Will the servers on ARM make Intel nervous?

    My acquaintance with the ARM version close to the server one began with installing the server version of Ubuntu on CubieTruck and deploying the test site on Drupal. I was wondering if the operation of the engine on Cubie will differ greatly from the operation of the same engine on a server based on Intel Dual-Core G850, which I use for my projects. About the results of the experiment, as well as about the different differences between the ARM and x86 platforms as server hardware, I will tell in my post.

    A set of processor instructions ARM began to emerge in the 80s of the twentieth century. By 1985, the first version of an ARM-based computer appeared, and a year later the second version saw the light. In those years, the market was just forming, companies conquered the market and chose their paths. As a result, Intel became the world leader in the production of processors for user PCs and servers, which has its own serious practical experience and production base. ARM, on the other hand, focused mainly on architecture, cores, and third-party licensing.

    Another difference is in the direction of development. Intel started with more powerful user solutions and took the path of lowering power and energy consumption. ARM Holdings started with weak and cost-effective solutions, such as processors for mobile phones and routers, then for smartphones, tablets and other devices, mainly powered by batteries and often without room for active cooling.

    So, let's look at what advantages and disadvantages are attributed to ARM platforms in comparison with other existing ones (mainly with the industry leader - Intel).

    Production policy

    Intel is the sole developer and manufacturer of its chips and processors. Accordingly, he and only he sets the direction for the development of chips based on his own considerations. ARM by itself does not produce anything, it only sells copyright on the architecture and design of the chips. Chips are produced by all and sundrywho received the license from the copyright holder, ARM Holdings. This allows the technology to be more flexible and adapt to the needs of both a specific consumer and the market as a whole. Manufacturers are given the opportunity to create SoC adapted specifically for their needs (System on a Chip - systems containing on the chip not only the processor itself, but also any additional hardware components necessary to solve the problem). A license fee is charged for each chip issued and is, according to public data, several tens of cents per unit.

    In the case of Intel, the company provides a full cycle, from production to marketing products. That is why the company is quite rigidly focused on sales volumes. How could it be otherwise if the new chip factory, which will become obsolete in a few years, costs about $ 5 billion? Naturally, the cost of such plants falls on the shoulders of future buyers of processors. In addition, the company must produce universal mass solutions and cannot afford to concentrate on niche highly specialized tasks.

    Interesting fact: since October 2013, Intel also has a license for the production of ARM chips. The companies agreed that Intel will build a 4-core 64-bit ARM Cortex-A53. It should also be mentioned that AMD announced its plans to release an ARM-based server back in 2012.

    In January 2014, AMD presented the first samples of the AMD Opteron A1100 Series line of server processors, codenamed “Seattle”. These are 4- or 8-core ARM Cortex-A57 processors containing up to 4 MB of level 2 cache and up to 8 MB of level 3 cache, supporting DDR3 and DDR4 ECC memory with a capacity of up to 1866 MT / s, 8 SATA ports, 2 10Gbps Ethernet ports and other buns . Together with the processor, a platform for developers was introduced.

    I could not find information about the price and ways to get it, because if anyone is in the know, please share. Based on the press release, we can conclude that the company plans to divert the server market from mass solutions, where approximately the same configurations are used for different tasks, to more specialized options.

    Simpler architecture

    ARM supports fewer microprocessor instructions and features than x86. On the one hand, x86 supports more built-in protections, buns and enhancements, on the other hand, ARM has a simpler topology, simpler assembler and requires fewer transistors on a chip. And this, in turn, gives another bonus.

    Smaller size and lower power consumption

    Fewer transistors consume less power and take up less space with the same technology. This became a key factor when choosing processors for mobile platforms in the early stages of their development. The batteries weren’t as good as they are now, and most processors consumed so much power that they needed mains power.

    At the moment, technology is changing not in favor of ARM. Intel is one of the few companies that can fund fundamental development. The size of the transistor also affects power consumption. Intel scientists and engineers have already reached 22 nm and even 14 nm technologies, while most other manufacturers are forced to settle for more affordable 28 nm.

    The X-Gene family released by AppliedMicro is evidence of less power and space.. Their solution based on the ARM architecture incorporated in one chip a 64-bit processor with a frequency of up to 2.4GHz, a memory controller, PCI, SATA, 2x10Gbps Ethernet and other useful things. According to the developer, the solution, adapted specifically for web applications, cloud and blade systems, is close in performance to modern processors of the E3 and E5 series, while consuming 50% less energy and allowing you to place equipment up to 4 times denser due to the fact that almost everything necessary for work is placed on one chip. Thanks to this, owners can significantly reduce TCO (Total Cost of Ownership; equipment maintenance costs). At the moment, the product is available as a kit for developers.

    However, a prototype blade version has also been created. The size of the entire board can be estimated by imagining the size of the memory bar.

    AppliedMicro together with partners also prepared ARM assemblies of popular Linux branches and server software for them, so those who wish can now experiment with the development kit and experience all the advantages of the system.

    Less heat

    This feature follows as a result of less energy consumption and the laws of physics. The principle of operation of modern chips is approximately the same, therefore, the efficiency is also approximately the same. It is logical that chips that consume less energy, with equal efficiency, will dissipate less heat. And this reduces cooling costs and saves space.

    Higher performance

    Actually, the net performance of most ARMs is lower. Only some of the latest developments are catching up with Intel processors. However, if we consider the performance per watt, then according to many data, ARM still wins. Also, according to some reports, ARM wins in the performance / $ ratio. The only caveat is that most of this data is presented in tests of the developers of ARM platforms themselves, so some consider it doubtful. The average layman is unlikely to afford the independent testing of platforms for developers due to their cost.

    Personal practice

    The only device I had the opportunity to practice was CubieTruck. This small but proud device carries a 2-core ARM Cortex-A7 processor, 2 GB of DDR3 480MHz RAM (960MT / s), a SATA port, a micro-SD slot, Ethernet, Wifi, Bluetooth and flashy diodes. in progress. Unfortunately, I did not have an extra SSD, so I had to install the OS on a micro-SD class 10. I suppose that with an SSD the experiment would be better. Also installed Apache, PHP + APC, MySQL.

    As a test site, I used Drupal 7 with automatically generated 10,000 pages of content, of which 100 were randomly selected using the Views module. Testing was performed using the ab utility with a total number of requests of 1000 and the number of competing requests of 10, 20 and 30. Actually, with the number of requests 30 testing failed because ab crashed out timeout. Similar tests were conducted on a server based on Intel Dual-Core G850 with 4 GB of RAM and SATA drives. Some values ​​from the tests are shown in the table below.

    The average time a page was generated by an idle server (not during tests) in both cases was 3-4 seconds, however, Cubie experienced spontaneous instability manifestations expressed in such an absolutely unexpected generation time:

    Thus, although CubieTruck loses the Intel G850, it shows itself quite worthy as for hardware that was not designed as a specialized server one. On the basis of it, small applications may well work, while the cost of equipment (board plus SSD drive) will be about $ 150-200. It is also worth mentioning the fact that the board consumes only about 10-15W. The only problem with the practical use of Cubie is the placement problem. Data centers refuse to colocalize such non-standard equipment, and it is not economically feasible to occupy even a full-fledged 1U. If suddenly now someone from Kiev data centers is reading these lines, who is ready to take part in the experiment and put Cubie in their place for a month or two, I will be glad to cooperate.

    I am looking forward to the releaseCubieboard 8 , which, unfortunately, was somewhat delayed. The device will contain 8 cores with an operating frequency of the order of 2GHz and support more than 2 GB of RAM, which brings it closer to the characteristics of entry-level servers. I hope that in the next few months the developers will complete the final debugging and introduce their brainchild to the world.

    In conclusion

    Summing up, we can say that although ARM-based platforms are still far from mass production, they already show some competitive advantages in comparison with key market players. All that remains is the question of how the players will behave: will they resist, or will they still join the new technology? It seems that the second scenario is becoming more likely. According to some experts, by 2019 ARM architecture will cover up to 25% of the server market, and AMD has already announced that it claims to be the leader in this segment.

    Now the ARM market is in its infancy-experimental state. Someone develops test platforms, someone collects blade systems, someone runs Cubieboard-based computing clusters , which processes files in tens of gigabytes.

    However, now you can seriously think about the commercial application of ARM-based solutions in the near future, as a small French startup NanoXion did, which offered colocation service for Cubieboard . The range of commercial ideas can be enormous - from a micro-dedicated service based on the same CubieTruck, which can be stuffed a lot even in 1U, to actively participate in marketing solutions like X-Gene and AMD Opteron A1100, while the market is still practically free. It remains to turn on the fantasy and correctly enter it.

    Also popular now: