The parallel elements are themselves made up of multiple fully 32-bit floating-point processing units in MIMD (Multiple Instruction, Multiple Data), each with an array of SIMD (Single Instruction, Multiple Data) vector units. Think of the hardware as an array of vector units, likely 4x4 for this first iteration of PhysX hardware, which fits in with their likely make-up and implementation in silicon. Ageia is reluctant to tell us how it works and we're left poking at patents and developer information to glean our ideas about the hardware.
The units aren't generally programmable, at least in the way you might be used to thinking about with a GPU. You can't (easily) 'shade' the physics interactions and be certain the hardware will (mostly) execute an instruction stream you can further control. There's no compilation of physics programs as you would a shader program on a programmable GPU.
In terms of its data rate, the hardware is supposedly capable of six 5D vector MADDs per cycle, per vector unit. In the 4x4 design we suspect first hardware has, that's a near 50Gflop (all 32-bit) rate when fully utilised, at 250MHz.
Our sample is one of BFG Tech's PhysX boards. Measuring 167x99mm, the PCB for these first BFG PhysX boards is pretty much identical in dimension to a modern sound card. Indeed, my Audigy 2 ZS is much the same size. I've yet to come across a sound card that needs an active cooler and auxiliary power input, though.
The 40mm fan and heatsink combo reminds me of those fitted to graphics boards in days gone by. It spins devoid of any thermostatic control and rotational regulation. It's loud enough to intrude on a silent PC and I estimate something in the order of 32-36dBA - yes, it's just a guess and we know it's logarithmic but we don't have the tools to accurately measure.
The heatsink gets pretty damn hot over time, indicating the aluminium heatsink and fan aren't the best pairing to get rid of the current PPU's heat, but a well ventilated chassis should see you right. Oh, and it lights up.
The PhysX PPU's memory controller supports the same GDDR 3 memory that current graphics hardware does, and the BFG sample is equipped with four 32MB 500MHz Samsung DRAMs. Each populates a 32-bit channel and they combine to give the chip a maximum read bandwidth of 16GBps and 128MB total on-board storage.
