With the recent announce of the new AMD “Napples” processor, a few things have changed in regards to options for Storage Spaces Direct. Let’s have a look to see what’s this new CPU is about.
A few key points:
- Between 16/32 threads and 32 cores/64 threads per socket or up to 64 cores/128 threads in a 2 socket server
- Intel Skylake is “only” expected to have 28 cores per socket (** Update 2017-03-19 ** There are now rumors of 32 cores Skylake E5 v5 CPUs)
- 2TB of RAM per socket
- 8 channel DDR4
- Bandwidth is expected to be in the 170GB/s range
- Intel Skylake is expected to only have 6 channel memory
- 128 PCIe 3.0 lanes PER socket
- In 2 sockets configuration, it’s “only” 64 lanes that will be available as the other 64 are used for socket to socket transport
- In other words for S2D, this means a single socket can properly support 2 x 100GbE ports AND 24 NVMe drives without any sorcery like PCIe switches in between
- That’s roughly 126GB/s of PCIe bandwidth, not too shabby
Here’s an example of what it looks like in the flesh:
With that kind of horse power, you might be able to start thinking about having a few million IOPS per S2D node if Microsoft can manage to scale up to that level. Scale that out to the supported 16 nodes in a cluster and now we have a party going! Personally, I think going with a single socket configuration with 32 cores would be fine sizing/configuration for S2D. It would also give you a server failure domain that’s reasonable. Furthermore, from a licensing standpoint, a 64 cores Datacenter Edition server is rather pricey to say the least… You might want to go with a variant with less cores if your workload allows it. The IO balance being provided by this new AMD CPU is much better than what’s being provided by Intel at this point in time. That may change if Intel decides to go with PCIe 4.0 but it doesn’t look like we’ll see this any time soon.
If VDI/RDS SH is your thing, perhaps taking advantage of those extra PCIe lanes for GPUs will be a nice advantage. Top that with a crazy core/thread count and you would be able to drive some pretty demanding user workload without overcommitting too much your CPU and while also having access to tons of memory.
I’ll definitely take a look at AMD systems when Naples is coming out later this year. A little competition in the server CPU market is long overdue! Hopefully AMD will price this one right and reliability will be what we expect for a server. Since it’s a new CPU architecture, it might take a little while before software manufacturers support and optimize for this chip. With the right demand from customer, that might accelerate the process!