SEAMICRO IS UPDATING their Atom based server system with several much needed features, all of which stem from a new Atom SKU. These new updates simplify, speed, and add headroom to the 512-core rack mount SM10000 box.
The short story is that the SM10000 is a 10U box filled with cards that contain 8 Atom based systems. Each of these ‘systems’ was a 1.6GHz Atom CPU, chipset, and memory, connected to a SeaMicro ASIC via PCIe. The Pineview Atoms only had one PCIe2 lane off the chipset, and it was running at half speed. Not optimal for a server’s only breathing tube, but in this case, you take what Intel offers.
The new simpler SeaMicro CPU board
A few months later, Intel introduced, or in this case will probably introduce really soon, a new N570 dual core Atom. It replaces the older Pineview Z530 parts, and is basically the same tried and true dual core Atom on the market, but with the chipset replaced with the new Tiger Point chipset. Tiger point was supposed to ship with Oaktrail, but that part is delayed yet again, so Intel is playing mix and match.
From SeaMicro’s point of view, there are three important things that the N570 brings to the table, a full speed PCIe2 1x slot for unrivaled speed(1), 64-bit addressing, a 4GB memory limit, and VT-x extensions. All of these things, except the memory limit, have been in the current Atom chips, but Intel felt the need to keep them fused off, for our protection of course. It is amazing what competition will do to the market.
The new boards all lop the CPU counts in half though, there are now 256 sockets, but with the doubling of core counts, the total number of cores stays at 512. Memory per socket doubles, as does external bandwidth. Net result, the new boards are the same speed, have the same per-core memory, same per-core bandwidth, and as far as gross numbers go, are basically a wash with the older system.
Why did SeaMicro bother then? The component count went down by almost half in the critical pieces, dropping cost and power. Half as many CPUs, half as many chipsets, and half the DIMM count adds up quickly when you are looking at the sheer number SeaMicro can pack in. 15% power use goes out the door too, not a trivial amount. 64-bit OSes are also a must for any self-respecting server, as is virtualization, both of which the ‘new’ CPUs are finally allowed do.
One thing most people don’t think about is that while the per-core numbers don’t move much, the potential available to each core does double. Any CPU that has access to the shared resources can use it all. An OS instance loaded on a CPU can get by with one copy instead of two, dropping some memory overheads dramatically. On top of that, if you need the full 4G for a single app, it is available, something you couldn’t do with the old system. PCIe bandwidth is the same, it can be doubled if the other core is not doing much I/O, so there is the potential for some real wins here.
Tiger Point connects to the SeaMicro ASIC, which talks to the backbone fabric. From the ASIC on out, there are really no changes to the system, it is the same chip, and it plugs in to the same box. If your application is suited for an environment like the SM10000, it will be a bit faster with the new CPU boards, and a whole lot more efficient. If it is not suited to the paradigm, well, go look at traditional servers.
The end result can be best described as adding a lot of efficiency, and uncorking a few painful bottlenecks. Nothing radical, and quite a bit more than incremental improvement, but it is fairly nuanced. With any luck, Oak Trail will be out in short order, and we will see the SeaMicro SM10001 with a lot of new features, and the current open board space put to interesting use.S|A
(1) I did it, I said that with a straight face!
Latest posts by Charlie Demerjian (see all)
- Intel talks about Knights Landing architecture - Mar 31, 2015
- Gigabyte’s ARM servers represent a sea change in the market - Mar 30, 2015
- VESA shows off working USB-C Alt Mode with Displayport - Mar 24, 2015
- UzBrain’s Rail Gun turns a toy weapon into an FPS controller - Mar 18, 2015
- Fotonation uses computational imaging to focus faces - Mar 17, 2015