A Phoronix reader has talked about the efficiency of using Raptor Computing Systems’ open-source Blackbird POWER9 desktop system for Monero cryptocurrency mining in 2019.
My Blackbird from the Black Friday sale had finally arrived. How to best burn it in, if not
with some crypto mining?
Back in 2017, when Monero was using the CryptoNightV7 algorithm,
Phoronix reported that POWER9 had higher mining efficiency than the common x86
processors. There have been two algorithm changes since then, and a third one is coming this
October. How does POWER9 fare now?
You’re probably familiar with the Blackbird by now, with Phoronix having a 4-core unit for
benchmarks. Mine was bundled with the 8-core option, but is otherwise similar. Build specs:
- 8-core POWER9, 3.45GHz base / 3.8GHz turbo, 160W TDP, SMT4
- 16GB DDR4 ECC, 2666 MHz, one stick
- 3TB spinning rust
Unlike my POWER8 server, the Blackbird cannot measure its own system power consumption (only
the processor’s), so I used a simple watt meter to take measurements. When off, with just the
BMC on, the system took so little power my meter could not measure
it. It kept showing 0 W, so presumably it’s under a Watt. At idle, 55 W.
For Monero’s current version, using the CryptoNight/R algorithm, there’s a fully featured
miner for POWER by nioroso-x3, xmrig. Git version fcf639e8274 was used for the tests. Huge
pages were enabled, spectre/etc protections at their defaults.
For each SMT mode, I tried six thread options. The SMT scaling is as expected,
at SMT1 there are eight threads, and performance drops after; at SMT2 16 threads, and a
corresponding drop after. The “more resources for each thread” effect is also slightly
visible, with SMT1 having the highest result at eight mining threads.
In SMT4, the efficiency scaling is quite nice, showing that a mere eight-core is not even
close to the bottleneck here.
Efficiency per Watt continues to scale well, but it’s far from the x86 options. Intel
processors get about 5 hashes/s/W, while AMD gets about 10. In its best config, this POWER9
cpu only got 2.86. Temps stayed quite reasonable, about 70 C at max.
smt4 threads rate power 1 36.5 77 2 76.7 91 4 147 117 8 294 170 16 423.7 202 32 575.1 201 smt2 1 37.5 77 2 77.1 90 4 144.8 116 8 285.8 167 16 417.9 199 32 418.7 200 smt1 1 35.7 78 2 72.7 89 4 151.1 117 8 304.4 170 16 295.1 166 32 283.3 165
So, for the current algorithm, CryptoNight/R, POWER cannot match the x86 options. According
to the miner author, it is fully optimized already. What about profitability?
According to whattomine.com‘s calculator, 575 hashes/s
would result in 0.00003 BTC/day, 0.27 $ at the current prices. A day’s electrity use would be 4.824 kWh, and at a cost of ten
cents per kWh, cost 48 cents. A loss of 21 cents a day, at current prices anyway; someone
mining to hold may have a different view.
Would winter heating change the picture? For me with district heating,
one kWh of heat costs about 6 cents. The system output 4.824 kWh a day, which would save
28.9 cents of heating costs/day. That would tip the scales to 7.9 cents profit/day.
Before wrapping things up, we do have to consider the October algorithm change. The RandomX
algorithm penalizes GPUs even more, and has the potential to change things in CPU rankings
I tested git version 5d815c57c086 of tevador’s randomx-benchmark.
./randomx-benchmark --mine --largePages --threads N --init 8
In its current state, the RandomX implementation does have support for POWER crypto
acceleration, but it lacks the JIT support of x86 and arm64. The README states the JIT has
at least 10x of an effect on hash rates. It affects the startup costs too; on x86, the
memory initialization phase takes 5-10 s, here it took 192 s.
smt1 threads rate power 1 15.1 77 8 120.9 139
Intel Core i9-9900K is reported to get 5770 hashes/s using eight threads, while AMD Ryzen 7
1700 at the same thread count gets 4100. Our 120 is quite far off, but should someone port
the JIT to POWER, the situation may change.
- Monero mining on POWER no longer slam dunk
- Winter heating or holding for appreciation may change picture
- October algorithm change needs a JIT