Editor’s Note: This is the third article in our Atom S1200 coverage, following our launch story and the critical response from the founder of micro server business, SeaMicro (now a wholly-owned AMD subsidiary). We also have an interview coming with Mr. Andrew Feldman, ex-CEO of SeaMicro. For now, enjoy in the in-depth analysis from the feather of our semiconductor analyst, Mr. Ashraf Eassa.
On 12/11, the head of Intel’s DCG group, Diane Bryant, publicly announced the Atom S1200 series processors, formerly codenamed "Centerton". This is a brand-new server-oriented system-on-chip that uses the 32nm "Saltwell" cores coupled with a single-channel DDR3 memory interface. The lowest TDP goes to 6W, although there are higher clocked, higher TDP parts available that go up to 14W.
Intel Atom S1200 Server Processor
Immediately, the problem here is obvious: the ultra low voltage Xeon E3 part based on the 22nm "Ivy Bridge" at 17W has a very significant performance-per-watt advantage over any of the Atom parts. This is due in part to the fact that the 22nm Xeon chips are built on a much more power-efficient process technology, in addition to a much slicker and modern micro-architecture. To put it bluntly, the Xeon has a much better memory interface, significantly more refined and powerful cores, and a much more refined system architecture that includes a very fast L3 cache. It is as though the products came from two very distinct and very unevenly matched companies.
However, it is clear that a 5 year old micro-architecture built on a last-generation process node would not be competitive with the best mainstream designs available. Intel has traditionally invested much more heavily into the Core/Xeon side of things because that is where the money is. The S1200 seems to be an attempt to get a design out onto the market before any of the ARM vendors do in order to claim a first-mover advantage. It’s not going to set the world on fire, but it is a statement of intent in the space. However, a major positive that I took away from the launch is that, going forward, Intel will take low power cores much more seriously.
Atom Will Not Be Held Back
Atom was quite obviously held back for financial reasons. The return on investment of developing the "big" cores has traditionally been much more attractive than on the "Atom", which was designed to go into low cost, low performance net-books. It is a miracle that the 5-year old design is still competitive with the latest ARM-based designs such as the Cortex A9, the Qualcomm "Krait", and even the Cortex A15 in integer performance.
However, as the secular growth story in tablets and smartphones continues, and as micro-servers could potentially emerge as a serious threat to the Xeon product line, it is critical that Intel not hold back performance and financial investments into its "Atom" core. In the ARM-compatible micro-server space, Applied Micro, Cavium, AMD, and others are making significant strides, enabled by the ARMv8 architecture. While performance characteristics on these upcoming products are not available and while these companies are likely to be hamstrung by last-generation process technology, they also have no vested interest in keeping a higher margin, higher ASP product line alive.
2014 Intel Valleyview-T SoC is based on the 22nm Silvermont core, while Centerton/Cloverview utilize 32nm Saltwell x86 core
It seems that Intel is making the correct strategic move by not limiting the capabilities and performance of its upcoming "Silvermont" processor core. It has been revealed in a number of leaks that "Silvermont" will feature a fully out of order execution engine, a shared 2MB L2 cache between pairs of cores (similar to the Core 2 Duo generation), be 50-60% faster than "Saltwell" in single-threaded performance, lack hyper-threading, and be imbued with quite a bit of interesting power-gating logic. Also, according to Intel’s own website, next generation "Atom" will feature < 1mW idle power in its mobile system-on-chip incarnations.
The more important thing is the actual system design. Right now, the folks touting the ARM micro-servers with unique interconnect fabric have very interesting technology. While the next generation "Avoton" micro server will feature a much higher level of integration according to management, the current generation "Centerton" is fairly modest on this front, sporting only PCIe lanes built onto the processor. Applied Micro’s "X-Gene" will sport Ethernet, PCIe, and storage connectivity. The rumors surrounding "Avoton" state that it, too, will sport these features, but nothing has been confirmed by Intel other than that the fabric will be integrated onto the processor.
Luckily for Intel, X-Gene and the other ARM designs will not see widespread availability until late 2013/early 2014 at the earliest, which is about the timeframe that "Avoton" is slated to launch. Should Intel properly leverage its 22nm process advantage as well as its experience with high efficiency, high performance CPU design to really give the new Atom core a strong boost, then the competitive landscape will likely shift in. Should Intel, however, hamstring Atom to make Xeon look better, the market may very well move in favor of the ARM-based vendors in the micro-server space.
Sales and Margins
The micro-server space is all about density, so when it comes to revenues, Intel will not see a direct replacement of a $200 low-power Xeon with a $54 Atom. Instead it is likely to be 3-5 $54 Atoms replacing a single low power Xeon, which could translate to flat-to-higher revenues. On the margin front, however, it really depends on the die sizes of the Atom micro-server parts. Better aligning the Atom chips with the Xeon chips on the process technology side will help, but the Atoms will have additional transistors used for tighter levels of integration (although at
the press conference, Ms. Bryant noted that the Xeon line would also see higher levels of integration), which could negate the transistor count savings from using a smaller core.
Management seemed optimistic that margins on the Atom products would be good and that they would have no problems if Atom did well in micro-servers, but until the financial results start rolling in, and until Intel is selling a clearly better product (to keep pricing pressure from the ARM guys at bay), this will need to be watched closely.
The micro-server space is getting interesting. The ARM instruction set moving to 64-bit has enabled a slew of competitors to come in and try to take a piece of the micro-server solution pie. Intel will vigorously defend its server market share, but it will need to fire on all thrusters to stay at the top of the competitive heap. That may just mean eating into Xeon sales at the low end, but when it comes down to it, Intel will either cannibalize its own sales, or Applied Micro, Calxeda, Nvidia, and others will do it for them.