FMA3 bug discovered in (Ry)Zen

[zMeul]

source: https://www.heise.de/newsticker/meldung/Bug-in-AMD-Ryzen-Kompletter-Systemabsturz-bei-manchen-FMA3-Anwendungen-3641409.html

http://forum.hwbot.org/showthread.php?t=167605

 

 

HWBot users have discovered a reproductible hardware bug that seems to affect FMA3 instructions

the bug appears to show up when it gets to

Quote

Single-Precision - 128-bit FMA3 - Fused Multiply Add:

 

if this is a instruction set but and not something else, this could be potentially patched via a BIOS micro-code update 

so far, this bug has been confirmed on:

Quote

• 1800X + Asus Prime B350M-A (BIOS 0502)
• 1700 + Asus Prime B350M-A (BIOS ???)
• 1700 + Asus Crosshair VI Hero
• 1700 + Asus Crosshair VI Hero (BIOS 5803) (two sets of memory G.Skill + Kingston - also fails with overvolted SOC)
• 1800X + Asus Crosshair VI Hero (Windows 7) - Once pass, mostly failures.

Heise.DE was able to reproduce the bug with Ryzen 7 1700X on the MSI X370 XPower Gaming Titanium and a Ryzen 7 1700 on the Asus Crosshair VI Hero

 

instruction set bugs are not uncommon, as Intel's own Skylake launch was troubled with AVX bug that was discovered couple of months after

 

at this time, AMD did not respond to the inquiry made by Heise.DE

 

---

 

update: 16th of March

 

more info on this from PCPer https://www.youtube.com/watch?v=dDrOQrjJTSs

it seems that the FMA3 set isn't actually bugged, but it seems it doesn't have enough power delivered to it as 25% of the people who OCed their CPU were able to ran the test successfully

 

sources inside AMD have said that it's a known issue and a BIOS update is indeed planned

Edited March 16, 2017 by zMeul

[vorticalbox]

So a bug was found in a benchmark software that 

Quote

 uses highly-optimized code with modern command set extensions such as SSE and FMA3 (Fused Multiply-Add 3)

So outside these types of programs literally not a problem.

[vanished]

2 minutes ago, vorticalbox said:

So a bug was found in a benchmark software that 

So outside these types of programs literally not a problem.

Well who's to say other programs won't also need/use that thing?

[zMeul]

3 minutes ago, vorticalbox said:

So a bug was found in a benchmark software that 

So outside these types of programs literally not a problem.

the bug is not in the software, it's in the CPU's FMA3 instruction set - not the same thing 

[samcool55]

Well as far as i can work out they dropped FMA4 so i don't think that many programs use it...

Can be wrong tho, i have no idea how this stuff works.

[zMeul]

Just now, samcool55 said:

Well as far as i can work out they dropped FMA4 so i don't think that many programs use it...

Can be wrong tho, i have no idea how this stuff works.

as far as I can tell, both AVX2 and FMA3 is used by ZFS file system - so, that's kinda' big in Linux world

[vorticalbox]

8 minutes ago, zMeul said:

the bug is not in the software, it's in the CPU's FMA3 instruction set - not the same thing 

I understand that but if any given software doesn't end up calling that function then no problems. If a lot of programs did use this function there would be a shit storm right now, rather than some benchmarker that happened to find it. 

I'm not saying this isn't a massive issue, it is, it's just not going to cause that my problems.

[zMeul]

@leadeaterregarding this: https://linustechtips.com/main/topic/746289-amd-responds-to-1080p-gaming-tests-on-ryzen-supports-ecc-ram-win-10-smt-bug/?do=findComment&comment=9457973

I remember what bug was about, it was about TSX implementation bug discovered in Haswell and Broadwell - Intel had to disable the set in the 1st two generations of Haswell and 1st generation of Broadwell

 

[zMeul]

3 minutes ago, vorticalbox said:

I understand that but if any given software doesn't end up calling that function then no problems. If a lot of programs did use this function there would be a shit storm right now, rather than some benchmarker that happened to find it. 

I'm not saying this isn't a massive issue, it is, it's just not going to cause that my problems.

that issue is that you called it a SW bug, it was quite incorrect

whether the software you use actually needs FMA3, or not, doesn't mean the issue is not there

[zMeul]

2 minutes ago, Jorgen297 said:

Would it be possible to fix this with a microcode update? Or is it embedded in the actual circuitry in the CPU? 

depends on the gravity of the bug,

it is possible to fix it via microcode update - Intel did the same with Skylake's AVX bug

but, if the issue is way to complex, AMD might be forced to disable the entire set - Intel did this for TSX set on two generations of Haswell and 1st gen of Broadwell

[vanished]

Wow, Ryzen has sure had a rocky launch... one thing after another.  Seems new problems are cropping up faster than they can fix the ones we already knew about  

[zMeul]

2 minutes ago, Ryan_Vickers said:

Wow, Ryzen has sure had a rocky launch... one thing after another.  Seems new problems are cropping up faster than they can fix the ones we already knew about  

that's why you have engineering samples and test the fuck out of them before going full production

and after production you do fucking validations to confirm the fucker isn't screwing around - especially since AMD aimed to sell this as compute CPU replacement/competitor for Broadwell-E

[vanished]

1 minute ago, zMeul said:

that's why you have engineering samples and test the fuck out of them before going full production

and after production you do fucking validations to confirm the fucker isn't screwing around - especially since AMD aimed to sell this as compute CPU replacement/competitor for Broadwell-E

It's unfortunate though... they really need this to go well.  We all do, even if you don't buy from them - more competition only helps the market.

[zMeul]

1 minute ago, Ryan_Vickers said:

It's unfortunate though... they really need this to go well.  We all do, even if you don't buy from them - more competition only helps the market.

they only have themselves to blame tho

I wonder who will AMD point the finger at this time  those pesky compilers .. sure

[The Sloth]

not bad as you might think, bugs happens to new platforms, hope optimizations and bugs gets streamed out 

 

[porina]

17 minutes ago, samcool55 said:

Well as far as i can work out they dropped FMA4 so i don't think that many programs use it...

Can be wrong tho, i have no idea how this stuff works.

There were two problems with FMA4: it was only on AMD processors so already there is a limited use case, and the only person I know who tried to use it (George Woltman, of Prime95 fame) couldn't extract any performance boost from using it over not using it. 

 

FMA3 was implemented by Intel in Haswell, and it did provide a huge performance boost and gained huge traction in certain mathematical uses.

 

There seems to be some activity to update Prime95 to support Ryzen, although for the moment the latest test version I've seen adds better CPU detection and uses Intel FMA3 code still. I will say there haven't been any problems I'm aware of running the stress test on Ryzen in FMA3 mode. I don't know if things will change much if an AMD native/optimised FMA3 was implemented. My impressions so far are that if you do want high FMA3 performance, stick to Intel. Any Sky/Kabylake quad core i5 would beat an equivalently clocked R7 in this use case. At best R7 has half the IPC.

[HalGameGuru]

Isn't a benchmark/stress-test run like that exactly what you would use in the R&D department to test your implementation of the instruction set? How would this have gone under the radar? I HOPE this is something fixable in the code rather than the lithography.

[zMeul]

3 minutes ago, nerdslayer1 said:

-

please remove that trash 

[Sakkura]

26 minutes ago, samcool55 said:

Well as far as i can work out they dropped FMA4 so i don't think that many programs use it...

Can be wrong tho, i have no idea how this stuff works.

This bug concerns FMA3, which is different from FMA4.

 

AMD introduced FMA4 first, then switched to FMA3 when it became clear Intel was going to be using FMA3.

 

There likely isn't much software that uses FMA4, but FMA3 usage should be higher because it's what both Intel and AMD are pushing.

[The Sloth]

Just now, zMeul said:

please remove that trash 

what "trash"?

[zMeul]

Just now, nerdslayer1 said:

what "trash"?

AdoredTV 

[The Sloth]

1 minute ago, zMeul said:

AdoredTV 

so whats wrong with it? he shows evidence and tells the truth.

[matrix07012]

2 minutes ago, zMeul said:

please remove that trash 

I can't ban you. I would have to ask someone.

Edited March 15, 2017 by matrix07012
grammar

[Sakkura]

Just now, porina said:

There were two problems with FMA4: it was only on AMD processors so already there is a limited use case, and the only person I know who tried to use it (George Woltman, of Prime95 fame) couldn't extract any performance boost from using it over not using it. 

 

FMA3 was implemented by Intel in Haswell, and it did provide a huge performance boost and gained huge traction in certain mathematical uses.

FMA4 and FMA3 do basically the same thing, it's just a small difference. What matters is AMD ditched FMA4 when it became clear Intel would be going for FMA3. AMD actually launched processors with FMA3 support before Intel.

[zMeul]

Just now, matrix07012 said:

I can't ban you. I've have to ask someone.

why would you ban me? because AdoredTV is actual trash!?!? are you even aware of the background of that channel?