Watched the first day IDF Keynotes? One point must have been noticed from Kirk Skaugen's and David Perlmutter's talks – it's LATE 2013 Haswell in the 'Core' processors… and therefore even later in high end Xeons.
Now, it seems Kirk Skaugen have gotten the short end of the stick by moving from the ever-strong, ever-ready Enterprise division, to the former-glory Client division within Intel. The first one has as smooth riding as ever, with every piece of Xeon and related server platforms selling even before they are made in fab, at four-digit US$ prices. The Client side is squeezed between (somewhat over-empasised) diminished desktop, under-performing UltraBook and the growing ARM and possibly MIPS threat on the mobile front. Of course, compared to say AMD, it still dominates by a mile, at least until next year.
Despite all that, the Haswell-focused – and, mind you, power- not performance-focused at that – keynotes imply late 2013 Haswell delivery in those 'Core' parts for UltraBooks and desktops, compared to the expected early 2013. Whether Intel needs more time to fine tune the power states and bring down the power consumption to 10W on UltraBooks, or simply there is no desktop & mobile PC CPU competition worthy retiring the existing Ivy Bridge parts early, is another question. CPU performance wise, keep in mind that, unless you use its new instruction enhancements, the Haswell core is expected to be only 10% faster clock for clock on average than Ivy Bridge for the existing codes. The GPU should be a little extra faster, even for the mainstream GT2 version, but still, these are not massive improvements – especially since the desktop GT3 version with L4 cache for the GPU seems to be off chart now… OK, Linpack would gain something like three quarters extra if you re-optimise it for Haswell FMA fused multiply add instructions.
So, in that sense, both the delay and the power-focused talk make sense. As there won't be a noticeable general application performance – same as in IVB vs SNB – there's no substantial performance benefit for the user from switching the generations, but there will be enough power saving to justify the move. In the meantime, since there's no competition to speak of in these market segments within this period at least, why bother rushing? Come to think of it, if you exclude the instruction set extensions, the real instruction per-clock core performance gain from 2006 Core 2 to 2013 Haswell only increased by, what, 50%? And the clocks didn't really move up much – from 3 GHz then to 3.5 GHz now…
Leave a comment