I’m profiling for alexNet/caffeNet on JetsonTX1.FOund there is property called ‘MAX IPC’ which is 6 for jetson tx1. This ipc is per core or per sm. Because I Changed different application and got same number for this property so I’m pretty sure this propery not application specific.
If anyone can explain what is this propery that’ll be great help.
please find the link for screenshot.
[url]Dropbox - Screenshot from 2017-06-18 19:23:50.png - Simplify your life
It is per multiprocessor (per SM).
It indicates the maximum number of instructions that can be issued per SM, per clock.
It is a function of the hardware, not of your application code.
A given core (or we could generalize and say a given functional unit - FU - in the SM) can have at most a single instruction issued to it per clock. Since FUs (and therefore instructions) are mostly pipelined, it’s possible of course for a given FU to have multiple instructions “in flight” in a given clock cycle (i.e. at different pipeline stages) but with respect to a single cycle, only a single “new” instruction can be issued to a given FU in a given cycle.
Having an IPC greater than 1 indicates that an SM design is such that instructions are issuable to independent groups of FUs. Most modern SMs (after Fermi cc 2.0) have this capability.
It was my understanding too , in case of maxwell architecture there are 4 warp scheduler per SM, each warp scheduler capable of dispatching 2 instruction that means max IPC should be 8 why 6?
I don’t know the precise answer, but I imagine there is some other hardware design issue that prevents all 4 warp schedulers from dual-issuing in a given cycle.
The scenario would probably be difficult to achieve in practice. Since the Maxwell SM has 128 SP cores, you could only issue 4 instructions of the FADD, FMUL, or FMA type. So you wouldn’t get more than quad-issue (probably) for dense FP operations. You’d need a careful mix of instructions just to achieve 6 with any meaningful frequency.
But I don’t know the precise answer. Possibly someone else will answer. I bet Scott Gray knows.
Well The scenario you are talking about it’s a practical scenario while MAX IPC represent Ideal case. so it shouldn’t account account those scenarios. Thanks txbob for inputs.
