Intel held its 2019 investor meeting today, it's first since 2017, and CEO Bob Swan announced that Intel would launch its 7nm process in 2021 to challenge TSMC's 5nm products.
Intel's Xe graphics will be the leading 7nm product, which will come on the heels of Intel's first discrete 10nm GPU that arrives in 2020. The company also unveiled its first block diagram of the Ice Lake architecture and announced that its new 10nm Tiger Lake processors will come to market in 2020, which we covered here.
The Road to 7nm
Intel hasn't officially revealed a timeline for its 7nm process in the past, but Swan indicated that this fast-follow to the 10nm process reflects an acceleration of the production of the process node. That isn't entirely surprising given the difficulties the company has experienced with its long-overdue 10nm node that finally arrives in volume in June.
Intel also announced that an Xe GP-GPU (General-Purpose Graphics Processing Unit) for the data center would be the first 7nm product to come to market, arriving in 2021. The next 7nm processor from Intel will come in the form of a data center Xeon.
Swan indicated that the company's gross margins would contract as it adjusts its 10nm production and ramps 7nm. Intel's 7nm will be the company's first mass production of a process node using EUV (extreme ultraviolet) lithography technology, providing a 2X improvement in density compared to 10nm. It will also use EMIB and Foveros packaging heavily with its 7nm products. Intel also expects a 15% improvement in transistor performance and a 20% improvement in performance-per-watt.
Dr. Murthy Renduchalata dove in deeper on the difficulties that Intel encountered in its move to the 10nm process, such as taking on too much risk on 10nm design goals. The company plans to address this in the future by redefining expectations for process development. Intel also plans to focus on intra-node optimizations of its new nodes, meaning more "+++" steps for all of its future process technologies. In the future, Intel plans to deliver one Moore's Law gain at the beginning of each node, and another at the final revision of the node.
As you can see in the chart above, Intel will overlap its 10nm++ node and 7nm in 2021, with 7nm taking full center stage in 2022.
The Xe Graphics Architecture - 7nm in 2021
Intel will use its EMIB (Embedded Multi-Die Interconnect Bridge) and 3D Foveros technologies for the Xe Graphics Architecture. Intel also revealed that the new GP-GPU design would power the exascale Aurora supercomputer. We have extensive coverage of Aurora here.
Intel's Dr. Murthy Renduchintala also unveiled a conceptual design of the company's new MCM (Multi-Chip Module) design (right) with several XPU chips, which indicates these are not CPUs, tied together with a fabric. Given Intel's reference to a heterogeneous architecture via EMIB for its Xe Graphics, this could be a potential render of the Xe Graphics Architecture. We're reaching out for confirmation.
10nm Ice Lake Emerges - Shipping In June
Intel also unveiled the first block diagram of its 10nm Ice Lake processors. We already know this design features the Sunny Cove core architecture. Intel says Ice Lake will feature 2X the graphics performance (courtesy of its Gen11 integrated graphics), 2.5X to 3.5X the AI performance, and 3X the wireless speeds of the Coffee Lake chips. As we can see, Intel is sticking with its ring bus architecture instead of migrating to its newer mesh design, and has an integrated USB Type-C controller on-die.
Intel will also release 10nm CPUs for client and server markets, 10nm AgileX FPGAs, 10nm Nervana NNP-I, a 10nm general purpose GPU, and the Snow Ridge 5G-ready SoC, all in 2019 and 2020.
Cooper Lake and Ice Lake for the Data Center
Intel will also ship 10nm servers in the first half of 2020. Intel announced that Ice Lake-SP data center processors are already shipping to leading customers, and that its 14nm Cooper Lake processors are on track for a 2020 delivery. The company says these new processors will come with more cores than the existing models. From a larger perspective, Intel is working to shorten its delays between server processor launches, to roughly one release every four to five quarters.
The Shortage, NAND, and Other Bits
Swan said that 14nm chip supply would continue to improve, with full availability expected by the fourth quarter. Intel also announced that it will not build any new NAND production capacity this year, but will deliver 96-layer NAND in the second half of the year. Intel is also exploring partnerships before it moves forward on the flash memory front. Intel has been rumored to be speaking with Tsinghua Unigroup, a Chinese-government backed corporation, as a potential NAND production partner.