AMD is at a pretty massive deficit for single threaded power, but when you look at virtualization, TCO and performance per watt, or the more theoretical performance per watt per dollar, are the metrics that rule. In this market, system price can be pretty persuasive and single threaded performance matters less than aggregate values. Why? Because if you are running 10+ VMs per box you are not really that concerned with the performance of any given thread just overall capacity.
In light of that, AMD is claiming a win over Intel, but take those numbers with a big grain of salt. Why? The 30% claimed cap-ex reduction and $130K saved per rack are for servers with dissimilar counts of very expensive memory. The 16GB DIMMs in the AMD machines are between $250 (Kingston) and $500 (most others) each, the Opteron has 8, Intel box has 16. The Intel servers in question score 11.05 on VMmark 2.1 and cost a bit over $15K, the AMD boxes score 8.31 and cost almost $7K. AMD claims about $6.5K saving per box, and that $130K is about right for a rack of similar VMmark performance.
Now if you back the memory out, the Intel boxes cost between $2K and $4K less apiece, but would also likely lose performance. If they dropped to the same 8.something as an AMD server, and cost an an average $3K less for similar RAM sizes, that still leaves AMD out in front by over $60K a rack, not bad, but not six digits savings. It also goes to show AMD can do a decent job with less memory, or at least it heavily points to that conclusion.
The same holds true for other tests, they are stretched but still hold up if you pick at them. We said the same about the server FirePros, TCO tells the story, and that is a grey area at the best of times. Once you sift through the fine print it looks like AMD does still have a niche market for Opterons where they can do pretty well. Luckily for them that niche is both large and growing, but it will still be a serious uphill battle for them. That said, now you know what their sales pitch is centered around for servers.S|A
Latest posts by Charlie Demerjian (see all)
- Intel kills off the 10nm process - Oct 22, 2018
- HyperX ships it’s 60 millionth enthusiast memory module - Oct 15, 2018
- Bittware/Nallatech water cools 300W of Xilinx FPGA - Oct 12, 2018
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018