You are on page 1of 2

Team Red Miner MTP Mining

=========================
The MTP algo is primarily used by Zcoin (XZC). We're writing this
document since the algo is a little different from the other
algorithms available in TRM.

4GB Cards
=========
4GB cards are NOT supported as the MTP scratchpad will not fit in the
available vram. The miner will automatically exclude GPUs where the
driver reports less than 4.4GB available vram. If you still want to
force the miner to run you can add the --allow_all_devices argument.

Characteristics
===============
MTP runs HOT on your gpus and is truly a power hog algo. On most gpus
it will consume close to 100% of the compute resources available while
also activating the mem controller to 75-85% of full load. Hence,
MTP's profile is both that of a compute bound and a mem bound
algo. For example, if you limit the compute resources too much with
e.g. a lower core clk you can end up running the mem clk unnecessarily
high and vice versa. In practice, this means you need to find the
correct balance between core and mem clk for max efficiency.

Due to the power hog profile where e.g. a Vega64 LC at 1407@900 core
clk, 1107@900 mem clk easily can pull 250W, we cannot stress this
enough: DO NOT START THE MINER OF A FULL RIG AT "NORMAL" CLOCKS unless
you know your PSU can handle a very high load. Run it on 1-2 gpus
first using -d 0,1 and observe, or lower your clocks and work your way
upwards instead when tuning.

Tuning
======
For Polaris (470-580s) cards, a typical configuration would be a core
clk around 1200 MHz. Then, start with a mem clk around 1900-2000 MHz
and lower it in steps of 50 MHz and measure the hashrate for each
change. The first steps should have a very small impact on the
hashrate, but at some point the mem clk will have a more significant
impact. This point varies a lot between gpus, you have to find your
own. After finding your hard limit for the mem clk, adjust it upwards
a little bit to a hashrate of your liking given the power draw. Last,
dial down both core and mem voltage as much as possible to save power.

For Vegas, as stated above this algo runs HOT. To improve efficiency,
the key is to lower your mem clk significantly, to 800 MHz or even 500
MHz. You should also try to force the mem power state to P2 or P1
since this will lower the soc clk, resulting in lower power draw as
well.

In a full throttle configuration on my Vega 64 LC using e.g. 1400 cclk


at 860 mV, 800 memclk, I can push it to 3.5 MH/s. This will be a 250W+
power draw. This can absolutely be a net win in the end, but you need
to check your numbers and account for your power costs.

For a full Vega rig, a more reasonable approach is to clock down mem
to 500 MHz in mem p-state 1 and the core clk to anything between
1180-1250 MHz. Start with the core clk in the low range and set
voltages to 800 mV (for both core and mem), then increase core clk to
see how far you can push it while remaining at 800mV. This will lead
to 2.8-2.9 MH/s for a Vega 64 and 2.6-2.7 (maybe) MH/s for Vega 56s,
still pulling around 200W per gpu.

Nicehash Mining
===============
TRM supports Nicehash mining for MTP out of the box. However, there is
a confirmed bug with Nicehash for ntime rolling which they are aware
of. Meanwhile TRM will automatically apply the option
"--no_ntime_roll" whenever Nicehash mining is detected.

Dev Fee Switching


=================
In all other TRM algos, we run the dev fee hashes with a very fine
granularity, many times truly concurrently with the user's work. This
means that there are no interruptions for the user, a smooth poolside
hashrate and zero downtime for pool switching (which can have a very
negative impact with pools that penalize pool hoppers).

For MTP, this approach is not possible. The user's and dev's work will
never use the same 4GB scratchpad. Therefore, MTP must be implemented
using a more traditional switching mechanism. We've worked actively to
reduce the impact for the user with this less sophisticated approach:

- Heavily optimized code for the pad rebuild, fully done on gpu
(1.6-1.9 secs on both Polaris and Vega GPUs).

- We're flexible with when we switch, always trying to switch when a


new user job arrives which means that the pad would have to be
rebuilt anyway.

- We remove -0.1% from the listed dev fee, i.e. devs take the hit for
the extra pad rebuild, not the user.

Last, we don't log explicitly when we switch to dev hashing. That


said, we do not switch often, and the user should indeed see a
poolside hashrate of 97.5% of the average hashrate over time. Another
difference compared to other miners is that different GPUs are not
necessarily switched at the same time. However, please note that there
will always be a single early dev fee switch on all GPUs to avoid easy
circumvention of the dev fee by restarting the miner repeatedly. This
means that the (very) early poolside hashrate for the user might be
lower than expected.

You might also like