4 ms·

I have a AMD GPU, a 380X... and to be honest I am just sure AMD power management is crap in general. 380X in particular has bigger demand than 380, but AMD didn

I have a AMD GPU, a 380X... and to be honest I am just sure AMD power management is crap in general. 380X in particular has bigger demand than 380, but AMD didn't bother raising power limit, causing some weird issues (some games for example get unstable unless I use some overclocking app to reduce voltage and raise power limit... mind you, the normal 380 people already used it with maximum power limit by default, because it was already too low...). Then we get the R480 melting cables for the same reason... and so on.

nottorp 7y ago
Hmm whatever is melting cables is drawing more power than the cables are specced for. I doubt increasing the power limit will help with that...
- speeder 7y ago
  You are correct. But what I meant is that the source of the problem was the same. 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.
  - jotm 7y ago
    Wow, that's ridiculous...
paulmd 7y ago
> Then we get the R480 melting cables for the same reason... and so on. 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-drawing-much-power-over-pcie-slot/index.html 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.
- jotm 7y ago
  The stock VBIOS?! Wth were they thinking when programming it...
  - paulmd 7y ago
    Basically, the chips drew more power than expected. You can tell from the fact that they only put one 6-pin on there. That gives a total board power limit of 150W, which the card actually exceeds pretty easily (pulls about 200W). They made the mistake of overdrawing the slot, the fix was to switch to overdrawing the power cable instead (which can tolerate it much more easily, you can very safely do up to twice the rated power). 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.

nottorp 7y ago

Hmm whatever is melting cables is drawing more power than the cables are specced for. I doubt increasing the power limit will help with that...

speeder 7y ago
You are correct. But what I meant is that the source of the problem was the same. 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.
- jotm 7y ago
  Wow, that's ridiculous...

paulmd 7y ago

> Then we get the R480 melting cables for the same reason... and so on. 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-drawing-much-power-over-pcie-slot/index.html 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.

jotm 7y ago
The stock VBIOS?! Wth were they thinking when programming it...
- paulmd 7y ago
  Basically, the chips drew more power than expected. You can tell from the fact that they only put one 6-pin on there. That gives a total board power limit of 150W, which the card actually exceeds pretty easily (pulls about 200W). They made the mistake of overdrawing the slot, the fix was to switch to overdrawing the power cable instead (which can tolerate it much more easily, you can very safely do up to twice the rated power). 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.