Slashdot Mirror
AMD Confirms It's Issuing a Fix To Stop New Ryzen Processors From Crashing Desktops (digitaltrends.com)
AMD says the company has been able to figure out why FMA3 code is causing system hangs on PCs using a new Ryzen desktop processor. From a report: Although AMD didn't provide a detailed report on the problem's root cause, the company said that BIOS changes will be distributed to motherboard manufacturers to resolve the issue. Customers are encouraged to keep an eye on their motherboard vendor's website for an update. "We are aware of select instances where FMA code can result in a system hang," the company said. "We have identified the root cause." AMD released three Ryzen-branded desktop processors at the beginning of March that plug into motherboards based on AMD's new AM4 socket. The trio of processors include the Ryzen 7 1800X, the Ryzen 7 1700X, and the Ryzen 7 1700. However, all three reportedly cause a hard system lock when running certain FMA3 workloads. The problem was replicated across all three processors and a variety of motherboards.
Kidding, but I'm old enough to remember when running Windows 95 on the old AMD K6 boxen was a no-go...
That said, does this fix affect performance any (no matter the OS)?
Quo usque tandem abutere, Nimbus, patientia nostra?
Apparently, I had to look into the forum posts that the FA referenced, and FMA instructions are Fused Multiply Add, whatever the fuck that is.
After looking at Wikipedia for 5 seconds, FMA instructions perform round(a+b*c) in a single operation, so you can a) speed up and b) get more accurate results whenever you need such a mathematical operation (which is actually reasonably frequently, in numerical computing).
"None can love freedom heartily, but good men; the rest love not freedom, but license." --John Milton
FMA instructions are Fused Multiply Add instruction. Usually their are on SIMD registers and allow you to do "a += b *c". Modern cores can do that on a vector in a single cycle. Actually, they may be able to do more than one FMA on a vector register in a single cycle.
FMA are most commonly used to compute dot product, and are therefore very helpful in linear algebra. (And so they are useful in a ton of data mining algorithms.)
Here ya go I mean it doesn't say, "WONTFIX, LOL" verbatim, but there are 128 items in there labeled "No Fix", so they may as well say LOL after it.
Why bother? Just do what time and again I've found works best for a trouble-free life: totally avoid any/all AMD CPUs or GPUs.
Heh, if you want to avoid bugs, better stay away from Intel also. They have a 1000 page errata list with every single processor.
Also known as the Multiply-Accumulate (MAC) instruction in DSPs. MAC is an extremely common instruction in signal processing kernels (the inner loop that does the calculations). It is vital to be able to do a lot of them per clock cycle. In fact, it's often why DSPs have special looping registers so you can do zero-overhead loops and thus doing a sequence of MACs without incurring branch (and branch prediction) times thus being able to do nothing but this instruction for very little overhead
But I can use FMA3 instructions on multiple Intel microarchitectures without it ever causing my system to freeze up. That's the kind of "quality" only AMD can provide.
Try TSX instruction in a Haswell or Broadwell.. Well if your bios haven't been updated by Intel to disabled the instructions completely.
Errata is unfortunately pretty common in CPUs these days. I hope this one is fixed without having to disable the entire extensions like Intel does.