NVIDIA is scheduled to release ExaScale machine architecture in 2017,Bill Dally,chief scientist of NVIDIA at the Global Road Show organized by Institute of Process Engineering on Oct,27th.The way of introudction of architecture of eight-year later seems a bit of suspicion.However, Nvidias ambition to enter the general-purpose computing market is clear.
According to Bill Dally, NVIDIA will put 2400 throuput cores and 16 CPUs on a single chip with 300W TDP.It is well known that NVIDIA will make CPU sooner or later.This is obvious evidence in definite roadmap at first. However, Huang stresses that NVIDIA will not make CPU, while the Bill Dally also takes Tegra as a shield in answering relevant questions.
This CPU + GPU chip from NVIDIA has also confirmed that they will achieve Fusion in the opposite direction of Intel and AMD.
Each throuput core include three single-precision floating-point unit and a double-precision floating-point unit, the chip is expected to provide a total of 40T single-precision floating-point processing power and 13 double-precision floating-point processing capabilities. Node using this chip also has 128GB of memory, and 512GB of phase-change memory, or flash memory as a high-speed local memory. This node architecture is very similar with current super-computer.Although the figure is amazing,this product can be realized as long Moores Law takes effect.
The wole system is made up of 128 cabinets,which can provide up to 2Efplop computing capabilities with a TDP of 10W.
NVIDIAs future product appears to be powerful.However,Green giant wont give detailed time frame just like the preparation of Exaflop, phase-change memory, heterogeneous processors.