During the debut appearance of the SKARAB agile neworked FPGA supercomputer at IEEE-HPEC 2015, a seemingly secondary item at our demo booth, the Hybrid Memory Cube mezzanine card that is in late development for the SKARAB caught a lot of attention. As much as folks were impressed by the throughput and performance of SKARAB's four 40 GbE interfaces screaming away while barely breaking a sweat (only 35 W power consumption); many of the high performance computing researchers in attendence were dazzled by the sighting of a mythical creature: a Micron Technologies Hybrid Memory Cube (HMC) chip.
Comments quickly followed:
- I thought HMC was unobtainium!
- Where did you get that?!?!?
- Does it work?
The short answers are: Not quite, South Africa, and We're working on it.
Here's the scoop. The engineers at Peralex, makers of the SKARAB, kindly sent a prototype Hybrid Memory Cube mezzanine card designed to plug into the product's motherboard for us to show as an example of the mezzanine capability...but they kept the functional ones for integration and test in their labs. So yes, HMC exists, and it works...in this case, near Cape Town!
The SKARAB HMC mezannine prototype has spots for 31x31 mm BGA packaged chips with 4 Gb capacity each. Only one HMC is populated, but the board supports two chips. The first HMC chip is interfaced through two half-width links. A "link" is an external high speed interface (16 TX & RX high speed serial lanes, each running at up to 15 Gbps). Lanes can also be "half-width" (8 RX/TX lanes). Links allow multiple devices (CPUs, FPGAs, etc) to access the memory simultaneously, making HMC a useful way to pass information between processors.
The links also allow HMCs to be meshed up, so one can access an HMC device through another (to expand the memory capacity without needing more lanes on the processor). The two SKARAB HMC mezzanine chips are interfaced through one full-width link (16 TX/RX lanes @ 10 Gbps). Overall, this produces about 16 x 10 Gbps bandwidth to either device!
Hybrid Memory Cube is exciting cutting edge technology ideally suited to the high processing and memory bandwidth capabilities of SKARAB's huge FPGA. With the possibility of having 24 GB of this memory available to high perfomance computing algorithms that can work with 4 x 40 GbE data streams, insane applications can be imagined---and accomplished!