Blog AMD’s sudden purchase of SeaMicro came out of the blue on Wednesday afternoon (West Coast US time) and was followed by a conference call with CEO Rory Read, SeaMicro co-founder Andrew Feldman and Lisa Su, AMD’s newly dubbed head of Global Business Units. This is a purchase that has the potential to put AMD back into the forefront of server technology – a place it hasn’t visited since the early Opteron years.
For the uninitiated, SeaMicro makes very small, very dense systems typically populated by dual-core 64-bit Intel Atom processors. They fit four sockets/servers, each with 4GB RAM, onto a PCIe board about the size of a typical PC sound card (3”x2”) and still have space for “more cool stuff,” as the company put it in a 2011 analyst presentation.
Thirty-two of these boards combine to make up the SeaMicro SM10000-64 system, which gives customers a 10U box with 512 total cores running at 1.66Ghz, a TB of RAM, 64 SSD or SATA drives, and 64 1GbE (or 16 10GbE) uplinks – all connected with SeaMicro’s proprietary fabric interconnect. For webby workloads, the full system uses about a quarter of the power and space of traditional servers while providing equivalent (or better) application performance.
The secret sauce in SeaMicro’s server stew is its server board ASIC and the interconnect fabric that ties all of the server boards together.
The ASIC reduces the server components to exactly three: the CPU, RAM and the ASIC, radically reducing the size of each server and allowing SeaMicro to put four sockets into a notecard-sized space.
Its 1.28Tb interconnect fabric gives high bandwidth – about 1Gb per core – and has some proprietary functionality (in the latest version) that allows it to turn off individual components, such as the SATA, USB, or graphics controllers in the Xeon chipset or processor. The power savings are small when looking at a single chip, but significant when the numbers get into the thousands or tens of thousands.
AMD’s Rory Read said that purchasing SeaMicro is an example of AMD “skating to the puck” to address today’s needs for highly dense, less complex, power-sipping systems that are optimised for specific workloads. I had thought that the age of hockey metaphors ended when Scott McNealy left the business, but I guess they are making a comeback, since Read used that phrase a few times.
While this is all pretty cool technology, I think that SeaMicro’s best move was to design its ASIC so that it could take virtually any common CPU. In fact, SeaMicro’s last big announcement introduced its SM10000-EX option, which uses low-power, quad-core Xeon processors to more than double compute performance while still keeping the high density, low-power characteristics of its siblings.
AMD, in their initial “Puck you” to Intel, will exercise this option by de-Atomizing and de-Xeonizing the SeaMicro boxes as soon as possible.
AMD says that it will be putting Opteron processors into SeaMicro boxes by the end of the year, but will be looking at other options in the future. AMD was adamant that it is not going into the server business, but that the SeaMicro systems will be building blocks that AMD OEM partners can use as a foundation for their own unique end-products. When you think about it, this is a very interesting opportunity to mix and match different processor combinations to optimise the resulting systems for different tasks.
Just within AMD’s own products, we could see SM variants with higher core counts, using their Bulldozer-based quad- to 16-core processors for serious processing power. They could also use their Llano APU (combination CPU/GPU) processors to push faster number-crunching at lower power levels. I’m not sure if these different servers could co-exist in the same SeaMicro chassis, but I can’t think of any obvious reasons why they could not.
Pushing the envelope a bit further, I can see where ARM processors and possibly even AMD Firestorm GPUs in future SM servers could be combined to yield extremely efficient HPC or Big Data optimised systems that could run price/performance rings around current systems.
One potential limitation, at least in the short- to medium-term, is that all communication outside the SeaMicro chassis is confined to either 64 1Gb or 16 10Gb Ethernet connections. Infiniband can deliver more bandwidth at significantly lower latency than current 10GbE implementations; when the node count rises, this could limit the SeaMicro box to mid-sized installations that use small numbers of chassis to provide a few thousand sockets. This won’t get you to the upper reaches of the Top500, but it’s a sizeable market to be sure.
If AMD plays its SeaMicro cards right, it could use this new technology as a vehicle to drive it back into the server industry limelight. The flexibility and openness of the SeaMicro architecture gives it the ability to put together many different configurations that are optimised to address customer facility pain while providing equal or better performance at lower acquisition cost. Just with its own processors, the new AMD SeaMicro systems will allow AMD OEM partners to deliver highly efficient systems that are difficult for current Intel-based systems to touch.
In their current form, these boxes address niche needs for huge scale-out web serving. It’s not a tremendous number of customers when compared to the audience for general purpose 2-way or 4-way systems, but the numbers of systems that these customers buy is truly massive.
This is going to put AMD back on everyone’s radar again (particularly Intel’s) – just like the introduction of the 64-bit Opteron did back in 2003. ®