Background — High Performance Attached Processors Handicapped By Architecture
The application of high-performance accelerators, notably GPUs, GPGPUs (APUs in AMD terminology) to a variety of computing problems has blossomed over the last decade, resulting in ever more affordable compute power for both horizon and mundane problems, along with growing revenue streams for a growing industry ecosystem. Adding heat to an already active mix, Intel’s Xeon Phi accelerators, the most recent addition to the GPU ecosystem, have the potential to speed adoption even further due to hoped-for synergies generated by the immense universe of x86 code that could potentially run on the Xeon Phi cores.
However, despite any potential synergies, GPUs (I will use this term generically to refer to all forms of these attached accelerators as they currently exist in the market) suffer from a fundamental architectural problem — they are very distant, in terms of latency, from the main scalar system memory and are not part of the coherent memory domain. This in turn has major impacts on performance, cost, design of the GPUs, and the structure of the algorithms:
Performance — The latency for memory accesses generally dictated by PCIe latencies, which while much improved over previous generations, are a factor of 100 or more longer than latency from coherent cache or local scalar CPU memory. While clever design and programming, such as overlapping and buffering multiple transfers can hide the latency in a series of transfers, it is difficult to hide the latency for an initial block of data. Even AMD’s integrated APUs, in which the GPU elements are on a common die, do not share a common memory space, and explicit transfers are made in and out of the APU memory.
My colleague Ted Schadler and I recently completed a six month investigation into social business and collaborative transformation. As the title of the report -- The Road To Social Business Starts With A Burning Platform -- suggests, these complex workforce programs work when there is a compelling motivation to change behaviors among employees, business sponsors, andRead more
[For some reason this has been unpublished since April — so here it is well after AMD announced its next spin of the SeaMicro product.]
At its recent financial analyst day, AMD indicated that it intended to differentiate itself by creating products that were advantaged in niche markets, with specific mention, among other segments, of servers, and to generally shake up the trench warfare that has had it on the losing side of its lifelong battle with Intel (my interpretation, not AMD management’s words). Today, at least for the server side of the business, it made a move that can potentially offer it visibility and differentiation by acquiring innovative server startup SeaMicro.
SeaMicro has attracted our attention since its appearance (blog post 1, blog post 2) with its innovative architecture that dramatically reduces power and improves density by sharing components like I/O adapters, disks, and even BIOS over a proprietary fabric. The irony here is that SeaMicro came to market with a tight alignment with Intel, who at one point even introduced a special dual-core packaging of its Atom CPU to allow SeaMicro to improve its density and power efficiency. Most recently SeaMicro and Intel announced a new model that featured Xeon CPUs to address the more mainstream segments that were not a part of SeaMicro’s original Atom-based offering.
Earlier this week Dell joined arch-competitor HP in endorsing ARM as a potential platform for scale-out workloads by announcing “Copper,” an ARM-based version of its PowerEdge-C dense server product line. Dell’s announcement and positioning, while a little less high-profile than HP’s February announcement, is intended to serve the same purpose — to enable an ARM ecosystem by providing a platform for exploring ARM workloads and to gain a visible presence in the event that it begins to take off.
Dell’s platform is based on a four-core Marvell ARM V7 SOC implementation, which it claims is somewhat higher performance than the Calxeda part, although drawing more power, at 15W per node (including RAM and local disk). The server uses the PowerEdge-C form factor of 12 vertically mounted server modules in a 3U enclosure, each with four server nodes on them for a total of 48 servers/192 cores in a 3U enclosure. In a departure from other PowerEdge-C products, the Copper server has integrated L2 network connectivity spanning all servers, so that the unit will be able to serve as a low-cost test bed for clustered applications without external switches.
Dell is offering this server to selected customers, not as a GA product, along with open source versions of the LAMP stack, Crowbar, and Hadoop. Currently Cannonical is supplying Ubuntu for ARM servers, and Dell is actively working with other partners. Dell expects to see OpenStack available for demos in May, and there is an active Fedora project underway as well.
At its recent financial analyst day, AMD indicated that it intended to differentiate itself by creating products that were advantaged in niche markets, with specific mention, among other segments, of servers, and to generally shake up the trench warfare that has had it on the losing side of its lifelong battle with Intel (my interpretation, not AMD management’s words). Today, at least for the server side of the business AMD made a move that can potentially offer it visibility and differentiation by acquiring innovative server startup SeaMicro.
SeaMicro has attracted our attention since its appearance (blog post 1, blog post 2), with its innovative architecture that dramatically reduces power and improves density by sharing components like I/O adapters, disks, and even BIOS over a proprietary fabric. The irony here is that SeaMicro came to market with a tight alignment with Intel, who at one point even introduced a special dual-core packaging of its Atom CPU to allow SeaMicro to improve its density and power efficiency. Most recently SeaMicro and Intel announced a new model that featured Xeon CPUs to address the more mainstream segments that were not for SeaMicro’s original Atom-based offering.
Emerging ARM server Calxeda has been hinting for some time that they had a significant partnership announcement in the works, and while we didn’t necessarily not believe them, we hear a lot of claims from startups telling us to “stay tuned” for something big. Sometimes they pan out, sometimes they simply go away. But this morning Calxeda surpassed our expectations by unveiling just one major systems partner – but it just happens to be Hewlett Packard, which dominates the WW market for x86 servers.
At its core (unintended but not bad pun), the HP Hyperscale business unit Project Moonshot and Calxeda’s server technology are about improving the efficiency of web and cloud workloads, and promises improvements in excess of 90% in power efficiency and similar improvements in physical density compared with current x86 solutions. As I noted in my first post on ARM servers and other documents, even if these estimates turn out to be exaggerated, there is still a generous window within which to do much, much, better than current technologies. And workloads (such as memcache, Hadoop, static web servers) will be selected for their fit to this new platform, so the workloads that run on these new platforms will potentially come close to the cases quoted by HP and Calxeda.
I'll go out on a limb here and say that businesses are not more social - at least, not in the broad-based fashion people envisioned when we first started talking about Enterprise 2.0 in the heady days of the mid-2000s. How could it be? According to our recent survey of 4,985 US information workers, 28% of the workforce uses a social technology. While you may be thinking to yourself this is a good start, allow me a moment to point out some key differences between Enterprise 2.0 users and the rest of the workforce:
They're your highest paid employees. Over half of this group earns more than $60k a year, compared to just 36% of non-users.
They're the most educated members of the workforce. Sixty-five percent of this group has completed at least a 4 year college degree compared to 55% of the rest of the workforce.
They're the leaders in your office. It's not surprising to see 49% of this group are managers are executives given management's enthusiasm about social technologies. Just 31% of non-users are in similar positions.
The world of hyper scale web properties has been shrouded in secrecy, with major players like Google and Amazon releasing only tantalizing dribbles of information about their infrastructure architecture and facilities, on the presumption that this information represented critical competitive IP. In one bold gesture, Facebook, which has certainly catapulted itself into the ranks of top-tier sites, has reversed that trend by simultaneously disclosing a wealth of information about the design of its new data center in rural Oregon and contributing much of the IP involving racks, servers, and power architecture to an open forum in the hopes of generating an ecosystem of suppliers to provide future equipment to themselves and other growing web companies.
The Data Center
By approaching the design of the data center as an integrated combination of servers for known workloads and the facilities themselves, Facebook has broken some new ground in data center architecture with its facility.
At a high level, a traditional enterprise DC has a utility transformer that feeds power to a centralized UPS, and then power is subsequently distributed through multiple levels of PDUs to the equipment racks. This is a reliable and flexible architecture, and one that has proven its worth in generations of commercial data centers. Unfortunately, in exchange for this flexibility and protection, it extracts a penalty of 6% to 7% of power even before it reaches the IT equipment.
Calxeda, one of the most visible stealth mode startups in the industry, has finally given us an initial peek at the first iteration of its server plans, and they both meet our inflated expectations from this ARM server startup and validate some of the initial claims of ARM proponents.
While still holding their actual delivery dates and details of specifications close to their vest, Calxeda did reveal the following cards from their hand:
The first reference design, which will be provided to OEM partners as well as delivered directly to selected end users and developers, will be based on an ARM Cortex A9 quad-core SOC design.
The SOC, as Calxeda will demonstrate with one of its reference designs, will enable OEMs to design servers as dense as 120 ARM quad-core nodes (480 cores) in a 2U enclosure, with an average consumption of about 5 watts per node (1.25 watts per core) including DRAM.
While not forthcoming with details about the performance, topology or protocols, the SOC will contain an embedded fabric for the individual quad-core SOC servers to communicate with each other.
Most significantly for prospective users, Calxeda is claiming, and has some convincing models to back up these claims, that they will provide a performance advantage of 5X to 10X the performance/watt and (even higher when price is factored in for a metric of performance/watt/$) of any products they expect to see when they bring the product to market.