Update: Golem’s ICO is over. It raised 820,000 ETH (roughly $8.6 million) in twenty minutes.
Golem aims to build a P2P computing network across the globe: anyone can rent out computing power. Blockchain technology will handle payments and data routing. They are holding an ICO for their GNT tokens: the sole means of payment.
Here is an overview of the ICO details, including links to other documentation and instructions on participating.
Below are our thoughts about the project and its viability.
- Golem is more accurately thought of as a distributed supercomputer than a cloud services provider. In the foreseeable future, the network will provide raw CPU power for computation tasks, not live web hosting or interactive software. Golem will utilize those CPUs in specific types of tasks, like CGI rendering. There are still plenty of technical hurdles until Golem is P2P SaaS.
- Even ignoring the cloud services that Golem won’t be providing, the market for supercomputing is small(er) but real. CGI rendering is computationally hungry and its scale and frequency is set to grow: animators, designers, and architects increasingly need them. Golem will probably not be processing scenes from Disney, but it could open up low-cost rendering options.
- Golem has many competitors in the blockchain industry and outside it. Traditional cloud computing does pose a significant source of competition. Costs for cloud computing broadly are dropping fast, and in the short-term, Golem is primarily competing with supercomputing (analytics, processing) not providers of the biggest services (SaaS, Iaas, Paas). That said, it appears they have room to compete in the CGI rendering market, and it’s possible many of these render farms might ultimately join Golem.
- The team is well positioned to compete in its initial rendering services market because they have direct experience in it and there is a market need. They also have a track record of delivering on projects and the technical background to be credible.
Product: What is Golem Providing and How?
Golem provides a P2P network of computing power. Users download Golem software, which runs in the background and creates a virtual environment. People who need computational resources can farm out jobs to Golem, which breaks them into smaller discrete tasks and uses the virtual environments to execute them. Computers are compensated in GNT based on how much they contribute to the task.
In the foreseeable future, Golem is an input-output process — people send jobs to the network and get back results–and is more accurately thought of as a distributed supercomputer, not a cloud service. Enabling reliable real-time access to untrusted computers around the world will take much more technical work (as Golem admits).
Why the GNT token is needed
The network does need a vehicle for conveying information from project requestors to the computation suppliers. Because the system is built on top of Ethereum, Golem could have handled payments in Ethereum using smart contracts. However, this would have gotten unwieldy with their micropayment model, which allows for payments to slowly accumulate as computer suppliers perform small tasks. In addition, the team wanted control over the long-term future of their token and they may assign additional rights to GNT. As such, the GNT token is the sole means of payment on the network.
Important details: security, work accuracy, pricing
The team needs to tackle three big challenges in a P2P computing network.
- Security: computation shouldn’t infect the host computer, and the host computer shouldn’t be able to view the computation tasks. The team has a two-part solution for the former. First, each task is test run on the requestor’s computer. Second, tasks run on docker containers, an industry-standard way of running virtual servers: tasks don’t leak out of the docker into the host. At present, there isn’t a great solution to the latter, and they recommend not running highly sensitive/secretive tasks through it. That seems unlikely on an open P2P computation market anyway.
- Work accuracy: all contract labor markets have some way of verifying work accuracy. Given that humans aren’t doing the computations, it’s likely to be less of an issue, but some heavy computation could break computers in processing. In the short-term, Golem will likely duplicate tasks and compare outputs.
- Pricing: network capacity and GNT token values will be fluctuating, creating a dynamic market. How will computers maintain competitive GNT per cycle rates? In the short-term, task requesters will put prices on entire projects, which will be divided among worker computers by the Golem software. Worker computers will accept the bid or not.
The user experience is still a major question–not just the interface but also the bidding system and how lite the Golem software is.
Estimation of capacity
We can develop some scenarios on the potential capacity of a Golem network, relative to other benchmarks. There are multiple ways to measure computation capacity. In the supercomuting world, it is done in FLOPS – the number of floating-point operations per second.
The Market: who needs this and how much might they pay?
In a recent post on the economics of GNT, Andrzej Refulski (COO) cites Gartner turnover and growth estimates for the public cloud services market and its IaaS, PaaS and SaaS components. In sum, they are estimated at $204 billion for 2016 and growing fast.
These estimates include many market segments for services Golem will not be providing for now. Golem will not be a decentralized version of Amazon’s Web Services platform (AWS). As a market platform for CPU resources specifically, Golem will more closely resemble AWS’s EC2 and Lambda services. AWS offers these services as well as a full suite of integrations including data storage and transfer, making it a one-stop shop for infrastructure and software – a level of convenience that will take Golem, indeed Web 3.0 at large, years to replicate.
Moreover, Andrzej himself links to an article describing the current race-to-zero in the cloud storage space. CPUs, like storage hardware, are getting cheaper, begging the question of whether Golem will be able to sustain lower prices than the current establishment, while still being profitable for infrastructure (computer) providers.
Estimated cost of distributed computation
Accordingly, we conducted back-of-the-envelope calculations to explore the value Amazon would place on the computing capacity of Golem, as well as the average compensation per provider, using GFlops as the common unit. See Appendix for details on the methodology.
How much could the Golem network be worth?
The exercise above suggests that it would currently cost $175m a month to leverage the full capacity of BOINC, which gives us an average reward per provider of US$11.35. In the high adoption scenario with increased decentralization of providers, it falls even lower. Taking electricity costs into account, margins would surely be slim or negative, indicating there are economies of scale at play which enable big players to sustain profitability. So, while Golem does somewhat lower barriers to entry for providers, scale and capital barriers remain, suggesting the IaaS market will continue to be oligopolistic.
In addition, no more than 30% of BOINC’s hosts and 20% of its teams have been active at the same time over the past 6 months, suggesting there is no shortage of supply of distributed computing power. Curecoin and Gridcoin are examples of platforms which monetize contributions of CPU and GPU resources to BOINC projects. However, both have sub $2 million market caps 2 years from genesis – not surprising due to the altruistic culture of the scientific research niche.
Another way to understand market for decentralized supercomputing is through the CGI rendering market–Golem’s initial target market. In this market, architects, animators, and designers need data files processed into images and videos. This is computationally intensive. Rumors suggest that each frame in Disney’s Frozen took 30 hours to render (at 24 frames per second of animation). Even lower quality architectural rendering can take 5 – 20 min per frame to render on a home computer.
CGI rendering services is a crowded marketplace with integrations for many base apps and renderers. Brass Golem promises only one of each, Blender and Luxrender. Market pricing is measured in GHz/hour rather than FLOPS/hr. The primary providers of these services are render farms–shops of dedicated server equipment that just focus on rendering jobs. Generally, their pricing model is highly variable, based on service type and turnaround time. Quotes for these services can range from $0.0015 to $0.05 per GHz/hour. This is understandably cheaper than using professional clouds services. One of AWS’s EC2 instances costs an estimated $.067 per GHz/hour, but this is a generalized computing environment, not specialized to a particular task.
Within the context of this competitive marketplace, the profits earned by a provider with an average PC on the Golem network can be easily estimated. The ‘average PC’ mentioned above produces approximately 7.4 GHz per hour. Presumably, for an untested P2P system with minimal service support, the cost should be below the lowest on the market. At $.0015 per GHz/hour, this means $.01 per hour for a PC hooked up to the network–or $8 per month for 100% up-time and utilization. That same CPU requires 51W–drawing 38.5 Kwh per month. At $.1 per kwh, this is almost $4 in power costs per month, not including power costs for the rest of the computer or in how much relative power the Golem software itself uses.
This suggests that there is some room to compete on price, even at the lowest end of the spectrum, but it will depend on the cost of quality control and task coordination. It will probably also require cheap enough electricity for the providers that they’re not worried about losing money.
Ultimately, the determinants of success will be related to non-price competition, as evidenced in the race to zero. Golem will need to provide utility by creating an ecosystem of unique, well integrated applications, just like any other PaaS. Realistically, this will only be possible after the development of features such as external data links, reputation systems and developer toolkit, 3-4 years after the crowdsale.
Competition: Who else is trying to provide services like this?
Apart from the major cloud IaaS, PaaS and SaaS providers as well as public grid computing platforms, Golem will face competition from niche service providers and existing crypto projects.
Cloud providers: As stated before, cloud providers won’t directly compete with Golem yet. Golem also doesn’t call cloud providers their competition: they write that AWS could potentially participate in Golem. This seems less plausible, given how carefully curated these environments are and how much can be charged for service, UX, and modern protections around financial contracts.
Existing render farms: As the most immediate competition, Render farms will probably set the price ceiling on Golem’s CGI services. They are also the most likely to join the Golem network if it becomes successful. These providers have several advantages: fiat payment infrastructure and the ability to compete not just on overall price but through price packages. Tiering prices on Golem will be possible but difficult in a distributed system.
Open-source alternatives: There are plenty of distributed computing projects that already exist today. Two of the most famous are SETI@home, in which user CPUs process radio telescope data looking for signs of extraterrestrial intelligence, and Folding@home, in which user CPUs try to fold proteins as part of scientific research. Both use BOINC software that lets people launch distributed computing projects. There is even Bitcoin Utopia, which uses distributed computers to mine Bitcoin to donate to BOINC projects. This service is very popular and includes scientific-scale computers, home computers, and smartphones. Users aren’t compensated but most projects are hobbies or altruistic causes, like scientific research. It seems likely that Golem could absorb some of this infrastructure, however it won’t be able to compete with the altruistic BOINC for scientific or non-profit projects.
Traditional supercomputing: Traditional Supercomputer manufacturers are starting to explore SaaS solutions, but given the cost of supercomputing, it doesn’t seem likely this market push will meet Golem’s anytime soon.
Other cryptocurrency projects: Golem is not alone in trying to enable distributed computing with blockchain technology. Suchflex monetizes unused computer resources for existing projects: CPU cycles for BOINC projects, GPU resources for cryptocurrency pool mining and storage for Storj.io respectively. Contributors earn FLEX, an asset with voting rights on the Bitshares blockchain that become transferable at the end of testing in March 2017. The project offers a more convenient approach than running Gridcoin and Storj clients while contributing hashpower to Dwarfpool, but no it’s not a perfectly competitive PaaS for supercomputing. Although Maidsafe may accomplish something worthy of the label down the line, their current focus is clearly on data-sharing.
Team: can they pull this off?
Julian Zawisowski (CEO) and Andrzej Regulski (COO), have been working together since 2008–first at Polish economic think-tank IBS, and then at Imapp, a consulting and software development company they founded together in 2013. At Imapp, they went on to contract for the Ethereum Foundation as well as Omise’s Blockchain Lab, affording them the opportunity to work with other blockchain technologies including Factom, Hydrachain and Raiden. Imapp has also been working on Black Vision, a real-time rendering engine for Polish production company Blackburst, making CGI rendering a logical use-case focus for Brass Golem. Piotr Janiuk (CTO) is credited in the whitepaper as the “father of Black Vision”, along with a long list of pioneering technical achievements.
Also worth highlighting is the involvement of Alex Leverington, the architect of devp2p, Ethereum’s underlying P2P protocol, as well as Paweł Bylica, creator of EVMJIT, a library for just-in-time compilation of Ethereum Virtual Machine (EVM) code and active member of Ethereum’s C++ team.
Extended bios for the team can be found at the end of the whitepaper.
Smith + Crown does not collect money or bounties for posting these announcements. All of Smith + Crown’s announcements for ICOs should not be construed as investment advice or an endorsement. In addition, some of the information contained above may be changed after we posted this; we urge the reader to read all source material before investing or passing judgment. If you see anything that has become obsolete, let us know.
Primary Authors: Andreas Weiler, Matt Chwierut
Market Sizing Information: AWS’s largest computing-optimized EC2 instance type, c4.8xlarge, is composed of 36 virtual CPUs with 9-core Xeon E5-2666 chips, boasting performance of 1.232 GigaFLOPS. Using this value, we estimate the number of instances that would be required to match the computing power of Golem at what we believe to be low, medium and high levels of adoption. Subsequently, we use the number of instances and the AWS calculator to get a dollar value for 100% uptime over a month, which is then divided by the estimated number of providers.
Medium Adoption Scenario: At 165 PetaFLOPS, Golem matches the performance of the world’s fastest distributed supercomputer, BOINC. The ~15,286,000 hosts and ~106,000 teams are treated as equal and mutually exclusive providers.
High Adoption Scenario: A mature iteration of Golem coincides with exascale computing in the 2020’s. There is further decentralization of providers, with 150 million giving a provider/GFLOPS ratio of 15% vs. the medium adoption scenario’s 9.33%.