In TRM, "smooth power" means carefully managing gpu activity to produce the smoothest possible power draw over time on a fine-grained level. The technique is used for all gpu types and on all algos in the ethash family (ethash, etchash, kawpow, firopow). The objective is to be as nice as possible to the whole power delivery system, from PSU to risers to GPU VRMs. The assumption is that if we make life easy for the voltage regulators we should be able to survive on a lower voltage than in previous TRM versions and other miners. The way we address the problem in TRM also has the nice side effect of slightly increasing hashrate on most gpus, normally in the 0.1-0.2% range.
To illustrate what the above means in practice, please see the following scope captures of 5 secs of uninterrupted ethash mining with TRM 0.9.4.2, Phoenix 6.2c, lolminer 1.51a, and TRM 0.10.2, all on a Sapphire Pulse 580 at a ~31 MH/s tuning. All captures were performed using a Siglent SDS2104X Plus and a 300kHz CP6220 current probe.
| TRM 0.9.4.2 |
|---|
 |
| Phoenix 6.2c |
|---|
 |
| lolminer 1.51a |
|---|
 |
| TRM 0.10.2 |
|---|
 |
Clearly visible, all miners except TRM 0.10.2 suffer from repeated up/down swings, i.e. deviations from the average power draw. Lolminer is exhibiting the worst behavior. Phoenix and TRM 0.9.4.2 behave similarly with Phoenix having slightly less pronounced swings but occuring more frequently. The outlier is TRM 0.10.2 that has a smooth continuous draw without any swings at all deviating from the average level.
To verify that the new behavior in TRM 0.10.2 does indeed translate into more robust mining at lower voltages we tested running the same Sapphire Pulse 580 at 1138 MHz core, 2080 MHz mem, ref 30 for 30.9 MH/s at a low core voltage of 793mV.
| Miner | Result |
|---|---|
| TRM 0.9.4.2 | Crashed immediately after dag build. |
| Phoenix 6.2c | Crashed after 10 secs of tuning. |
| lolminer 1.51a | Crashed immediately after dag build. |
| TRM 0.10.2 | Ran for 2h without crashing. |
NOTE: the smooth power feature was released in v0.10.0, but in hindsight it needed additional polishing and fine-tuning, especially for Polaris and Vega gpus. We hope that the version released in v0.10.2 will address earlier issues.
This feature normally doesn't need tuning, but a few pointers are still interesting:
-
As you try to minimize voltages, switch to a static --eth_config for all gpus. This will skip the tuning part and enter the best continuous smooth power setup we can produce immediately after the dag build.
-
Our Polaris tuning had a preference to increase intensity on certain mem straps in 0.10.0 and 0.10.1. This could trip some rigs, and was not intended. This is fixed in v0.10.2, but we still recommend using a static --eth_config when trying to tune for a low voltage.
-
As always with advanced features, some gpus simply aren't reacting the way theory dictates they should. If you can't seem to get a gpu stable regardless of what you try, you can always test turning off the smooth power feature. Using e.g.
--eth_smooth_power=1,1,0,0,1would turn it off for gpus 2 and 3 in the rig but keep it enabled for gpus 0, 1, and 4. Passing--eth_smooth_power=0will turn off the feature for all gpus in the rig.
GPU-specific pointers:
-
Polaris: we've found that most Polaris gpus can be lowered 10-15mV, all else being equal, in some cases more as well. Unfortunately, there are always some bad silicon samples that will need a certain amount of voltage for stability.
-
RX Vegas: same as for Polaris, many gpus can shave off additional core voltage but still remain stable. If you've had issues remaining stable on RX Vegas in general on later TRM versions, you should try switching to v0.10.2 but keep the old tuning. If you still have issues, try turning off smooth power.
-
VIIs: no expected change in voltage required, no need to retune.
-
Navi: our testing indicates that smooth power results in increased stability on the vast majority of Navi gpus. For the remaining ones that rather seems to be crashing more often, turning off smooth power is a good test.