HPE announced Silicon on-demand last week — the switching on and off of Xeon cores and Optane Persistent memory to make scaling up CPU power and Optane capacity as easy as a key press or two. There were few details, and we asked Intel and HPE some questions about how it worked.
An Intel spokesperson kindly provided some answers.
Blocks & Files: Is this switching-on of CPU core capacity at the per-CPU level, or the per-core level? If at per-core granularity, which Intel CPUs are supported and how does it work? For example, in a 40-core CPU does it start at one, add cores in units of one and stop at 40?
Intel: “Intel scalable Xeons offer the capability to enable/disable cores at a per-core level; However, each OEM uses this capability as they see fit. While technically, a 40-core scalable Xeon can go from one core all the way to 40 cores, not all combinations may be enabled by an OEM.”
Blocks & Files: Is this switching-on of Optane Persistent Memory at the DIMM level or sub-DIMM level? If it is at the sub-DIMM level, which Optane PMem products are supported and what is the unit of granularity please?
Intel: “Optane Persistent Memory allows configuration of the memory within each DIMM via established HW interfaces. The capacity of memory of each DIMM designated as 2LM (Two-Level Memory) can be configured per DIMM as a percentage value. For example, a 256GB DIMM can be configured to have only 75 per cent of the capacity exposed as 2LM and the rest either reserved or as App Direct memory. On top of this, OEM BIOS has the ability to not map an entire [DIMM]. At this point, the capabilities are within a single system but also regular DDR DIMMs on a per slot/channel basis).
“An OEM may choose to use any combination of this configurability to offer their platform specific capability.”
Blocks & Files: Is this Silicon-on-demand technology unique to HPE and Intel, or will Intel make it available to other system OEMs?
Intel: No other OEMs have announced a Silicon-on-demand capability.
Blocks & Files: Are Intel’s hyperscaler customers using it?
Intel: “No hyperscalers have announced they are using silicon on-demand capability; While not the same as Silicon-on-demand, phoenixNap has previously announced a vertically scalable solution using 2nd-generation Intel scalable Xeons.”
Blocks & Files: Is there a composable systems angle to this technology?
Intel: “Currently the capabilities are within a single system.”
It’s apparent that both the Xeon core enable/disable feature and partial Optane DIMM reservation are Intel features that other OEMs could use if they wished. So far no one else apart from HPE is using these features.
It seemed to Blocks & Files that Xeon core switch on/off and Optane PMem capacity switch on/off could be theoretically used to compose server systems more precisely tailored for specific workloads, but Intel indicated that this option is not being considered at the moment. Nevertheless, once the CXL bus/fabric arrives the possibility might become real.