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

Point being that even if those Intel CB scores are real, they could have been taken with a sample that was possibly overclocked and compared it to a stock TR.
And even if the Intel 18 c is not overclocked, overclocking the TR to 4.25 GhZ apparently gets you within spitting distance of Intel's 18 core CPU (scoring just 800 points less - which is less than 10% difference) - equalizing for clock speeds between what the CPU's could handle, though clocking TR to 4.2 as well would have been better and more accurate.

P.S. That's actually odd. How can TR with 16 cores at 4.25 GhZ reach Intel's 18 cores to a spitting distance that should operate on 4.2 GhZ boost (presumably across all cores?)?

It seems a bit... off... and flies in the face of Intel's alleged superiority.
Shouldn't the extra 2 cores give Intel a distinct advantage that goes beyond 10%?

 

It isn't 800 points. I am at 3380-3406 on 3.85GHz. Many at 4.0 are above 3500. If you know what you are doing and tune to not throttle, you would see 600-700 point difference. What are you using for comparison?

Edit: also 1/8th, or 12.5%, of the score, should be 2 extra cores.

 

I'm using the video hmscott provided

 

Fair enough. But, considering the two extra cores, you adjust to 16 cores and it would get 3733 for CB15, if the score was 4200 with the 18-core (it was slightly lower than that). You then compare with many getting over 3500, some getting over 3700 in CB15 (I've seen one or two, and the clocks, etc. Had indicia of real results). So, it is better, just not by much.

 

Ah... I see.
Still, it gives an indication that TR is better value for the money.
Because even with those CB scores, the differential is what... 12.5% between 16 and 18 core scores.

Doesn't that mean that IPC-wise, AMD is actually on par with Intel and the only advantage for intel is the higher single core clocks along with industry software optimizations (that take advantage of virtually none of the features available for Ryzen/TR).
Makes me wonder what might happen if/when software optimizations arrive for Ryzen/TR from the industry at large.

 

I went to [email protected] and saw CB R15=3523 best score. This is with the 3466 memory so I wonder if the silicon heat and work builds too high at getting those scores at over that point. So in other words SL comes to play but what is important is cooling and how hard the CPU is working in order to get that high score.

 

I was seeing 40-60 points per 50mhz jumps, so your score is in the right area. But I've seen a couple hit over 3700 @4.2ghz range...

 

Another issue, is with the Zeinith Board on the video I was with HW monitor where he claimed it was set to 1.425 that the voltage at times jumped over 1.5. With the Taichi we do not get those extreme jumps to support the higher clocks. I am not setting my board to 1.5 any time soon is for sure.

 

There are many things about the taichi I like. Setting the voltage and it staying is one of them.

Edit: skipped past the intro. I beat him in wprime with 3.85 or 3.9GHz. He's forcing speed and throttling somewhere, or the firmware isn't right. He isn't one I would take bench advice from, personally.

 

I am mainly interested is what is safe voltage. I think overall I am just going to set 4.0 GHz and call it be a day.

 

AMD recommended 1.35V daily driver, not over 1.4 or you'll get degradation. So, it comes down to a heat/voltage estimate on degradation.

 

If they recommend 1.35 are they doing so knowing the chips are being driven to 1.5 plus as on that Zenith board at times? It seems most boards have that behavior in overclocking. I am sure when they were trying it out it did the same thing then too. Maybe spikes at up to 1.5 are OK then but it would be nice to have an answer on that from AMD.

EDit; I set [email protected] and HW monitor shows a max of CPU VCORE as 1.376 but running R15 it goes to 3.34, So I guess that is fine. I am trying 1.375 and under load it goes to 1.328 in HW monitor, so I am guessing this is better.

 

Could you undervolt on stock clocks as low as it will go and see the temps you are getting along with power draw?

 

Do you mean at 3.4 GHz fixed? I know with auto the voltages @ stock go insanely low even under full load.

 

Yup... just curious if its possible to undervolt TR to the point of increasing its efficiency even more.

 

My max OC is 4.175. If I try 4.2, no matter the voltage, the system crashes. My most recent stable clock is 4.125 @1.38 but I am not going to use that as my daily. I'll stick to my 4.0/4.025 for everyday use.

 

Same version Wprime? Can’t compare 2.10 with 1.55

What’s the correct IPC difference between Intel vs. Ryzen? 12.5% more cores and IPC on top should push numbers a bit up.

 

I used 1.55, meant for submission to HWBot. So, I don't know if comparable. I use what is meant for competition, you know that (if he is benching and bragging, like here, he should be using the right version also).

IPC is supposedly just under 10%. But, when comparing multi threaded scores, you have to look at latency and SMT vs HT, which SMT is much faster, closing the IPC gap in part. This is why I did the 18 to 16 core discount by multiplying by 8 the score achieved, then dividing by 9 (it is 12.5% more cores compared to a 16 core, but is the opposite ratio to scale down the score). But his 4.2 on air on a Noctua is less than a custom loop 4.2 score in many accounts I've seen (I'm saying he likely has throttle issues). That is why you cannot accurately compare them. We do not know the skill level of the person benching the 18-core either, which is problematic.

 

Yeah, That he don't know what version being used on hwbot, ain't exactly promising

 

It seems with the Zenith using those higher voltages where needed from the CPU itself that most are obtaining overclocks at lower set voltages. The Taichi may not OC as high but I know the voltages are always at the safe level under load etc. under 1.350 with 4.0 GHz. I am not sure how others feel with HWMonitor reporting over 1.5 on the VCore at times. I wouldn't be able to allow it.

I found a bug here. When overclocked in W10 Pro I get R15=3462 after boot. If it goes to sleep the OC shows all good but performance drops in the 2900's until reboot. If at CPU stock I get R15 = 3025 even after awoke from a sleep.

 

The taichi firmware also gives higher results at lower multipliers. So don't worry about that as much as not getting the higher OC.

For sleep, it could be os, could be a microcode bug, firmware, etc. But definitely good to know. I'll pass it along to someone working on helping with the taichi firmware in case.

 

AMD might be replacing the Polaris line of GPU's with Vega's smaller counterparts:
https://www.guru3d.com/news-story/amd-might-replace-rx-500-cards-with-rx-vega-32-and-28.html

And it would seem that they all might come with HBM2.

Makes me wonder if the Raven Ridge APU will also have HBM.

 

The FIRST custom AMD RX VEGA 64 Card is here... do custom PCBs help?

 

The question is, do you blame the AIB suppliers or AMD for what looks to be mediocre custom GPU's? Some could argue that Vega requires more than off the shelf board designs (as in the case of the Asus RX Vega 64 Strix) that are shared with Nvidia's GPUs. With the likes of MSI (going by rumors) supposedly skipping making custom RX Vega 64 cards, it does appear that AMD could be in some trouble in regards to Vega 64. I still want to get a Vega 64 and put it under custom water, but I won't spend above $500 for it.

 

The automatic performance tuning still seems to be the missing part of AMD RX Vega, even watching Jay fiddle with settings, and discovering improvements for seemingly counterintuitive settings - reducing voltage to get higher frequency stability - and I don't think Jay - or anyone else - is anywhere near expert enough yet in getting the best performance from RX Vega tuning.

Jay also skipped the other benefits of an AIB card, quiet, better cooling, and additional features - he did do a short run down, but no real comparison was done as we've seen in the past between AIB and Reference designs.

Also toward the end he said the AIB board tuned matched the out of the box AMD RX Vega 64 LE, and if he approached it all again I think he would get there sooner - it shouldn't have taken 3 whole days - I think he had a lot of gaps in attention during that time

I don't think Jay did the card justice, but I don't think that's his fault, there is a lot of ramp up time to understand the options for tuning and their effects, you just gotta burn the time trying things out to get there. By the time he started getting close to some understanding of the tuning he probably exceeded his time commitment and didn't have time for the rest of the review niceties.

 

The Portable AMD Build!! September PC of the Month

 

How “ZEN” is Made

AMD Ryzen™ PRO Launch Event Highlights

AMD's Vision for the Future of Technology

 

Finally figured out why THREADRIPPER has so many PCIe lanes (en)
8x NVME RAID - 27GB/sec READ! - Asus BIOS Update / Free RAID driver released by AMD!
Benchmark results, starts @ 8:50

Update: der8auer has pulled video's before, he's probably not happy with it's incomplete nature. Not really stable at tested speed, not reproduceable with Crystalmark / ? - but repeatable with iometer, and he only shows READ performance not WRITE performance. I was surprised he posted it

Hopefully he will reshoot it and put out a better video.

Watch his videos for new ones:
https://www.youtube.com/user/der8auer/videos

Guru3d posted some shots from the video:

 

Can't see the video... must have been removed or something.
Link to another one?

 

der8auer has pulled video's before, he's probably not happy with it's incomplete nature. Not really stable at tested speed, not reproduceable with Crystalmark / ? - but repeatable with iometer, and he only shows READ performance not WRITE performance.

I was surprised he posted it

Hopefully he will reshoot and put out a better video.

Watch his videos for new ones:
https://www.youtube.com/user/der8auer/videos

 

ENERMAX LIQTECH TR4

ENERMAX LIQTECH TR4 Liquid Cooler Installation Demo (For AMD Threadripper CPUs)

ENERMAX LIQTECH TR4 AIO Liquid Cooler Installation Guide

Enermax seems to really like this footage, they've posted it in their Global site 3x

Product page:
http://www.enermax.com/home.php?fn=eng/product_a1&lv0=109&lv1=118

 

It appears as though Vega cards are somewhat back in stock at retail prices.

https://www.newegg.com/Product/ComboBundleDetails.aspx?ItemList=Combo.3623495

Is this news or am I out of the loop?

 

News, and it's a great improvement, almost back to MSRP of $399 + $99 (games) = $499.99

"MSI Radeon RX Vega 56 Radeon RX Vega 56 8G Video Card, AMD Gift with Purchase- Wolfenstein II: The New Colossus and PREY
In stock. Limit 1 per customer.

was: $629.98
$499.99
save: $129.99

I love how they are spinning the return to MSRP as "saving $".

 

#173 - AMD Threadripper 1950X Quick look and Overclocking to 4.1Ghz

 

It looks like Zen and Vega refreshes on 12nm will be arriving next year.
Same for Raven Ridge (the Ryzen APU with Vega 11).
It seems like Raven Ridge might be made on 12nm LP process as well... which of course we don't know, but it coincides at least with Ryzen/Vega refresh being released next year.

https://hothardware.com/news/leaked-amd-roadmap-zen-2-matisse-12nm-finfet-cpu-2019

And it also looks like Zen 2 will be arriving in 2019.
So, it looks like its been pushed (as I was under the impression it was due for a release next year).

 

It isn't pushed. 7nm tape outs on some chips Q4 this year. It takes 9-15 months from tape out. That put Q3 at the earliest. GF has said volume manufacture of 7nm in the second half. You have to have a couple months minimum to build inventory. Hence why I have been saying Q4-Q1 2019.

 

New 400 chipsets? What happened to backwards compatibility?

https://seekingalpha.com/news/3297830-digitimes-amd-launching-new-ryzen-processors-february

 

New chipset doesn't mean that it won't be compatible with old chips or old chipset cannot use new chips. It means you can use the chipset as shorthand for features changes found on the board. It also may be to help board manufacturers sell more (although mitigated if compatible with different generations of chips), which will increase support from board manufacturers.

 

These are the APU's, using Ryzen + Vega, which require video output(s) to access the GPU, and the 3xx motherboard chipsets don't have them.

AMD Ryzen Pinnacle Ridge 12nm FinFET Processors Allegedly Arriving In February 2018
https://hothardware.com/news/amd-ryzen-pinnacle-ridge-processors-arriving-february-2018

 

It does not look like 12nm will make it, at lest at first, to Epyc or TR. Not that I am looking to drop another $1,000 this soon. It would have been nice to see some added pressure to Intel's 7980XE offering.

 

GIGABYTE Has No Plans to Release a Custom Radeon RX Vega 64 - Techpowerup.com

"Due to the inconsistency in quality of chips that AMD are providing, AIB partners are having a difficult time establishing a standard GPU frequency for their factory-overclocked cards. Furthermore, temperature reporting is broken. The actual GPU temperature is different from the temperature reported by the GPU which can become a big problem for stability in the long run."

 

Sounds like excuses for being late, then giving up, hopefully more level heads prevail at Gigabyte and get their project back on track. I wonder if this affects both Gigabyte and Aorus?

I mean, if AMD can do it, the others should be able to make it work too

 

The OEM's need to wisen up to the fact that they need to undervolt Vega as low as it will go and sell undervolted versions of both stock and overclocked cards (with OC-ed ones largely focusing on smaller core and larger HBM overclocking).
OEM versions are going to come with higher quality silicon, which means, most if not all of them could easily accomplish the following:
https://translate.google.de/translate?sl=de&tl=en&js=y&prev=_t&hl=de&ie=UTF-8&u=https://www.hardwareluxx.de/index.php/artikel/hardware/grafikkarten/44084-amd-radeon-rx-vega-56-und-vega-64-im-undervolting-test.html&edit-text=

I suspect that Vega 56 achieves better undervolting because it also has less stream processors.

Perhaps OEM's could work with AMD to disable a certain amount of stream processors on both Vega's which would likely not affect performance and are seldom used in games - or give users the ability to disable stream processors.

I am sick to death seeing overclocked GPU's from OEM's that are throttling and not achieving boost speeds and consuming ridiculous amounts of power.
They should know limitations of 14nm process AMD is using... meaning they can't overclock to too high levels for minimal gains... especially on the core.
The biggest changes we saw performance wise for a minimal power increase were on HBM overclocking.
If they can clock Vega 56 HBM to 1100 MhZ and Vega 64 HBM on say 1200 or 1300 MhZ and slightly bump up the clock speeds or keep them at their designated stock rates with much needed undervolt, I think they can achieve better results... but they would need to customize the bios so enable such speeds on HBM.

The reference GPU's are running into throttling issues due to being overvolted and are being compared to Pascal gpu's that have no issues with voltages because Nvidia set their base voltage to lower levels and also has auto-voltage tuning - which means, Pascal is achieving it's advertised boost clocks easily enough, while AMD does not (because a no throttle Vega 56 is faster than 1070, and no throttle Vega 64 is faster than 1080).

Out of curiosity... has anyone tried to artificially disable certain % of stream processor units in a GPU?
We see that AMD has a lot more of them vs Nvidia's cuda cores for example... hence higher compute performance, but this compute potential is not used or optimized by devs in games at all.

 

You mean make new 970's

 

No... the reduction in stream processor counts would be there to reduce overall power draw.
1080ti for example has 3584 cuda cores for example... Vega 64 has 4096 stream processors, while Vega 56 has 3583 stream processors (and hence the better undervolting most likely because there's less of a power demand).

When Vega 56 and 64 were examined at same clocks on both core and HBM, they both produced equal performance in games despite a difference in stream processor count, but Vega 56 still drew less power (probably due to reduced amount of stream processors).

I know AMD can't really produce a lot of customized GPU's that lacks specific hardware as its likely cost prohibitive.
And Vega 56 when flashed with Vega 64 bios does not unlock more stream processors on 56.

So, perhaps the OEM's could work with AMD to further optimize power consumption by disabling the amount of active stream processors on Vega 56 and 64 by say 500 on each (either via BIOS or through drivers) and undervolt the GPU's in the process (on the core and HBM) while overclocking the core slightly and HBM by a lot.

I think this would alleviate a lot of Vega's power issues in the current 14nm process even with voltages set to 1.2 maximum for gaming hardware.
The 12nm LP process however, could demonstrate to be a lot better though as it might be more akin to what TSMC has and suited for higher clock speeds at lower power draw?

 

It becomes so when AMD is trying to steal the clothes of Nvidia. I do not like that Nvidia or Amd starts fiddling/tampering with this. We've had 970... Now max-Qrippled. I think Enough is enough. Maybe low production of Amd graphics means variable silicon quality. Aka one of the reason the AIB partners are having a difficult time establishing a standard GPU frequency for their factory-overclocked cards. Something is up.

For more information about custom designs of the Radeon RX Vega56-Tomshardware.de

 

That is really unfortunate. Inconsistency is never a good thing. AMD cannot afford that kind of damage, as they are still in somewhat of a fragile state in terms of reputation and the potential to be haunted by their checkered past. Stability and a steady and consistent product needs to be one of their highest priorities considering it has been their Achilles Heel for so many years.

 

But, isn't Vega supply issue down to HBM 2 production (which isn't AMD's fault) but the fault of a company supplying HBM2?

 

Or AMD still argues about pay lower in prices for HBM2 memory Put what's fit.

 

I think given the info that Papusan linked, this would be very costly or impossible for AIB's to accomplish. Papusan's link said " Due to the inconsistency in quality of chips that AMD are providing, AIB partners are having a difficult time establishing a standard GPU frequency for their factory-overclocked cards.". This means that individual chips are showing a wide range of stability or instability with certain voltages & frequencies - so they're having a hard time standardising a voltage & Mhz - this by definition would mean that undervolting would be nigh impossible, instead they would need to add voltage if anything to overcome the inconsistent quality of silicon coming from AMD.