As Nova elaborated in
his exhaustive and insightful article with Ivy Bridge-EP in 2013, Intel will bring out higher TDP-draw processors, and that trend is set to continue with up to 16-core Haswell-EP monster which targets the introduction in the second half of 2014. The main competitor for Haswell-EP will not be AMD with their unannounced successor of the yet unannounced Abu Dhabi Server CPU (confused yet?), it will be a combination of NVIDIA Maxwell GPU paired with an Project Denver-based Server CPU.
The processor will be available at different clock speeds and with the different amount of enabled CPU cores, but the main power characteristics are
120W, 135W, 145W and 160W. There will not be any sub-100W TDP parts, as the company was told that it has to compete against NVIDIA Tesla GPUs, which all come in 225 and 300W range. With Tesla K10 (Dual-GK104, server version of GTX 690 with ECC and 8GB of GDDR5 memory) coming at 375W, and GK110-based K20 drawing 300W with 12GB of GDDR5 memory, it is obvious what Haswell-EP has to do.
Power draw characteristics require dedicated four and five rail solutions to feed the CPU. We were shown a prototype of the platform which draws less than 100 Amps in typical working conditions, with Turbo Mode jacking that up to 120 Amps. What will make overclockers wet is the fact that Haswell silicon is designed to accept draws of up to
190 Amps on AIR, meaning we might get another overclocking king if its able to sustain 190A when you freeze it and increase the voltage to let's say, 1.575V i.e. 300W.
As far as the platform goes, The 2013/14 server and workstation platform is called Grantley. It will support
DDR3 or DDR4, with extensive memory options. We were a bit intrigued by the amount of DIMMs/SO-DIMMs which Intel supports with the
single 72-bit (64+8) lane (times four and six), bringing this author back into 1999 and the 8-channel RDRAM DEC Alpha… funny to see history repeating itself with the amount of supported DIMMs per channel.
Bookmarks