Hmm are we having another "AMD motherboards are crap" moment?
Or is it simply that delivering 200+W at load through a CPU socket can't be reliably done at consumer prices?
Anyone has had this problem with less high end CPUs? Something at 95-65 W?
Or is it simply that delivering 200+W at load through a CPU socket can't be reliably done at consumer prices?
Anyone has had this problem with less high end CPUs? Something at 95-65 W?
The "CPU defective by design" in the title might be a bit misguided since the suggested workarounds do not address a CPU issue but a motherboard one.
My motherboard is the highest end consumer motherboard GIGABYTE has ever built (as far as I know), and I wouldn't say that they tried to keep costs down (it's listed at $849 on PartPicker, and it was introduced at $999) and the power delivery stage is insane. Here's BuildZoid's _in-depth_ review and analysis of its VRM: https://www.youtube.com/watch?v=HMUWzDSAS9c
I now literally go out of my way not to get the flagship motherboard models, even if it means holding off on a purchase until a lower-spec'd model comes out, and have never regretted it since. (I also will never again buy Gigabyte motherboards, either.)
Just because it's high end doesn't mean they won't cut corners or fail to test properly. My ROG motherboard was (at the time) the most expensive motherboard for the socket. Yet it behaved overall worse than many mid/low-end motherboards I owned, an experience shared with other owners of the same board. Even if this is a problem with the CPU that is mitigated by changing parameters of the motherboard it should have been caught during the motherboard design and testing. I can't imagine Gigabyte's engineers noticing this stuff and saying "just ship it like that, nobody will notice". So the best interpretation I can have is that they missed it during testing (the worst is that the marketing department said "we have to put it out there fast, all else be damned", and all else was damned).
High end in consumer stuff, and especially anything gaming related, is basically scam.. look mah, 16000 dpi! And blingy leds! (But no engineering)
Sure they might drop an over-specced part somewhere in it but it's just marketing when the rest of the product is crap and still has no proper engineering behind it.
It's incredibly frustrating. I think there was a time when you could generally assume that expensive = high end = actually good, but now it's just a cheap thing with crazy markup and a premium part (but nowhere near premium enough to justify the markup) or two somewhere (where it probably doesn't matter much anyway) along with other gimmicks. Now it's just expensive = expensive, good or maybe not.
Google's stressapptest runs fine for long durations, building a kernel with make -j32 succeeds (and can boot it), every parallelized archiver like 7z,pbzip2,pigz,xz is ok, and even gaming on a Windows VM using 8c/16t + GPU passthrough works well.
This is my first Ryzen system so I chalked it up to a possible carry over of DDR4 issues from earlier generations. I didn't investigate further since the 3950x had just been released and I couldn't find any other reports of prime95-only instability until now. And just to reiterate, it is perfectly stable otherwise.
Heavy AVX2 workload breaking things fits better so will have to try to collect more data.
Then we get the R480 melting cables for the same reason... and so on.
In this case, AMD wanted to keep using 6pin cable despite it not being approprite, on the 380 it just slowed it down.
On 380X, it is unstable, it still don't draw enough to melt things, but you need hackery to make the GPU useable
The 480 they outright made a GPU that used more power than the cables specs would allow, and insisted in using the same power delivery design the 380 had... Their "fix" to the issue was make patches that would just make the GPU run slow, and make it misbehave like the 380X does.
RX480 wasn't melting cables, it was melting motherboards. The stock VBIOS was pulling more power than the PCIe spec allowed you to pull from the slot. It was pulling over 100W, vs the 75W spec, not massive but probably enough to push some older/weaker boards over the edge (most of which were just ready to fail anyway and were going to fail from a solid 75W draw too).
https://www.tweaktown.com/news/52871/amds-radeon-rx-480-draw...
It's pretty hard to melt a cable. You can overdraw the connectors by roughly twice their rated power safely, and the cables will do more than that as long as you're not using splitters or some other hack.
The 295x2 pulled almost 500W through a pair of 8-pins and the slot (nominally that's 375W rated) while overclocking.
Part of it was likely that NVIDIA's performance was so good. Some anecdotal rumors from the AMD vlogosphere suggest that AMD thought that the RX 480 would be competitive with the GTX 1080 - I consider this kinda dubious because the RX 480 is basically a GTX 980 tier chip, so that would mean AMD thought that NVIDIA wouldn't make any progress at all from a node shrink, which seems dubious. Maybe they figured the RX 480 would be a lot faster or more efficient than it actually ended up being.
Polaris was the last generation with "old style" voltage control where you just set a target and go, it is possible that they figured the chips would hit close to 2 GHz like NVIDIA's but ended up with validation problems and didn't hit the expected clocks. This could possibly be the reason they nuked Big Polaris and ended up lengthening the pipeline so much in Vega to try and get the clocks up (along with adding the Pascal-style "smart" power conditioning management).
Anyway, regardless, the point is that it seems likely that AMD was forced to push the RX 480 much farther than they intended to, at the last moment. Like, after the PCBs had already gone out for manufacturing, and it was too late to switch the 6-pin to an 8-pin (which would have solved the whole problem, that would have allowed 225W rated total board power).
It was still a bit of a showoff move to run from only a 6-pin, there is no way that the card would have been significantly less than 150W total board power, but perhaps defensible as a compatibility move - although some marginal PSUs still might not have handled it.
It could be that the spike is the problem not the end load. The scenario in play here is going from entirely idle to maximum load in an instant. So the CPU is going to shoot straight from ~20-30w usage to 280w (or even higher if the power management is a bit sluggish at reducing the clocks). That's a pretty drastic swing if anything isn't entirely up to spec, assuming the spec even handles this properly.
In theory the 3970X could even spike all the way up to 430w. The individual cores top out at 13.5w, so if they all end up running at single-core turbo frequencies even for a split second that's going to be brutal.
> Anyone has had this problem with less high end CPUs? Something at 95-65 W?
My 3700X in an old X370 board has been flawless. As has many others, Ryzen 5/7/9 are _widely_ recommended and have been for a while. Any systemic issues in the "regular" consumer end would have cropped up by now.