Samsung Puts the Crunch on Emerging HBM2 Market
The memory market can be a volatile one, swinging from tight availability and high prices one year to plenty of inventory and falling prices a couple of years later. The fortunes of vendors can similarly swing with the market changes, with Samsung recently displacing Intel at the top of the semiconductor space as a shortage in the market drove up prices and, with it, the company’s revenues.
High performance and high-speed memory is only going to grow in demand in the HPC and supercomputing arena with the rise of technologies like artificial intelligence (AI), machine learning and graphics processing, and getting a growing amount of focus in the industry is High-Bandwidth Memory-2 (HBM2). There also is demand for GDDR5 and plans by many vendors to soon start rolling out GDDR6. There are tradeoffs – HBM2 is more expensive but enables lowers power consumption. All that said, Samsung, Intel, Advanced Micro Devices, Nvidia, NEC and others all are coming out products with HBM2.
As we’ve noted at The Next Platform, Intel is getting ready to finally bring “Lake Crest” deep learning chip – which will include 32 GB of 3D stacked HBM2 – that is based on the architecture the company inherited almost two years ago when it acquired Nervana Systems and will compete with Nvidia and its cadre of deep learning GPUs and appliances. The Lake Crest chip is the beginning of an aggressive roadmap Intel is laying out for the Nervana technology, which includes what will be the follow-up to Lake Crest, called “Knights Crest.” Intel also has integrated HBM2 in its Stratix 10 MX field-programmable gate array (FPGA), with the ability to put up to two HBM2 devices into a single package for a maximum memory bandwidth of up to 512 GB/s.
Nvidia is leveraging HBM2 in a range of products, including its powerful Titan V PC GPU, which includes 12GB of HBM2, and Tesla “Volta” V100 datacenter GPU aimed at such HPC workloads as AI training and inference and high-end graphics. The Tesla V100 includes 16 GB of HBM2. Nvidia also has used HBM2 in other products, such as Quadro GPUs based on the “Pascal” architecture. AMD is arming its new “Vega” Radeon GPUs with HBM2, and at the recent CES 2018 show, both AMD and Intel rolled out chips that include the Vega Radeon GPUs. At the same time, AMD is said to be planning to use GDDR6 in future products.
Read More |
