Graphics processing units have found an important life beyond videogames as an accelerator for artificial intelligence, but they could soon share that spotlight with other groupings of silicon.
investors especially should pay close attention to this battle, as that stock has experienced the biggest AI-influenced bump, though Advanced Micro Devices Inc.
could also be affected. Both stocks dropped Tuesday after Susquehanna Financial Group analyst Chris Rolland wrote that there may be a better semiconductor for AI processing.
“2017 was the year of Artificial Intelligence GPU but might 2018 be the year of the ASIC?” Rolland wrote in his note on Tuesday. as the No. 1 theme of 2018 in the semiconductor sector.
Rolland believes application-specific integrated circuits, or ASICs, could be the long-term winners in the burgeoning AI arena. ASICs can be customized for their workloads, as can another type of chip, the field programmable gate array, or FPGA, a type of chip that has also been put forth as a potential beneficiary as AI progresses.
As an example of the progression, Rolland points to cryptocurrency mining, where GPUs have been used early in a digital currency’s lifespan before turning to ASICs once the task gets tougher. Bitcoin
miners typically use ASICs, and Ethereum miners are expected to make the switch to ASIC this year after a miniboom for Nvidia and AMD from early GPU-based efforts on that blockchain. RBC Capital Markets analysts made a similar point in a note Wednesday morning.
Rolland noted Nvidia Corp.’s rise in the last two years as the clear semiconductor beneficiary from the growth in AI workloads. This year alone, Nvidia has risen another 80%, after closing out 2016 as the top-performing stock in the S&P 500
Nvidia should still be an important figure in AI, and has its own ASIC-based tensor processing unit to pitch, but other chip makers could benefit from changes moving forward. Rolland noted that FPGAs designed by Silicon Valley veteran chip company Xilinx Inc.
are being used by Amazon Web Services
, but he believes that while FPGAs may benefit from more AI workloads, they may have a limited footprint in contrast to ASICs. For those chips, Rolland points to Broadcom Ltd.
which he notes helped create AI-focused chips for Alphabet Inc.’s Google
and potentially Cavium Inc.
which has discussed ASICs specifically designed for AI but is a bit of a mystery ahead of a proposed merger with Marvell Technology Group Ltd.
Another company that could benefit is a rare chip unicorn. Rolland hosted a fascinating conference call last month for investors with Andrew Feldman, Chief Executive and founder of a stealth chip startup called Cerebras, which now has a $1 billion private valuation. Feldman is developing ASICs for artificial intelligence workloads, and described on the call why GPUs are not the best chips for AI.
One limitation is that the cores, or threads of the GPU, do not talk to each other, he said, which creates a big communication constraint. He gave a metaphor as an example.
“If I have to do some work and run the answer to your office, and if you have to do some work, and run the answer to somebody else’s office, and I have a way to improve the amount of time it takes me to do the work, you say that’s good, but 99% of the challenge here is in moving the answer from you to the next guy.”
Feldman spoke about the limitations of the GPU architecture in AI, without discussing in any details what Cerebras is doing. He did note that the company has 60 engineers with a combined 1,200 years of experience working on processors.
“Wouldn’t it be serendipitous, wouldn’t it be a surprise, if after 25 years of tuning a part, or optimizing a part for one market, that it was a great fit for a separate and unrelated market?” Feldman said during the investor call, referring to Nvidia’s discovery that its GPUs were a great solution to running computing-intensive AI workloads. “That just doesn’t happen. The best way to think about it is, it is the best part available today. It has allowed the industry to take one step forward, it has shown a glimpse of the future.”
ASIC chips, like GPUs and central processing units made by Intel Corp.
, have been around for a long time, but Rolland said he has high hopes for this chip technology because of the ability to customize these chips for different customers and workloads.
“We think ASICs are the long-term winner here,” Rolland said in his note, which listed Broadcom, Marvell and Microsemi Inc.
as his best bets for chips in 2018.
Without any further information from Feldman about his company, who talked with Barron’s in September, it’s a bit hard to tell how far away the launch of its first chips may be, but it’s definitely an interesting development to see a unicorn chip company with a story to tell. Meanwhile, investors looking for AI (or crypto) bets in the chip sector seem to have a growing list of options beyond Nvidia and AMD.