3 ms·
Not to mention the CPU choice is one of easiest parts of making a datacenter company. I'm not counting heterogenous ones customized to workloads like I've alway
by nickpsecurity 11y ago
Not to mention the CPU choice is one of easiest parts of making a datacenter company. I'm not counting heterogenous ones customized to workloads like I've always advocated: just typical put work or VM's on Intel/AMD CPU's. What they need to do, from marketing to operations to internetworking, to outperform major players in datacenters is huge and not in their direct expertise.
I could see them establishing new revenue by making some custom SOC's for host workloads, I/O offloading, and networking analysis. There's already companies and academics in this space but barely tapping the potential. They could bring in some big bucks putting their engineering talent into some already proven methods. Not going to solve their company's problems, though. Too big haha.
- ksec 11y agoI thought Xeon-D already solve that SOC problem you mentioned. AMD's ZEN is the closest to get Intel some competition. The margin on Xeon product line are huge, making plenty of room for AMD to strike. And since the Cloud is still a growing market, I could see Intel lowering price and AMD still enjoying some success in it. This scenario should make it even harder for Qualcomm, or any ARM maker to set foot in server market.
- nickpsecurity 11y agoI didn't know about it. Thanks for mention. My baseline for this topic of SOC's w/ offloading is Cavium's Octeon line: http://www.cavium.com/OCTEON-III_CN7XXX.html http://www.cavium.com/OCTEON-III_CN7XXX.html The summary should give you an idea. Just offloading TCP/IP or filesystem ops by itself can knock a host CPU down from 90+% utilization to closer to 10%. That leaves a lot of room for compute. Such a situation helps whether the Octeon was the host or a coprocessor for a Xeon host. Now, apply that same principle to text algorithms, concurrency support, compression, SSL, Hadoop's main ops, key mechanisms of deep learning, and so forth. Performance you get per dollar and watt gets crazy. The only strategy I had that could deliver better was integrating FPGA's with CPU's so workload could be split. Intel is already on top of that with Altera acquisition. I can't wait to see accelerator logic on a Xeon. :) Meanwhile, a combo of multicore CPU and libraries that tap into accelerators is much easier for average developer to use. No hardware design experience necessary.
- ksec 11y agoExactly my thoughts on FPGA, it is going to be huge on server market. Although I read AMD is also working with Xilinx, but unfortunately due to the issue with AMD, they will continue to use GF rather then TSMC which is what Xilinx is using for the forseeable future.