correlation sieve for all the streams of information being presented). While this is a GREAT paper, it neglects a few critical points that someone who’s a bit more familiar with the platforms might pick up.
- the rise of APUs.
- Power consumption and Thermals
- Form factor
Let’s tackle these individually (and yes, in a cursory manner):
The rise of integrated CPU and GPU resources is nothing new. For the longest time, the idea of sub-par graphics capabilities coupled to a moderately powered CPU has been the ideal of SMB and “Home” based computing. Ease of integration and application were the norm. With the advance of more powerful GPUs, however, the ability to run onboard (albeit restricted) graphics architectures like GCN or IRIS in AMD and Intel processors (respectively) have given rise to the idea where the processing complex can do more with less real estate. (which will be point #3). The integration of eDRAM on-die (in the case of Iris Pro, for example) bring platform bandwidth (something required to feed the typically starved pipelines of GPUs by way of efficiency) directly to the most needed point of access. This reduces the latency/hops of having to traverse the PCIe bus. What’s been missing, however, has been full hardware integration where common resources (e.g. system DRAM) can be accessed in a completely fluid model. For reference see AMD’s discussion about hQ (heterogeneous queuing) and equal CPU/GPU dispatch.
Power Consumption and Thermals
Unlike discrete GPUs with their sometimes-massive power consumption requirements (~80w-210w differential) and dedicated power drops from the PSU (thus increasing complexity of system design…think OpenCompute), APUs are able to provide a consistent thermal and power envelope that can be planned into the system design without requiring excess cost. The downside of this design is the nature of the GPU core (usually a mid-range and generation behind; Kaveri being an exception) and not being as brute force powerful as, let’s say, current generation parts (nVidia Titan, AMD R9, etc.). Then again, is all that capability needed?
This is somewhat answered in the previous section but given that most servers would want a limited footprint in terms of RU (2RU being common for SIEM appliances) and that disk and network I/O tend to consume most of the slot resources, it would be somewhat of a challenge to fit a 2-slot GPU (with the full-length to match) into a server and maintain power, thermal, and I/O consistency. By utilizing APUs, “thin” 1RU Twin Servers (a la Supermicro or even Dell’s excellent VRTX) could be harnessed without excessive rebuilding times and additional power and cooling considerations.
Just some thoughts down on paper but I’m liking the direction things are going, specific to networking. Thoughts welcome.
- Findings in Parallel and Distributed Computing Reported from New York City College of Technology(hispanicbusiness.com)
- Penguin Computing Updates Densest GPU Compute Solution with NVIDIA Tesla K40 GPU Accelerators(sys-con.com)
- Designed for Big Data Applications, AMD Announces Industry’s First ‘Supercomputing’ Server Graphics Card with 12GB Memory(hispanicbusiness.com)
- Slidecast: Nvidia Launches K40 GPU, Teams with IBM on Power Supercomputers(insidehpc.com)
- imabonehead: Super Computer 13: GPUs would make terrific network monitors – Network World(networkworld.com)
- IBM Embraces Nvidia GPUs for Acceleration(enterprisetech.com)