Too bad that Intel is having no significant competition on the server/workstation front, so no need to rush either – the Ivy Bridge based Xeon E5 and E7 processors are finally rolling out towards the end this year, but with more cores than originally expected. Meanwhile, the Xeon Phi MIC/co-processor/accelerator gets segmented into more SKUs and a new stepping that clocks things up to 1.238 GHz (from 1.1GHz) with roughly the same TDP and form factor.
Xeon EN/EP E5-xxxx v2 series
Wondered why Intel is not exactly rushing with the new Xeon (or, for that matter, any Socket 2011 CPU) product announcements? Well, in the absence of any competition at present, the push to accelerate all the expensive product releases loses steam.
Nevertheless, in a few months, you'll finally see the first of the new Xeon offerings based on the Ivy Bridge core, well over a year after the mainstream desktop and mobile ones. The difference? Along with the usual slight IPC increases, the initial Xeon E5-2600v2 series parts will be 12-core monsters (+25% from Sandy) with 30 MB cache and quad-channel DDR3-1866 server memory paths per socket, and likely up to 3 GHz clocks, with a bit extra turbo frequency oomph for the higher TDP workstation parts. The 22nm FinFet '3D transistor' process should help a bit with the power consumption, and hopefully there shouldn't be major leakages there. Otherwise, it is the standard Socket LGA2011 platform, so a microcode bios update should allow upgraders to stay with the same motherboards.
Much later in the year, you should see the E5-4600v2 series, basically the same as above except enabled for 4-socket operation, where those dual QPI links make a square interconnection layout between the CPU. At the same time, the E7-4800v2 and 8800v2 – 15 core IVB-EX enterprise Xeon processors, should arrive as well, finally replacing the Westmere-EX workhorses from the past two years. These, however, will run at substantially slower clocks, likely no more than 2.4 GHz. Their memory, however, will be upgraded to DDR3-1600, and the chipset will be the same Patsburg one you see on the current Romley dual Xeon platform, with RAID TRIM finally enabled in OPROM.
Sorry to kill the dreams of overclockers – just like Sandy Bridge-EP, no unlocked multiplier SKU.
Xeon Phi – now with three SKUs and a new C0 stepping
Noticed the ongoing HPC & BigData war between Intel and Nvidia on the Tesla GPU vs Xeon Phi accelerator deployments worldwide? You've read here first even about their upcoming 2013 world's largest Xeon Phi deployment – the next generation of Tianhe machines – in Guangzhou, which will have close to 100,000 of these beasts together with the same number of Xeon E5-2600 v2 IVB-EP Xeons.
Now, to address Nvidia's DP FP performance, Intel seemingly has two new SKUs of Xeon Phi – as the yield improves, so do the obtainable frequencies, isn't it?
We found out about the last month's brand new C0 stepping of Xeon Phi, and the associated two new SKUs, operating at 1.238 GHz with the same 61 cores and 31 MB cache as the previous top part, the 1.1 GHz bin. the 7120X and 7120P (no cooling and passive cooling respectively), should provide in excess of 1.2 TFLOPs DP peak performance, yet keeping the same 8 GB memory capacity and 512-bit memory path advantage over the Nvidia and AMD accelerator offerings on the market currently. The GDDR5-5500 memory speed stays the same as the 1.1 GHz part, though, as well as the rated TDP of 300W.
Since this is a 22 nm part, Intel likely could still squeeze a bit more juice out of it during this year, likely through these same 'step by step stepping' approach. Now, will the Chinese ask for this part now? Almost the whole world's production of Xeon Phi is going there at the moment… (ED: would you rather the Wall Street guys get it?)