Samsung could upset the AI and HPC markets with GPU entry | Virtual Reality
Jon Peddie Research reports that Samsung has hired Dr. Chien-Ping Lu, a former Nvidia executive who oversaw development of discrete and integrated GPUs used in Apple MacBooks as well as integrated processors before the advent of GPUs in CPUs.
Now, many mobile CPUs have integrated GPUs, or the phone itself comes with its own GPU. Qualcomm has its own line called Adreno. But Peddie doesn't think Samsung hired a heavy hitter like Lu to make mobile phone chips.
“The world can't seem to get enough GPUs for all the AI ambitions being explored,” Peddie wrote. “Samsung already has a big AI team, [and] they must certainly be eyeing this development.”
It's no accident that Nvidia's GPU technology has rocketed to the lead of the AI processor market. AI is a massively parallel process, something GPUs are born to do. GPUs are designed to do mathematical acceleration in a massively parallel method. Whereas a Xeon has 22 cores, a Nvidia Volta GPU designed for AI has thousands, all running in parallel.
GPUs are also immensely popular in high-performance and supercomputing environments. In the most recent Top500 list, more processing power came from GPUs than CPUs, almost all of that powered by Nvidia.
What about AMD?
If you are wondering why AMD isn't a player, it's not because its GPUs are poor. Quite the contrary; they can hold their own with Nvidia and have about one-third market share to Nvidia's two-thirds market share.
The problem is that AMD was in an existential fight for its life up until the last year. It made good GPUs but didn't have the resources to fight Nvidia, which has spent almost two decades seeding the market for HPC and now AI for its processors. Nvidia may look like an overnight sensation in AI, but it's been at this for a long time, while AMD was fighting to stay alive. AMD's most recent quarter was excellent; the Zen processor has turned things around for the company. But it has a long way to go against Nvidia in AI and HPC.
Resources aren't a problem for Samsung. We're talking about a company that reported $50 billion in operating profit last year.
Peddie says Samsung's GPU development teams are in Austin, Texas, and San Jose, California, and are reportedly hiring engineering teams to handle hardware, software, testing, and verification of designs.
“Simulations have indicated an even better performance than predicted, and the batch instruction processing holds the promise of significantly reducing the latency in the motion-to-photon path, and bringing low-power (consumption) [virtual reality], with fast recovery very high-resolution dynamic display capability,” he wrote.
The question then becomes what kind of infrastructure will Samsung build around the chip. It needs the ecosystem similar to Nvidia's CUDA language, and it will require a major effort to educate developers the way Nvidia has done. So, even if the chip is introduced next year, it will be years before they start showing up on the Top500 list.
Samsung isn't the only chip maker with GPU visions in its head. Intel is reportedly working on a discrete GPU for 2020 release. But they have tried twice and failed spectacularly, and the ship of state is currently without a captain. A new Intel CEO might scuttle it to focus on core competency. At least, he or she should.