TIS THE WEEK for cloud products, and AMD is not one to miss the opportunity to attach lots of new Opteron 4100s to the cloud. The chip itself is half of a Magny-Cours/Opteron 6100, but the cloud is where AMD thinks they will sell the most units.
We showed you the boards from Supermicro a few weeks ago, and now there are officially CPUs to put into the waiting socket C32s. Since the chips are literally half an Opteron 6100, they don’t bring anything new to the game that the bigger dual-die part didn’t, other than socket count. The new parts are for 1 or 2 sockets, the 6100 was 2 or 4 socket. Advances over the Socket F Opterons include C1E sleep states, integrated APML (Advanced Platform Management Link), and 1.35v DDR3 support.
The raw specs
The chips themselves come in 4 or 6 core variants, and range in speed from 1.7GHz to 2.8GHz. Once again, we have three power ranges, ‘normal’ chips at 75W, HE at 50W, and EE at 32W, all ACP, not TDP. Prices range from $99 for the 4122 4 core 2.2GHz model to $698 for the 1.8GHz 32W 4164EE. 6 cores in 32W is a pretty amazing feat no matter how you look at it, and power constrained data centers will eat this one up.
High on the list of questions for this chip and socket is, “Why bother?”. No, that is not a snide remark prompted by high end sports cars paid for by Intel PR mysteriously showing up in my driveway, why would you want a 1S or 2S 4100 system over a 1S 6100? There are two answers to that question, speed and availability.
Speed is the main concern here, the 4100 can hit 2.8GHz, the best a 6100 can do is 2.3GHz, but it has double the core count. Two 2.4GHz 4176HE’s would suck 4W less than a single 6176SE at 2.3GHz, and two 2.8GHz 4184s pull 150W, so it is far from a clean kill, but data centers actually care about the odd watt here and there. If you are only concerned about raw single threaded performance, two 4100s would definitely do a better job. Then again, not seeing anything red and aerodynamically sculpted in my driveway, I won’t mention the word Xeon.
The other issue is that there aren’t any 1S 6100 systems out there, they start at 2S. This means you are getting the equivalent of a 4S 4100 system with a 2S 6100, amazing how that product differentiation worked out, almost like they planned it that way. 2S 6100 systems are overkill for a lot of customers, so that is a good reason for the 4100 to exist.
Ironically, the largest target market for 4100 systems are the big cloud compute guys who buy machines by the semi-truck load, and order CPUs and memory by the pallet, not by the part. There are 4100 systems floating around, but the overwhelming majority will be custom boards on custom systems.
If you are ordering 10,000 servers a month, you probably have a workload in mind for it, and pretty exact specs as to what you want on the board and what you don’t. Supermicro, Tyan and even Dell will be more than happy to make a custom machine to your exact specs. If you want a 1S 6100 system, they will make it for you, there is no technical issue preventing it.
And that is where most 4100s are going to end up, boards made for an exact purpose, with specs that seem very odd to the average user, but save the data center a few dollars per unit, and a few watts a day. For them, things like the 56xx chipset’s ability to clock gate to zero the SATA and USB ports in the BIOS are a big big deal.
To prove a point, AMD points to a SPECpower_ssj2008 score from ZT Systems. The 2S 4100 system has a the lowest power use from 10-100% loads, but not at 0%. Power gating is a good thing, and AMD isn’t there until Llano. Still, how many data centers run machines at 0% load? Most of them turn the machines off if there is going to be long idle periods.
Looking to the future, the new Opteron 4100s force a change from the older DDR2 Socket F boards to newer DDR3 Socket C32 boards. They will be socket compatible with Bulldozer whenever that comes out, so it is ‘futureproof’ to an extent, and begins catch-up on the management side for AMD.
This isn’t to say there won’t be ‘pedestal’ servers from everyone that matters, Dell, Gateway, HP, Tyan, Supermicro, Gigabyte and MSI have signed up already, so you can expect boxes for smaller business from just about everyone that matters. There is even an embedded 4100 from day one, so you can be assured of availability in 2017. Best of all, if you don’t want to buy 10,000 units a month, that is OK, they will sell them one at a time if you ask nicely.S|A
Latest posts by Charlie Demerjian (see all)
- Memblaze makes flash aware wide RAID solutions - Aug 22, 2014
- Bland title for a good GPU story on AMD and Nvidia - Aug 20, 2014
- Analysis: What does the automotive market look like for Nvidia? - Aug 19, 2014
- Analysis: Nvidia refocuses it’s Tegra program - Aug 15, 2014
- AMD’s Fiji GPU not arriving soon enough - Aug 14, 2014