In the recently launched Sandy Bridge "Xeon E5" series, the top bin part, E5-2687W, was 'allowed' 150W TDP, compared to the 130W TDP of its desktop cousin, Core i7-3960X. Will the TDP increase continue? On server/workstation parts – seemingly yes. What about the desktops then?
Remember how the 8-core die of the Core i7-3960X had to be limited to 6 active cores to fit the 130W TDP standard high end desktop limit while keeping competitive frequency? Well, to provide similar competitive frequency while keeping all 8 cores turned on – for the far more lucrative workstation market where you charge twice per pop – Intel went to 150W TDP per that same Socket 2011, for the first time. This enabled a 3.1 GHz 8 core part, which as of today is the fastest X86 CPU.
Will Intel continue to increase TDP limits on its high end CPUs? The customer pressure in scientific and technical computing to do so is there, and growing. Often, increasing the frequency – and therefore TDP – is the only tool left standing after all the I/O and compiler optimisations for extra performance are done, and adding more cores is not an option in many apps that just don't scale to more cores or, worse, charge license fees per core. Workstations and technical computing servers are the standard example, in a way similar to top-notch gaming rigs in this sense.
So, the sources pointed out to us that, over the Ivy Bridge and Haswell generations, there will be the high end CPU TDP increase across the board to 150W and then 165W – and here we're talking also about generic high end MP servers with 4 sockets or more, including the 'E7 v2 (Ivy Bridge-EX)' and 'E7 v3 (Haswell-EX)' generations. This is a noticeable jump over 135 watts allowed up to now, and could signal even more TDP for the workstation-specific 'W' parts.
For instance, the general expectations – at the same TDP level – is that Intel would provide roughly 40% performance increase when going from Sandy Bridge EP to Ivy Bridge EP Xeon E5 series. The first 25% of that would obviously come by jumping from 8 to 10 cores with the associated cache increase to 25 MB L3. The next 10% on top of that 25% (i.e. extra 12% from the base line) would come from the frequency increase – for instance, the successor to the 2.9 GHz Xeon E5-2690 would be a 3.2 GHz 10-core part in the same 135W TDP bracket. The remaining few % is, of course, minor CPU enhancements and DDR3-1866 server memory support.
Similarly, in the 150W workstation field, we'd get a 3.4 GHz 10-core part replacing the current 3.1 GHz 8-core one.
Now, what if the Ivy Bridge EP was to follow the same general server 150W limit as its 4-socket 15 core cousin, the Ivy Bridge EX, and add the same extra percentage to the workstation parts? The extra 10% TDP would comfortably give us some 7% higher frequency, enabling 3.4 GHz 10-core server and 3.6 GHz 165W workstation 10-core parts. So, if Turbo is used constantly, one would have 10 cores per socket running on 4 GHz effective frequency in a workstation – no overclocking needed. Not bad…
Watch out Haswell server parts then – 165W TDP limit is expected there for high end server parts, implying possible 180W TDP for the Haswell workstation parts, if Intel decides to follow the trend. So, we'd be guaranteed some nice GHz figures there, on top of Haswell execution enhancements.
What about the high end enthusiast desktop market then? For now, it seems, the 'recommended' 130W TDP there stays for the time to come – it may be just too expensive to justify extra validation work here, compared to the highly lucrative server and workstation markets. However, as long as the board makers go along, we can still push the unlocked 'X' or 'K' parts to sky high beyond 200W power – just without warranty…