AMD Flexes Manufacturing Might In Data Center Announcements

AMD announced a flurry of new products this week in what is best described as silicon beasts for the data center. With announcements of new data center processors and a GPU for accelerated computing, each takes the performance of a single chip/socket to new levels and demonstrates the impact of designing products for High-Performance Computing (HPC) applications. While the performance expectations of the new products are impressive, what stood out to me is how AMD is leveraging the latest semiconductor packaging and manufacturing technology 12 years after spinning off its own manufacturing group to form GlobalFoundries (GF).

The first announcement was of a new Epyc server processor called Milan-X. Milan-X shares the same Zen 3 architecture as the existing Epyc processors with up to 64 CPU cores, but Milan-X adds a stacked memory die on top of the chip, which AMD refers to as V-cache. The CPU and the V-cache dies will be manufactured using TSMC’s 7nm process node. The result is a 300% increase in processor cache for a total of 804 MB of cache per socket. The V-cache uses direct copper die connections without microbumps for increased interconnect density and energy efficiency. The new parts are compatible with the existing Epyc Gen 3 processors with just a minor BIOS upgrade. According to AMD, the result can be a significant increase in performance of technical workloads, such as a 66% faster semiconductor RTL verification. Microsoft’s Azure is using Milan-X in its HBv3 platform and indicated increases in workload performance ranging from 50% to 80%. The Milan-X will be generally available in Q1 2022.

AMD also introduced the first product to use its next-generation GPU architecture for computing – CDNA 2. The new product is the Instinct MI200 OAM, the second generation of the GPU product family. MI200 combines two of the new CDNA 2 GPU cores into a single package resulting in a whopping 58 billion transistors in a single chip. It features 220 compute units, 880 matrix cores, and 128GB of HBM2E (stacked high bandwidth) memory. The MI200 GPU will also leverage a 3rd generation of AMD’s Infinity fabric capable of supporting up to 3.2TB of total memory bandwidth. The chip is manufactured using TSMC’s 6nm process node and will consume up to 550W but, incredibly, it is packaged in an air-cooled module, although liquid cooled options will be available. The module is available today and a PCIe card version will be available in the future. The MI200 was designed in cooperation with Oak Ridge National Laboratory for the Frontier Supercomputer, and assembly of Frontier has already begun. If completed on time, Frontier should be the first U.S. Exaflop supercomputer.

The company also provided a sneak peek into the 4^th generation of Epyc processors featuring two new processors – Genoa and Bergamo. Genoa will be the standard processor that features up to 96 of the next-generation Zen 4 CPU cores plus support for DDR5 memory and the CXL and PCI Gen 5 interconnect technologies. Bergamo, will feature up to 128 Zen 4c CPU cores that are enhanced for cloud applications, including a new cache hierarchy and more power efficiency. Both products will be manufactured on TSMCs 5nm process node and will be socket- and software-compatible. Genoa is sampling to customers today, and will be in production in 2022. Bergamo is scheduled for production in 2023.

What is even more impressive to me than the products, is how AMD has vaulted to the forefront of semiconductor technology. While the company has always invested in new semiconductor technology, AMD often trailed behind its competitors in manufacturing processes and even spun out its manufacturing group in 2009. However, since developing the Zen architecture, AMD has continued to leverage leading edge semiconductor technology, processes, and packaging. With the Instinct MI200 announcement, the company highlighted the use of a new Elevated Fanout Bridge (EFB) to reduce the cost and complexity of building chips with multiple dies, known as multichip modules (MCMs), chiplets, or 2.5D packaging. Combined with being on TSMCs leading edge process technology, AMD has become a semiconductor technology leader. As we have seen in the past, having a leading product architecture and a manufacturing advantage is a combination for success in the technology segment.