We recently shared a post on three remarkable deep learning projects at Queen Mary University London (QMUL). SketchX, Warblr and folk-rnn were underpinned by a powerful IBM-NVIDIA system provided by OCF, called IBM Power System S822LC and codenamed ‘Minsky’.
The three projects were undoubted successes, leading to major media coverage and awards for QMUL. You may still be wondering, however, just what is so special about ‘Minsky’? And how did it support this level of innovation?
In this post, we’ll try to answer that question with a closer look at the technology inside ‘Minsky’.
Screwdrivers at the ready, let’s open the box…
POWER8 + P100 + NVLINK: a unique combination
The first and most significant thing to note about ‘Minsky’ is that it’s no run-of-the-mill server.
There may be a shift towards x86 systems in business IT lately. But academic research and deep learning are not standard computing. A competitive system must go beyond the norm.
Inside ‘Minsky’ we see three key components.
1. Two POWER8 processors
POWER8 is IBM’s newest RISC processor. It leads the industry in performance and is the first to support the OpenPOWER software environment.
Compared to other 24-core scale-out systems, POWER8 delivers up to 2.7x performance across key workloads.
POWER8 is also at the heart of IBM’s flagship cognitive computing system, Watson.
2. Four NVIDIA Tesla P100 GPUs
Built on the Pascal GPU architecture, Tesla P100 is the most advanced data centre GPU ever built.
Paired with the NVLink interconnect (see below), it delivers up to 50x better performance for data centre applications.
3. NVLink CPU-GPU interconnects
Finally, while POWER8 and Tesla P100 processors exist in other systems, only ‘Minsky’ allows them to share data at incredible speeds of 80GB/s via the NVIDIA NVLink interconnect.
NVLink allows CPU-GPU data sharing at up to 12 times the rate of 3rd generation PCIe interconnects. This is extremely important in deep learning and neural network applications, which rely on processing large blocks of data in parallel. GPUs are better at processing these large chunks of data than CPUs.
Deep learning applications divide processing tasks between CPU and GPU according to their specialisation. But slow data transfer between CPU and GPU can create a processing bottleneck.
Super-fast NVLink removes this bottleneck, making ‘Minsky’ an ideal choice for deep learning applications. Especially since NVLink is not available on x86 systems.
Watch a video about NVLink:
Faster time to insight
All this combines to deliver breakthrough performance for GPU-accelerated applications.
Compared to a similar x86/Tesla K80 system with no NVLink, ‘Minsky’ delivers:
- 5x more queries per hour running Kinetica “Filter by geographic area” queries
- 9x more GFLOPS based on running LatticeQCD
- 2x more “Base Pairs Aligned” per Second running SOAP3-dp with 2 instances per device.
- 3x better performance (57 percent reduction in execution time) running CPMD
- 7x better performance running the High Performance Conjugate Gradients (HPCG) Benchmark. (source: IBM Data Sheet)
These capabilities helped QMUL’s research projects to uncover innovation and innovation faster.
‘Minsky’ isn’t just about power, either – it also simplifies development of deep learning applications.
The system includes a feature called Coherent Accelerator Processor Interface, or CAPI. CAPI allows the CPUs and GPU accelerators to share the same virtual memory address system, so they can share data in-memory.
This means that:
- Developers don’t need to worry about manual data management between CPU and GPU
- CPU and GPU can share and work with very large datasets, even up to 1GB, with very low latency
Applications can, therefore, be built more quickly and easily. And they can handle larger datasets, and run faster, when complete.
Call to action
This was a brief insight into the key advantages of ‘Minsky’ over comparable systems. If you’d like to know more about the system and how it could support your research and innovations, we’re happy to chat. You can email me directly at email@example.com, or leave a comment below.