Intel taps high-performance computing
Earlier in this series, we learned that Intel (INTC) is moving away from the PC (personal computer) and mobile businesses and is investing in the data center and IoT (Internet of Things) space.
The data center market is a duopoly, with over a 99% share owned by Intel and the remaining share held by Advanced Micro Devices (AMD). However, the data center landscape is changing in favor of HPC (high-performance computing) and deep learning.
NVIDIA (NVDA) has been at the forefront of this trend, reporting 63% YoY (year-over-year) growth in data center revenue in its fiscal 1Q17. Google (GOOG) has developed a TPU (tensor processing unit) for its deep learning needs.
Challenging these two companies, Intel has recently launched a KNL (Knights Landing) version of its Xeon Phi processor that can power DNNs (deep neural networks).
Intel’s Knights Landing processor
Intel’s new Knights Landing processor has 30x better performance than its predecessor Xeon CPUs (central processing unit). KNL is better than Xeon in three ways:
- It’s the first CPU to offer Micron’s (MU) Hybrid Memory Cube, a high bandwidth stacked die memory.
- It’s a bootable CPU instead of a PCIe (peripheral component interconnect express) accelerator.
- It has an on-die omni-path interconnect that ensures high bandwidth, low latency scaling between nodes.
Intel has already secured orders from over 30 OEMs (other equipment vendors) for around 100,000 KNL processors, and these sales will be reflected in its 2H16 earnings. It has also secured orders from many supercomputers.
Intel hasn’t yet tested KNL for deep learning, but it claims that the processor may perform better than GPUs (graphic processing units) in certain deep-learning models used in medicine and image processing.
Intel versus NVIDIA
Intel claims that KNL can beat NVIDIA’s GPUs in deep learning by more than two times. Intel has demonstrated this by comparing its four KNL processors with NVIDIA’s four Maxwell GPUs. KNL completed the training of the Caffe Alexnet imaging neural network with 1.3 billion images in 10.5 hours, while Maxwell took 25 hours.
However, HPC faces the big issues of data transfer speeds and huge power bills. UK-based ARM Holdings (ARMH) has been looking to get in on HPC technology since 2011, and it has entered into partnerships with IBM (IBM) and NVIDIA for this purpose. Until now, no ARM-based server has been used in a supercomputer. While competition is still distant in the HPC market, it’s gaining momentum in the traditional and cloud server market.
We’ll look into the other two segments of the data center market in the next part of this series.