214
My father works in the aviation industry, and as a consequence, he's no stranger to the frequent -- often misguided -- question of 'what kind of airplane should I buy?' among his peers.
Fortunately, such a inquiry can be met with the straightforward and sincere retort of 'Well, what do you want to do with your airplane?'
Oftentimes, such pressing but justified scrutiny reveals the fundamental differences between romantic association and practical application. There may very well be plenty of people who have decided they'd genuinely consider the undertaking of ponying up the money for, say, a Learjet. But of all those people, even out of all those who could afford the actual dollar-cost for possession of the unit, just how many would be capable of the regular costs incurred for the basic maintenance of owning a private jet?
Much like a GPU, an aircraft may be viewed as an industrial tool, a means to an end. A device that requires power and maintenance to use, and a specific set of conditions to be used efficiently. And indeed, for those holding authority over either, there is a great incentive to lease them out to some other organization so as to offset some of the cost. A jet sitting idle in a hanger is doing little more than ticking the days along until the first layers of rust begins to develop, the weeks until paint begins to wear, and months until the cost of the regularly mandated inspection rolls around. All of these phenomena cut into the value of the assemblage by threatening to separate what might be an airworthy vehicle from a mere heap of expensive metals. A stationary jet simply isn't doing any good for anybody -- and "for parts" listings are always cheaper.
So when I made the admittedly somewhat-impulsive decision to purchase the P106-90 NVIDIA mining card, I did so partly out of curiosity. I'd already played around with my GTX 1060 card for years now and watched how it performed when tailored to a rather specialized function -- rendering graphics. Now that I had the two very similar cards working side by side on the same system, the opportunity for performance comparison was practically inviting itself in -- even if I had to keep a somewhat open mind about my newly acquired hardware's potential, given that it'd been tailored to perform a rather different specialized function.
Now of course, there are the usual performance benchmarks, often used in abstract, to gauge the faculties of hardware. Metrics may be analyzed, such as Floating Point Operations per Second -- know as FLOPS. Or physical qualities can be compared, like transistor size or die size. Or, we might turn to universally recognized standards to serve as a comparison. You know, like the usual "Bam! I got DOOM to run on a raspberry pi and a pregnancy test screen!" sort of metric.
A side by side specs comparison might be a handy place to start. Both units come with extremely similar physical specs, in great part due to chipsets from the same GP106 family. Not surprisingly, both work in the pascal architecture, and employ a process size of 16nm. Both types of cards interface with the motherboard through PCIe 3.0 x16. Both use the same exact variations of GDDR5 for their VRAM, ranging from 3gb to 6gb, each clocked at 192.2 GB/s through 192 bit interface. Initial impressions from these numbers alone, these cards would appear entirely identical. At least initially.
That's where the similarities end. While the GTX 1060 3gb is ascribed the processor name GP106-300-A1, its mining-focused alternative is registered, not surprisingly, as the GP-106-090.
GP106-300-A1:
Base Clock - Maximum Clock: 1506-1708 MHz
FP32 (float) performance: 3.935 TFLOPS
FP64 (double) performance: 123 GFLOPS
Power: 1x6-pin, 120 W
GP106-090-A1:
Base Clock - Maximum Clock: 1354-1531 MHz
FP32 (float) performance: 2.352 TFLOPS
FP64 (double) performance: 73.49 GFLOPS
Power: 1x6-pin, 75 W
Furthermore, a more obvious physical difference becomes apparent upon close examination of the cards. While one has HDMI, DVI and DisplayPorts, the other is missing its video output port entirely. The P106-90 card, and cards of similar background, are indeed stripped of some conventional functionality. Factory clock speed is limited to 200 MHz behind its twin. Theoretical performance is limited -- namely through the reduction of total number of shading units.
And yet, the Render Output Units, the process of applying mathematical functions to a pixel pipeline to display an output, remains entirely unaffected. Both units keep at a rate of 48 per second. Likewise, both the L1 and L2 cache are identical, at 48kb and 1536kb respectively. These differences in their sum total lead me to a few general conclusions about the mining card's origins. The capacity for each card to perform specific algebraic functions through linear operators is the one task the card excels at in particular. For everything else, restrictions are made, allowing the P106-90 to be under-powered to the point where it depends on nearly half of the wattage of their graphics card cousin.
The final product is a mathematically-efficient machine. A device built from the ground up to be electrically economical. Designed to turn those Watts into pure number-crunching.
For those who employ Folding At Home either altruistically or to earn Bananos and/or Dogecoin, the basic rendering of the protein folding process strikes me as identical when measured in credits completed for some period of time. This may be in part due to the nature of work load size and credit granted per work unit completed, and might be accounted for by comparing VRAM size. It's anybody's guess as far as I'm concerned -- the manner in which work units are exchanged for currencies is beyond me.
I will say that from a superficial perspective, the rate of BANANO earned over a given period of time seems to be consistent and almost exactly doubled rate from historic records. This suggests not only that the P106-90 is of comparable performance when applied scientifically, it is a more power-efficient alternative than its counterpart.
Additionally, I am proud to announce that games like DOOM and DOOM: ETERNAL appeared to be perfectly compatible with either chipset. Granted, this is in part due to the game's beautiful optimization, but also through some wizard-grade linux compatibility with a mining card that has no video output.
For some reason, my system recognized the card with greater VRAM, assigned it the task of rendering the game, and started sending data to my other card to act as its bridge to the HDMI port. And while I might have no idea what prompts vulcan shaders on NVIDIA's 460.80 linux drivers to want to select the mining card as it's default, I'm also not one to complain. I consider it a successful experiment, as the shaders still run every bit as beautifully as far as I can tell. If there are any complaints to be had, it's worth noting the technical reductions likely restrict the frame rate to a modest, but still quite tolerable level.
Alternatively, with two fully capable cards, the system is given more freedom to multitask. If you choose to dedicate both to folding in partnership with your CPU, you're welcome to do so. If you want to mine and game at the same time, you can pick and choose the appropriate GPUs for either function. If you want to use one card for protein folding while the other's running Photoshop, nobody can stop you.
And of course, we can analyze the unit's capacity for its designed function. While I think it's a healthier perspective to view any cryptocurrency attained via folding as little more than compensation for electricity used, the elephant's already in the room. How do we measure the primary function of the card on hand by its maximum potential mining profitability?
Like all things concerning mining, the real answer's a little fuzzy. Factors depend on the overclocking, ideal temperature, and the price of Ether -- or whatever cryptocurrency you're choosing to mine -- over a period of time. Since the process of mining is, in a way, merely the act of guessing at random numbers, luck is a noteworthy component for one's consideration. And we need not even speak of pool fees or downtime.
The numbers are in: the thing mines Ether at 11.58 MH/s when at base clock on a good day. And while this isn't hugely impressive, the numbers stay consistent -- an important consideration for any mining equipment. Interestingly enough, long-term profitability from rate of shares submitted seems to indicate a greater-than-average rate over time than what the unit reports. Either the card's just being modest, or I've just had greater luck than usual with it.
So What's The Final Verdict?
Buying a P106-90 card certainly won't make you rich any time soon -- but at the same time, it won't be burning a hole in your electric bill, either. With an average of seventy Watts consumed, it works more efficiently than some light-bulbs, while consistently outputting about a dollar a day. As far as I can tell, it functions every bit the same as a GTX 1060 when it comes to protein folding. And although gaming on the thing isn't a straightforward endeavor, it's not impossible either -- and the results aren't unpleasant to behold.
In short: there will be a fundamental difference in how much you pay for a GTX 1060 versus a P106-90. Those costs will stem both from the inherent restrictions of the card's chipset, combined with the pains a user will undergo in setting up their hardware to perform their function of choice. If you're looking at a mining card as a cheaper alternative to a gaming card as an 'easy fixer-upper', you might do well to remind yourself that the money you save might not be worth the man-hours and auxiliary hardware needed to accomplish your goal. If you're looking for an unobtrusive card that works quietly in the background to earn passive income or otherwise to do its part to FOLD proteins for a COVID-19 cure, you could do a lot worse than a humble and efficient mining card bought second hand.
