems to be offering a Power Plan that is more energy efficient yet retains high performance.

 

Spoiler

Ryzen wins

Ryzen is THE BEST CPU for Game Streaming? - $h!t Manufacturers Say Ep. 2

 

AMD Ryzen 5 1600 and 1400 Overclocking Results
HardOCP TV - Published on Apr 6, 2017
What to expect when overclocking your new Ryzen 5 CPUs! We overclock two Ryzen 5 1600 and two Ryzen 5 1400 CPUs that we purchased online. Also, AMD Reviewer's Kit Giveaway Preview for next week.


Ryzen 5 1400 vs Ryzen 7 1700 Test in 6 Games (GTX 1060)
Published on Apr 4, 2017


Ryzen 5 1400 vs i5 6400 Test in 7 Games (GTX 1060)
Published on Apr 5, 2017


Ryzen 5 1600 OVERCLOCKING Benchmarks, Cinebench, CPU-Z, Blender
Published on Apr 6, 2017
Benching the 1600 against the 5820K(6800k)


AMD RYZEN 5 1600 and GTX 1060 WITH FPS TESTED IN 4 GAMES
Published on Apr 6, 2017


AMD RYZEN 5 1600 VS INTEL i7-6700K | 4K Comparison
Published on Apr 6, 2017

 

Not directly related but AMD are taking on mobile solutions with a new Chromebook.

https://www.phoronix.com/scan.php?page=news_item&px=AMD-Chromebook-Kahlee

 

Here is Aida64 for 10 hours at 4.0, set to 1.375 at level 3. I believe according to Aida64 the CPU (Tdie) is around what the temperature should be but it is strange that under the motherboard temp the CPU reads max 47C and that is about what Aida64 recognizes it as. Little confused about the temps but going with CPU (Tdie). Also, had to go with 2933 as I cannot get 3200 stable yet, working on the timings still but anything under 18 seems to be unstable so far.

Edit: After several runs of Cinebench it is reading about the same in 3200 vs 2933, not much of a change, maybe if I could get better timings but not sure.

 

C'mon man, if I watched all that I'd be wasting nearly one hour of my life! What were the conclusions? (This is why I don't like video reviews!)