Americas

  • United States

Marvell exits the general purpose Arm server business

News Analysis
Sep 03, 20203 mins
Data Center

The company instead will focus on vertical markets and working with hyperscalers to produce custom silicon.

Marvell Technology Group announced last week that it has decided to cancel its ThunderX3 Arm-based server processor for general-purpose server use in favor of vertical markets and the hyperscaler server market.

Marvell was best known for making controllers for storage and networking devices before it bought Cavium, an Arm server developer, in 2018. The company announced the ThunderX3 in March and on paper it looked like a real monster, with 96 cores and four threads per core.

On a recent earnings call with financial analysts, CEO Matt Murphy said the company had always targeted hyperscale customers and is now acknowledging that by focusing on customizing its chips for them.

Of the “Super 7” hyperscalers – Amazon Web Services, Microsoft, Google, Facebook, Alibaba, Tencent, and Baidu – AWS and Microsoft have done their own custom Arm silicon for their cloud services. AWS has its home brew called Graviton while Microsoft has teamed with Marvell to use its ThunderX2 processor.

Raghib Hussain, chief strategy officer for Marvell and founder of Cavium, doesn’t believe the other five hyperscalers will follow Amazon’s lead. “Not everybody is as committed to do their own silicon themselves like AWS. Their silicon teams are not at scale as AWS and not only [do you need] the silicon team but you need a lot more investment in process node,” he told me.

He said the general-purpose market wasn’t an ideal fit for Marvell, but customizing  Arm processors for individual customers could be. “There is a lot of market for Arm vertical processors,” he said.

Marvell saw that with the hyperscalers, one size did not fit all and each wanted a custom hardware solution for their apps. Hussain said most of the IP in these custom solutions is Marvell’s – sometimes all of it. It’s just a configuration difference: a different size cache or different number of cores or in some cases some logic they want embeded in the product.

“A good part of having custom, specific model is we don’t start a project unless it is mutually funded,” said Hussain. “The risk is substantially reduced for a company like Marvell because the customer is getting what they want, and hey have skin in the game, and from our point of view we are developing a product for a targeted customer and ROI is very good.”

Jim McGregor, principal analyst with Tirias Research, says Marvell is continuing on with what’s been successful and a better move because they get up-front commitments. “The customer pays up front and places an order, so Marvell knows the minimum they will make,” he said.

And while Marvell is giving up on general purpose servers, there are still plenty of places to use an Arm socket, like networking and storage. “You don’t want an x86 server chip for those. It wouldn’t make sense,” he said.