Intel’s new CEO, Pat Gelsinger, has a huge task on his hands, to re-inject growth into the company at a time when it has intensified competition on many fronts from Samsung, Broadcom, Marvell and others; when it may face restricted access to the huge Chinese market because of geopolitics; when its PC market is fading and its domination of servers is threatened by new low power architectures. Even in the area where it once claimed untouchable leadership, manufacturing process, it has stumbled on advanced geometries and risks being eclipsed by TSMC, the foundry that supports most of Intel’s biggest rivals.
Gelsinger seems, publicly at least, undaunted by the challenge. He was a man of vision when he was previously at Intel, most notably as CTO, and he has gained real world executive and turnaround experience as head of VMware, bringing a welcome mixture of insider knowledge and external perspective to his new role, as well as front-line understanding of the cloud business.
There was a heavy dose of 5G and telco-centric thinking in his new strategy, which recalls his spell as CTO, when he banged the drum loudly for a more radical approach to the wireless network and device business. That largely failed – Intel’s WiMAX gamble was defeated by the strength of the 3GPP community; successive attempts to compete with Qualcomm in handset system-on-chip (SoC) products ended in exit from that segment and even from the modem business Intel had acquired from Infineon.
But of course, as 5G access, core and transport networks migrate slowly but surely to the cloud, Intel has the opportunity to leverage its server market domination. This makes the latest launch, of the Ice Lake processor, highly relevant to mobile vendors and operators. Ice Lake was the centerpiece of a string of recent Intel announcements that set out Gelsinger’s broader strategy.
That revolves around four main foundations:
- To build a standalone foundry business, while continuing to differentiate Intel’s own products by staying a step ahead in process. This was a surprise, given that former CEO Brian Krzanich’s plan to build a foundry operation scarcely got off the ground, and investors have been pressurizing Intel to outsource far more of its own chip production to save the cost and risk of maintaining its own foundries.
- To innovate in chip packaging and provide highly integrated products that will save cost and power for customers in many markets, including telecoms.
- To diversify geographically to a larger extent than Intel has done recently.
- To build a significant IP (intellectual property) ecosystem – again, something that has often been discussed in regard to Intel, with speculation that the firm might open up more of its architecture to licensees and so create a more ARM-style ecosystem. Intel has been progressively participating more in open initiatives such as those surrounding chiplets, and may follow this with broader licensing activities in parallel with opening up its production processes to customers through foundry services.
Most of the pillars of the strategy are not radical new ideas, but somewhat unexpected refocusing on old Intel strategies, and so success will require a modern new take on these.
In telecoms and 5G specifically, the focus of initiatives like Open RAN and white box transport networks is on driving infrastructure power and cost down, and agility and ecosystem innovation up.
Intel needs to ensure that it does not just enable cloud-based networks with its processors and software, but seizes pole position if the closed telecoms networks really do move to an open platform model. In that model, existing ecosystem owners like ARM have a clear advantage, while fully open, and even open source, efforts are trying to rewrite the rules of engagement.
Intel’s size and business model do not allow it just to be a participant in an open market – it needs to dominate the open platform as it did in PCs. The enablement of an open ecosystem for network hardware and software vendors, and even for makers of chips that complement or accelerate Intel’s own, would be the result of success in the foundry and IP strategies. That makes these pillars very strategic in 5G (and in other sectors of course), but the inbuilt suspicion of dominant players that goes along with efforts like Open RAN will give it plenty of challenges.
Overall, though, Gelsinger has a vision of a more agile Intel that not only enables the foundations of the cloud-based, software-defined world, but behaves in the same flexible way as its customers. But while the plan to build Intel as an integrated device maker (IDM 2.0, it was labelled) is all about flexibility, there are also heavy investments in foundries, with their long lifecycles and challenging ROI. Gelsinger’s response to the investors who want Intel to get out of manufacturing following missteps in recent years was to commit $20bn to leapfrogging TSMC again with two new state-of-the-art fabs in Arizona.
On the product side, the highlight was Ice Lake, which is Intel’s third generation of the Xeon Scalable processor family. The chip giant also announced that it is now shipping processors based on the recently announced Xeon D platform, which are targeted at low power edge cloud servers.
From a 5G perspective, Ice Lake is designed to address the biggest obstacle to advanced 5G vRAN – the inadequacy of general purpose processors (GPP), even with significant acceleration from merchant GPUs, to support demanding processing requirements in a vRAN with Massive MIMO (which can increase Layer 1 and 2 processing load by 10 times or more) and with requirements to deliver ultra-low latency and dynamic usage of many spectrum bands.
ARM-based vRAN chip suppliers like Marvell and are racing to fill that gap and now Intel has clearly shown its hand too. To make large-scale vRANs equal in performance to integrated networks currently needs complex and expensive accelerators to offload the more demanding functions from the processors – these tend to be based on FPGAs or ASICs and add significant cost. Nvidia is working on a GPU-based solution with Ericsson but acknowledges that GPU solutions currently come with a power consumption burden.
Intel used to claim that a whole vRAN could run on its high end processors, but that was in the 4G era, and it has acknowledged that current Xeons are not up to 5G Massive MIMO vRAN without a great deal of co-processing. Of course, Ice Lake will still need accelerators, but these will be more integrated and off-the-shelf, and will not increase power consumption, promised Dan Rodriguez, general manager of Intel’s Network Platforms Group.
He said that integrating 3rd Gen Xeon with Intel’s latest Ethernet 800 Series Network Adaptors, plus its AC100 accelerator, would support a 5G vRAN with 64T64R Massive MIMO antenna arrays and three 100 MHz spectrum bands. The AC100 is a dedicated accelerator to handle Layer 1 forward error correction, the biggest strain on processing capacity.
Intel claims the solution can deliver an average of 62% more vRAN performance, compared to the previous generation, and this could be boosted further if Xeon is paired with Intel FPGAs, Optane persistent memory, FlexRAN reference architecture, Open Network Edge Services Software (OpenNESS), Open Visual Cloud, and Intel Smart Edge.
There is also a different version of the processor targeted at a virtualized 5G core, and Intel said this could increase 5G user plane function (UPF) performance by over 40%.
The 800 Series network interface cards (NICs) made their debut, promising up to 200Gbps of total throughput over standard network interfaces. And Intel also announced its Agilex FPGA, which is capable of 400Gbps Ethernet and is targeted at 5G, cloud and edge applications.
The first endorsement for Ice Lake as a 5G platform came from Nokia, which has been more willing than Ericsson and Huawei to move large parts of its 5G platform to merchant chips. That caused problems in late 2019 when its first 5G base station SoC encountered performance issues and high costs, and it is now in the process of migrating all its 5G portfolio to new chipsets designed with multiple suppliers. The Finnish company said it would move its Airframe servers to the new Xeon platform in order to support its 5G AirScale vRAN and its 5G Cloud Core.
On the operator side, the new chips are being sampled by AT&T, Rakuten Mobile, SK Telecom, Verizon and Vodafone and according to Rodriguez, a flexible network that can adjust to many use cases is key to their requirement. “5G has been built from the ground up to not only support mobile broadband but also support a wide variety of vertical markets, so it’s been really built for different enterprise use cases,” he said. “When people think about a 5G network, they do think about a network that really needs to be designed for service flexibility.”
The new chips have been built “to support everything from cloud gaming for consumers to all sorts of use cases that could be in a factory floor, a retail environment, or smart city. I think that’s what’s causing a lot of great conversation about silicon in a 5G network, but it’s also an infrastructure utilizing silicon that enables it to be flexible and agile, really built from the ground up,” Rodriguez added.
Wind River, the virtualization software company that Intel used to own, also endorsed the launch, saying it would base its new vRAN blueprint on the combination of 3rd Gen Xeon, 800 Series adapters and AC100 accelerator.
“Telecommunications carriers are driving the momentum and actualizing the businesses that will grow 5G and the intelligent edge. Operators will be investing nearly $900bn in their networks in 5G technology over the next five years,” said Cyra Richardson, chief product officer at Wind River.
Designed in the 10-nanometer geometry, Intel’s most advanced process node, Ice Lake can support up to 40 cores and claims improved security and a revamped core for a wide range of data center applications. That promises 20% improvement in instructions per clock, compared to the latest iteration of Ice Lake’s predecessor, Cascade Lake, and a 46% improvement in “popular data center workloads”, said Intel.
“We’re excited about getting up to the 40 cores and offering that flexibility to customers,” said Lisa Spelman, general manager of Intel’s Xeon products and data center marketing. “The increased core count, combined with built-in acceleration features, allows us to deliver really strong gen-on-gen performance with an average increase of 1.46 times versus previous generations.”
The move to 10nm is a significant milestone for Intel, which has stayed at 14nm for over six years having encountered challenges with extreme-ultraviolet (EUV) lithography (EUV), which is used to fabricate the advanced chips. That allowed foundries like TSMC and Samsung to leapfrog it and implement 7nm and now 5nm nodes.
Ice Lake is based on Intel’s new Sunny Cove core architecture. This features an improved front-end branch predictor, larger Level 2 caches, enhanced vector processing, and other changes to boost performance in data center, cloud and edge workloads. As well as 5G vRAN and vCore, these will include other highly demanding workloads connected to AI and deep learning – a key target for Intel’s old rival AMD, which recently uncovered its high end EPYC 3 processors.
Rodriguez said: “A trio of inflection points including 5G, AI and edge computing underpin what Intel expects to be a total addressable market of $25bn for silicon in the networking industry by 2024.”
Using the Ice Lake launch to build on the new strategy he has been outlining, Gelsinger said: “It’s a new day at Intel. We are no longer just a CPU company. Only Intel can bring together software, silicon, and platforms packaging and process with at-scale manufacturing for our customers so they can build their next generation innovations.”
He said that ‘IDM 2.0’ would build on Intel’s “unique advantage as an integrated device manufacturer and “deliver a new era of innovation, manufacturing, and product leadership through the powerful combination of three capabilities” – the firm’s internal factory network, strategic use of foundry capacity, and its new foundry services.
He added: “Our IDM 2.0 strategy combines leading process, product, supply, and cost that we will use to design the best products and manufacture them in the best way possible for our customers in every category we participate in.”
Intel will still be on the back foot in terms of catching up with TSMC and Samsung in the latest processes, and even more so in competing as a foundry itself.
The Taiwanese and Korean giants are the real drivers of progress and competition in the semiconductor world outside China now, and they are not standing still in their move towards ever-smaller process nodes and advanced architectures and processes. They are ahead on 7nm and 5nm, which Intel will only reach at volume in 2022, and they aim to have started the move to 3nm within a year too.
The two companies recently unveiled investment plans for chip manufacturing capacity and process, which together add up to $130bn over the next 3-4 years, or 43% of all semiconductor capital investment, according to IC Insights.
TSMC will invest $100bn over three years to expand capacity to meet rising global demand driven by the “megatrends” of 5G and high performance computing. Samsung is planning to spend KR35 trillion ($31bn) this year alone to expand plants in Seoul; Xi’an in China; and Austin, Texas.
At least Intel is giving its foundry business more autonomy than it gained in the Krzanich plan. That did bring in Altera as a customer (as well as the smaller Achronix), but then Intel ate the fledgling division’s lunch by acquiring Altera. This time, Intel Foundry Services will be a standalone unit whose president, Randhir Thakur, will report directly to Gelsinger.
But Intel will not try to do everything at once. Gelsinger said it would contract with TSMC for 3nm production of microprocessors in 2023. This looks ahead to a likely model in which Intel will have chiplets manufactured in different foundries and will add value for customers by integrating different combinations of these chiplets in packages. The company has already been active in developing the interface protocols that allow chiplets to communicate and has released its Advanced Interface Bus royalty-free.
That will be important to ensure Intel is not hurt badly by the webscalers’ trend to design their own cloud processors. Now they could save further power and space by procuring not just an SoC but a full system in a single package made up of many chiplets. Intel is pushing a new definition of the old SiP (system-in-package) term that focuses on flexibility and very small, agile components, but with Intel interfaces and foundry services as the glue to pull them all together. To lend weight to Gelsinger’s view that the webscale industry is heading this way, he was joined on stage by Amazon, Google and Microsoft, as well as Ericsson and Cisco to represent the networks side.