Ryzen: Strictly technical

[thepaleobiker]

It's weird because another youtuber (JayzTwoCents, I think?) just posted a video saying that after a recent BIOS update he could do it on the same board. Linus didn't specify which BIOS he was using at that moment so there's no way to know if the problem was only that or something else.

He said he got his BIOS from "someone at Asus" and that his version of the BIOS would be out within a week (so six more days from today the 25th of MArch 2017)

I watched that video last evening, quite interesting indeed.

Regards,
Vishnu

 

[Ajay]

hmm, this is just becoming another general Ryzen thread. Not strictly technical anymore...

 

[Blake_86]

The advantage of being Intel is that you get to dictate how everything works to suit your own design failures. It's not that Intel has a better IMC as much as practically all the memory out there is designed to run on Intel's DDR4 IMC - all of their sub-timings are set appropriately.

AMD's IMC has been pushed to 3600MT/s.

Interestingly, I'm having the same results with dual rank memory as with single rank - I can't get past the 2667 multiplier no matter what, but I can overclock the reference clock and reach 2800. Both kits achieve the same timings and speeds with the same voltage. I still have more testing to do, but I have noticed that the ASUS Crosshair VI Hero runs CR2 for both kits with the 2667 multiplier.

There are two things i can do, take a 7700k and GG e or still and wait, one month maibe, and see if the situation grow better.
Meanwile the 1600x will be relased.
My only regret, is we dont even know if the next mainstream intel esacore will drop onto 1151. There are good probability, but if not im going to buy a dead platform.
Such a frustrating situation

 

[looncraz]

He said he got his BIOS from "someone at Asus" and that his version of the BIOS would be out within a week (so six more days from today the 25th of MArch 2017)

I had to scour his video to see if that was the case - because I watched it and didn't hear him say it.

But you're right, he's running BIOS 1001, built on 3/17 (current official is 0902 built on 3/13)... that's only four days of work...

So, I found the BIOS and am going to give it a run

https://rog.asus.com/forum/showthread.php?91766-Crosshair-VI-Hero-new-UEFI-build-1001

EDIT:https://rog.asus.com/forum/showthread.php?91766-Crosshair-VI-Hero-new-UEFI-build-1001

Well, that helped! Now running at DDR4-2933 16-16-16-38-1T Another 50 points in CB15 (to 1750), a good 4GB/s extra memory bandwidth, shaved ~3ns off latency and may have helped in a few other areas I haven't tested... in all, a minor change, but a welcomed one.

I got further along setting it to DDR4-3200 (Vengeance LPX 3200 C16), but still no go. Will probably need to play with the timings a bit more or push the base clock if I'm ever going to get this stuff at DDR4-3200 and 1T CR. May have to use the 2T BIOS to hit that, though.

 

[guskline]

looncraz, I was likewise able to upgrade the BIOS in my MSI X370 Titanium from 1.1 release to 1.31 and my memory now runs at 2933 BUT 3200 is a no go.

I think AMD has work to do with the MB vendors to improve the BIOS.

 

[DrMrLordX]

hmm, this is just becoming another general Ryzen thread. Not strictly technical anymore...

Okay. What do you want to see more of in this thread?

 

[Dresdenboy]

I cant see why a pair of ram should run slower, a lot slower, ther their spec. The same dual rank kit is fully compatible whit intel and also have no problem to run higher!
This is a duble fault by amd, creating a heavy ram dependant architecture whit a crappy ram compatibility.
I realy would to buy amd, but if it is true that is frustrating to go intel alwais and alwais quad core, it also frustrating to buy an 8 core and have to chose from perf and 32 gb of ram

What spec are you talking of? JEDEC compatible SPD timings or Intel XMP data (used for marketing 2133 mem as 3200 etc.)?

 

[coercitiv]

There are two things i can do, take a 7700k and GG e or still and wait, one month maibe, and see if the situation grow better.
Meanwile the 1600x will be relased.
My only regret, is we dont even know if the next mainstream intel esacore will drop onto 1151. There are good probability, but if not im going to buy a dead platform.
Such a frustrating situation

You now have more choice than you had in many years, yet regret is your only reaction. The irony is, whatever choice you were to make it would turn out just fine.

 

[SuperFist]

Well, I finally got Windows 7 Pro installed on my Samsung EVO NVMe SSD but now it hangs on the Starting Windows reboot logo. The logo still animates, it just doesn't proceed any further. I have to force a reboot, go into the BIOS and select the drive under Boot Override to get it to boot up properly.

 

[looncraz]

Well, I finally got Windows 7 Pro installed on my Samsung EVO NVMe SSD but now it hangs on the Starting Windows reboot logo. The logo still animates, it just doesn't proceed any further. I have to force a reboot, go into the BIOS and select the drive under Boot Override to get it to boot up properly.

It's just not worth fighting Windows 7, frankly. Install 10... then neuter it until it's Windows 7 with an updated kernel.

 

[Blake_86]

You now have more choice than you had in many years, yet regret is your only reaction. The irony is, whatever choice you were to make it would turn out just fine.

Exacly!

 

[Zucker2k]

The advantage of being Intel is that you get to dictate how everything works to suit your own design failures. It's not that Intel has a better IMC as much as practically all the memory out there is designed to run on Intel's DDR4 IMC - all of their sub-timings are set appropriately.

AMD's IMC has been pushed to 3600MT/s.

Interestingly, I'm having the same results with dual rank memory as with single rank - I can't get past the 2667 multiplier no matter what, but I can overclock the reference clock and reach 2800. Both kits achieve the same timings and speeds with the same voltage. I still have more testing to do, but I have noticed that the ASUS Crosshair VI Hero runs CR2 for both kits with the 2667 multiplier.

So JEDEC specs mean nothing? See Dresdenboy's post above. SMH.

 

[Kromaatikse]

JEDEC is pretty conservative.

Ryzen really likes high memory clocks due to that tie-in with Infinity Fabric. It's not that it uses memory bandwidth a lot, but that it relies on Infinity Fabric bandwidth to handle inter-CCX stuff.

 

[unseenmorbidity]

While AMD does not warrant overclocked systems, we can advise that 1.35-1.425V is a practical vcore range for OCed units.



Our official memory support:
-Dual Channel/Dual Rank/4 DIMM: 1866
-Dual Channel/Single Rank/4 DIMM: 2133
-Dual Channel/Dual Rank/2 DIMM: 2400
-Dual Channel/Single Rank/2 DIMM: 2667

In other words: Ryzen will automatically run at the rated speed with modules that use JEDEC SPD timings. Additionally, motherboard vendors like Asus already support "overclocked" speeds up to DDR4-3200. However, recall that XMP is exclusively an Intel memory technology designed exclusively for Intel memory controllers. Users may have to program the suggested timings of an XMP-based module manually on Ryzen.



Best regards,

Pugal.

AMD Global Customer Care

 

[looncraz]

So JEDEC specs mean nothing? See Dresdenboy's post above. SMH.

AMD has no problem working with most any JEDEC rated speeds. The problem is when it comes to OC speeds... which is anything over DDR4-2400, IIRC.

 

[Oleyska]

AMD has no problem working with most any JEDEC rated speeds. The problem is when it comes to OC speeds... which is anything over DDR4-2400, IIRC.

My primitive testing, which is for me a confirmation of results here and elsewhere show that 2666 mhz is a big deal.
But, dualrank 2666 mhz 2x 16gb boots fine on 1.2 V with latest bios for B350M-A which confirms that EFI/agesa updates help out.
16gb 2666mhz Corsair LPX dimms
"
PRIME B350M-A BIOS 0509
1.Improve system performance.
2.Make CPU temperature more precise."

 

[Elixer]

This has the problem that each core is running both reading and writing threads, so some of the results are dramatically high as the two heavy threads are switched back and forth, naturally, so I will be focusing on addressing that minor issue now that the basic test is working.

When you have time, can you compile & run https://github.com/gregs1104/stream-scaling (which uses https://www.cs.virginia.edu/stream/ as the main benchmark), and pastebin the results someplace? Must be done on a 64-bit linux system.

 

[Blake_86]

What spec are you talking of? JEDEC compatible SPD timings or Intel XMP data (used for marketing 2133 mem as 3200 etc.)?

What do you mean?

 

[Dresdenboy]

What do you mean?

If you see some DDR4-3200 RAM, then this clock is not its JEDEC specced speed, but that stored in the Intel XMP data. JEDEC speed is slower (likely exactly that what most Ryzen users get w/o changing the mem multipliers and timings).

 

[CrazyElf]

Canard PC says 16 cores @ 180W TDP, not 32 cores.

Here are the Tweets:

https://twitter.com/CPCHardware/status/843099198799187968

Il y a bien des Ryzen 16C/32T axés HEDT prévus sur X399 dans 4/6 mois. Clocks ~2.4/2.8G. 2 dies MCM. 4 chan DDR4. Socket LGA SP3r2. ~150W.
​

16 core is 150W, not 180W.

https://twitter.com/CPCHardware/status/844618089618722816

2nd Gen B1 ES for Ryzen 16C/32T on X399 platform incoming. Speed (and TDP) bumped to 3.1/3.6 GHz @ 180W on these samples. Still not QS.
​

32 core is 180W.

That's why I think the final silicon might be 3.2 or 3.3 GHz. Engineering Samples are always a bit slower.

The advantage of being Intel is that you get to dictate how everything works to suit your own design failures. It's not that Intel has a better IMC as much as practically all the memory out there is designed to run on Intel's DDR4 IMC - all of their sub-timings are set appropriately.

AMD's IMC has been pushed to 3600MT/s.

Interestingly, I'm having the same results with dual rank memory as with single rank - I can't get past the 2667 multiplier no matter what, but I can overclock the reference clock and reach 2800. Both kits achieve the same timings and speeds with the same voltage. I still have more testing to do, but I have noticed that the ASUS Crosshair VI Hero runs CR2 for both kits with the 2667 multiplier.

We really need those AMD RAM timings unlocked so we can get to the bottom of this. Probably not until May or June before we get it. Maybe a few more weeks after before G.Skill et al release AMD Ryzen RAM.

 

[imported_jjj]

What do you mean?

JEDEC is the standardization body that defines the specs.What you buy in retail is usually DDR4 2133 overclocked and designed for Intel's platform
The official support Intel offers is up to 2400MHz. AMD beats that by going to 2666MHz for single rank.
Everything above that is a memory overclock and technically voids the warranty.

 

[looncraz]

When you have time, can you compile & run https://github.com/gregs1104/stream-scaling (which uses https://www.cs.virginia.edu/stream/ as the main benchmark), and pastebin the results someplace? Must be done on a 64-bit linux system.

I'll try to do it in a couple of hours, I'll be natively running Linux Mint 18.1 while trying to figure out what I screwed up in my CCX benchmark utility - it will be a nice reprieve

UPDATE:

Results

Ryzen 7 1700X @ 3.9GHz, 16GB (2x8GB) DDR4-2933 16-16-16-38 1T.
Linux Mint 18.1, kernel 4.4.0-53-generic

 

[Blake_86]

Here are the Tweets:

https://twitter.com/CPCHardware/status/843099198799187968

Il y a bien des Ryzen 16C/32T axés HEDT prévus sur X399 dans 4/6 mois. Clocks ~2.4/2.8G. 2 dies MCM. 4 chan DDR4. Socket LGA SP3r2. ~150W.
​

16 core is 150W, not 180W.

https://twitter.com/CPCHardware/status/844618089618722816

2nd Gen B1 ES for Ryzen 16C/32T on X399 platform incoming. Speed (and TDP) bumped to 3.1/3.6 GHz @ 180W on these samples. Still not QS.
​

32 core is 180W.

That's why I think the final silicon might be 3.2 or 3.3 GHz. Engineering Samples are always a bit slower.







We really need those AMD RAM timings unlocked so we can get to the bottom of this. Probably not until May or June before we get it. Maybe a few more weeks after before G.Skill et al release AMD Ryzen RAM.

And we should fire the kit we already have? I mean, my trident z have indeed xmp profile, but if amd allow us to manualy set the subtiming there is no reason why they will not run at full speed. Right?

 

[looncraz]

And we should fire the kit we already have? I mean, my trident z have indeed xmp profile, but if amd allow us to manualy set the subtiming there is no reason why they will not run at full speed. Right?

In theory, that should be the case. This is probably what they have planned for the May update.

Ryzen has a RAM-learning implementation that seems to cause more trouble than anything (five power cycles, two soft resets, and a post-BIOS soft reset is a pretty insane process to have to go through every time you change a memory timing or frequency that the system finds not to be "stable").

Ryzen has crazy-sensitive error detection - I turned some of that off and hit 4.15GHz before having stability issues at 1.395V (highest CPU-z I did during that run was 4.05Ghz). I ran the DDR4-2400 divider because it was unstable at anything about 4GHz with 2667.

I have a feeling we are being limited by the data fabric more than the cores or process. That'd be a really good sign that clocks will improve with microcode updates and with Zen ver2. We're basically adding more voltage to prevent relay errors.

 

[looncraz]

On a tangentially related note: I think I underestimated the demand for Sandy Bridge i7s.

I put my CPU on eBay and it sold before I even began listing the motherboard (~30 seconds, I kid you not!). I got the motherboard up a few minutes later and began to list the memory - motherboard sold before posting (~5 minutes). Then I posted the memory and had an offer within the first minute... for half the price I asked, LOL! Some bartering failed to make a deal. A few minutes later another half-priced offer came in, and I countered with a minor reduction in price - no sale. Another five minutes and it sold at full price (which I reduced to $35/kit - both 8GB kits were purchased for $70).

I've decided to just put all that money into a Vega GPU, if it's reasonably priced. All of that plus the sale of the R9 Fury should make it a cheap move since I have room left in my computer upgrade budget as I ended up with the 1700X instead of the 1800X as intended (because I, wrongly, assumed CPUs would be harder to get than the motherboards...). That leaves me $500 for a GPU, which I didn't expect to upgrade.. but this Fury dumps a lot of heat into my case and blocks some airflow from my front and bottom intakes.

AMD getting all my money... better than the blue or green men getting it, I suppose.