AFTER QUITE A WAIT the latest generation Intel 'Sandy Bridge EP' Xeon E5 dual CPU socket 2011 parts finally came out earlier this month, on 6 March. The pinnacle of the Sandy Bridge 32nm processor generation, these CPUs have maxed out the limits of Intel's 32nm process.
The large 416mm square dies include eight full fledged cores with 16 threads in simultaneous multithreading, 20MB of L3 cache, four DDR3-1600 memory channels, and two 32GB/s QPI channels plus 40 PCIe v3 and four PCIe v2 lanes per chip.
We already covered the huge improvements the record breaking Xeon E5 series brings, starting from doubled floating point peak performance per clock via AVX - and yes, the code has to be recompiled to make use of it - to substantial in-core, cache, memory and I/O bandwidth and latency gains, dual QPI channel links between the CPUs for excellent inter-processor communications, and overall application speed gains of up to 80 per cent despite slight clock reductions compared to the similar TDP Westmere parts.
Other new features to highlight include very low latency Direct I/O, where data from the PCIe peripherals can be written directly into the L3 cache for the CPU cores to use quickly, rather than going through additional main memory writes and reads, as well as vastly improved power and frequency management. More Turbo bins are on offer for higher application performance, but within strict TDP limits where it can manage not just the CPU power usage but that of the whole system as well.
Our Lawrence Latif has looked at the midrange Xeon E5 in the reference standard Supermicro implementation, and now, here is the top speed bin of the new Xeon E5 range, the 3.1GHz E5-2687W - more precisely a pair of them - on the new Asus Z9PE-D8 WS workstation mainboard in its matching tower case.
The Asus board is very different from the Supermicro one, especially with its tuning capabilties. Size wise, it's not huge, but it is the standard dual workstation format and it accommodates eight memory modules, one for each CPU memory channel. This might reduce the total memory capacity to 'only' 256GB if using the highest-capacity 32GB LR-DIMM server memory modules. However it also helps optimise the memory path traces to the maximum, something very useful once memory optimisation kicks in. This much RAM should be enough for most workstation usage models anyway.
There are also four full bandwidth x16 PCIe v3 slots for graphics or other high-bandwidth I/O, with dual slot spacing to accomodate today's thick GPUs, and of course the usual plethora of I/O interfaces, which add the ubiqutous RS232 serial ports to the USB3, SATA3 and other common ports.