Digital Gold Rush Powered by GPUs
Coinlab, a Seattle-based startup that offers a unique monetization service, is considering a move into high performance computing. Today the company mines Bitcoins, digital currency to help online gamers make money. But they believe the same infrastructure that powers their operation can be parlayed into an HPC service business.
Bitcoins started as a small community-based project back in 2009 and has developed into an active online economy with users trading virtual currency for goods, services and even cash. Virtual currency has several advantages, the primary one being that it doesn’t have to answer to a central bank or national government.
Since there isn’t a central bank for the virtual currency, all transactions are generated and verified by the user base. Each computer running a Bitcoin client keeps a copy of the transaction data, which is updated every ten minutes. The transaction database serves as a central point of reference to all members of the community.
To reduce the chance of individuals hacking the database or attempting to rewrite its history, each update is encrypted with a cryptographic hash function. Here’s where the compute challenge comes in. Over time, the difficulty of processing the hash increases in order to keep pace with faster computers and more transactions.
Decoding currency transactions requires plenty of computational power, but the user that successfully solves the hash is rewarded 50 Bitcoins. This process serves two main functions; it encourages the community to validate the currency’s transactions, while simultaneously growing the Bitcoin economy. The users that offer up cycle time to solve the transaction encryption are essentially 21st century gold miners. And since they’re gamers, they tend to use GPUs as the computational engines to gather their treasure.
Coinlab is one such mining outfit that aggregates compute cycles from Bitcoin users, while acting as a currency exchange. Their contributors form a crowd-sourced supercomputer, powered by off-the-shelf gaming GPUs. The cluster is constantly crunching out the encryption to earn Bitcoins.
HPCwire recently caught up with Coinlab’s CEO, Peter Vessenes, and project manager, Chris Koss, to talk about their business model. They explained that individual users (their clients) receive a cut of the earnings made from their mining efforts. Given the inherent variety of equipment in this type of infrastructure, payouts are determined by the specs of their GPUs. For example, an iMac with an AMD Radeon 6970M could generate roughly a dollar per day using the company’s client.
Koss chuckled when asked to name the poorest performing GPU he’s seen on the network. “I saw a Radeon X1600, which I think came out in 2002,” he told HPCwire. “We round to the nearest 1/10th of a megahash per second, which is the unit used for Bitcoin mining, and it scored a 0.0. It produced nothing….” At the other end is the AMD Radeon HD7970, which generates the highest revenue; “I think that one makes $3.15 a day right now,” he said.
Mining can be a fairly lucrative operation, depending on exchange rates. In July 2010, Bitcoins traded as low as $0.04, but the virtual currency has experienced dramatic volatility, with the value of a single Bitcoin jumping to a high of $31.99 in June 2011. Today, the currency is valued at slightly less than half of its 2011 peak at $11.36.
Coinlab typically pays its users in dollars, but is willing to exchange cycles for a variety of other currencies. “We get paid by the Bitcoin network and we convert that to dollars and use those dollars to buy whatever asset people want,” said Vessenes. In one case, the company partnered with developers of a game called Wurm Online to deliver in-game currency for gamers who contributed GPU cycles.
There is a delicate balance between GPU performance, processing difficulty and Bitcoin’s exchange value that determines the feasibility of mining. If the hash values are too difficult to decode, or the exchange rate drops, the financial incentive to mine for Bitcoins starts to dwindle. Right now, miners can see a return on investment in less than a year. “A card will pay for itself in about eight months at current rates,” said Koss.
The returns are an important aspect for a set of determined miners who go so far as to build miniature datacenters dedicated to the process. Two of Coinlab’s users are of this ilk, contributing more than 100 cards to the company’s cluster. Vessenes said they are extremely focused on optimizing each component on their system to get the highest return from their hardware.
That being said, there are a number of Bitcoin miners who have gone so far as to develop purpose-built hardware with custom ASICs. The new equipment has some of GPU users feeling doubtful about the future of their mining operations. “People are pretty bearish about the long-term prospects of GPU mining on the Bitcoin network,” said Koss. “There are miners who built these $100,000 hardware installations and they’re asking ‘do I sell it right now or do I hold on?'”
Vessenes and Koss, aware of this shift in the mining culture, have started to look into new applications for their crowd-sourced supercomputer. The company has more than 500 GPUs at their disposal and looks to keep that number growing. Their hope is to have HPC users kick the tires on their cluster. If everything works out, they could eventually rent cycles out to researchers and former GPU miners would be able to keep their equipment running.
But that might be easier said than done. While computing over the Web has a number of advantages, the primary challenge lies in moving data in and out of compute nodes. This becomes more difficult when those computers are spread out geographically on consumer-grade Internet connections. That’s much less of a problem for embarrassingly parallel HPC workloads where node-to-node communication doesn’t matter so much. For those types of codes Coinlab might just find its niche.