On Wednesday we had the opportunity to talk to AMDs Andrew Feldman, founder and former CEO of SeaMicro who now runs the Data Center Service Division (DCSS) at AMD. We got some insight about their current business, the needs of their customers as well as a glimpse what products are around the corner in 2013.
As we reported earlier, AMD circulated a press release where it critizises Intel over their Atom S1200 launch. On the call Andrew made it clear that there are positives and negatives about the Intel launch. On the positive side he sees Intel?s commitment to this new product category of microservers. With the industry leader giving this segment the needed attention, it allows the ecosystem to grow more quickly.
Mr. Feldman explained that a few years back he approached Intel with SeaMicro to put Atom CPUs into servers. Back then, Intel was not very amused about the idea, to say the least. Eventually the company was allowed to design a microserver board based on Atom CPUs, which was quite a success even though the chip had a lot of shortcomings when it came to the server segment. Likewise from their point of view, Intels recently launched Atom S1200 is not a great product. Especially memory support is an area that is largely inadequate for some of the key workloads it is supposed to cater for. However, he admitted at least they have some product, whereas AMDs dedicated microserver chips based on ARM are only due in early 2014.
Current Business and Impact on AMD
When asked how their current business is going, Andrew explained that right now they are selling more Opterons. However, before the Opteron part was ready, SeaMicro sold a lot of the existing Xeon designs. Those are still being sold and some of their customers are very happy with them. The advantage of the Opteron part is twice the amount of RAM compared to the Intel-based offerings.
Andrew explained that many customers in the cloud like to have lots of RAM, so the Opteron has a natural advantage there. He even admitted that the raw CPU performance is a bit higher on the Xeon part, but their customers prefer to trade in some of that CPU power to get more RAM. In certain cloud applications, massive amounts of RAM are used to cache data for instant access, where fetching them from storage (i.e. arrays of SSDs or HDDs) would take too much time.
Another advantage of having a foothold in the system business is more close interaction with the actual customers, according to Feldman. They are delivering feedback much more quickly to the respective engineering units which allows to change roadmaps. In the end future chips should turn out to be better suited for their customers needs.
Smaller AMD x86 Cores Getting into SeaMicro Microservers
While at this point, the executive couldn’t directly talk about upcoming products he pointed out that the company has a new offering in the pipeline:
"We haven’t yet announced or made any statement about what we will be doing in the second part of 2013 with a smaller core [based on] x86."
Andrew was unable to provide us with further details at this point, but it seems to be clear that the chip we are talking about is a descendant of AMD?s upcoming Jaguar design. The Jaguar microarchitecture was first presented at the Hot Chips conference at the end of August this year and is poised to replace the Bobcat microarchitecture used in AMDs low power APUs in 2013. Jaguar brings the featureset in terms of instruction set extensions to the level of contemporary CPUs with the support of AVX, AES and some others. The instructions per clock (IPC) was improved a bit as well and the chip should provide slightly higher frequencies. Jaguar-based designs will feature up to four cores with a shared 2MB L2 cache. The physical address width was increased from 36-bit in Bobcat to 40-bit in Jaguar. This means that it can address up to 1024GB of memory (Bobcat: 64GB). Of course the actual limits are often lower and depend on other factors like available memory technology and the amount of modules employed.
Given that Jaguar-derived chips feature a lower TDP compared to Opteron or Xeon CPUs, this bring back to mind the board layout of the original Atom-based cards SeaMicro designed. Those boards had four to eight Atom chips installed and each had to be paired with a chipset. Given that Jaguar is actually a SoC, the chipset is potentially no longer necessary. So what you get when you pair those Jaguar SoCs with the SeaMicro fabric chip is a very dense, power efficient solution that should appeal to the needs of cloud customers. Unlike with the Atom parts, plentiful amounts of memory are supported.
AMD plans to launch Jaguar-based APUs in the first half of 2013. We actually expect the company to show it off at CES in Las Vegas, which takes place from January 8th to 11th. In our opinion it would only be natural to use the small core design aimed at tablets and netbooks also in microservers and this hint we got from the former CEO of SeaMicro that currently runs this business unit at AMD shows that AMD is actually committed to do so. We don’t know at this point whether AMD will move this chip into the microserver business in general or whether this is only related to the offerings of their SeaMicro business unit.
Andrew concluded the call with the following statement: "What you are beginning to see is the unfolding of a big strategy. The acquisition of SeaMicro, the announcement of an ARM part, these are just a few of the chess pieces that are moved forward and you get a chance to see the outline of a strategy."
One thing for sure, we will keep you posted as soon as we learn about new chess pieces in their strategy.