Epyc Birthday: AMD’s Xeon Challenger Turns One
A year ago today, AMD unveiled what’s arguably its most significant challenge to Intel across the markets where the two companies compete: Epyc. Consumer parts are great for volume and brand recognition, but the Epyc server CPU family is aimed squarely at one of the few PC markets that’s booming, year after year. It’s easy to see why. Server CPU margins are outstanding compared with your average desktop selling prices, and building long-term relationships with relatively conservative business customers is a great way to build a business that can weather economic downturns or less-than-stellar product launches. Intel’s data center business has often insulated the company from periodic downturns in the consumer space, and AMD obviously wants in on some of those dollars.
Over the past year, AMD has announced major wins with Baidu, Dell, Yahoo Japan, and Microsoft. The company has deals with two of the “Super Eight” data center customers (Next Platform gets the nod for that nickname, which bundles Google, Amazon, Microsoft, Facebook, Baidu, Alibaba, Tencent, and China Mobile). These companies collectively represent the largest and most demanding customers in the data center business. All of this is fairly impressive given that Epyc is just one year old.
Historically, the server market is extremely conservative; AMD launched its first Athlon CPUs for data centers in 2001, but didn’t begin to really take off in that market until spring 2005. The company’s CEO, Lisa Su, has stated that she hopes to take 4-6 percent of the data center market in 2018, while Intel CEO Brian Krzanich recently remarked it’s Intel’s goal to keep AMD below 15-20 percent of the server business. (Su’s statement was a prediction made at the end of 2017 about goals for the firm’s 2018 performance, Intel’s statement was more general).
Epyc is, of course, based on the same building block as Ryzen, an eight-core combination of two CCX units with 16MB of L3 cache per Zeppelin die. But there are some changes and advances baked into the core that Zen CPUs either don’t rely on to the same degree or don’t use. While AMD doesn’t built monolithic dies like Intel does, managing thermals and power consumption on a CPU with up to 32 cores is an extraordinarily complex endeavor. Epyc has a network of 1,300 sensors that deliver data on the CPUs internal temperatures, voltage, power consumption, and clock speed every millisecond.
Based on this data, the chip calculates the minimum voltages required for effective operation and uses LDOs (Low Drop Out linear regulators) to adjust the voltages and frequencies of its CPU cores, while metal-insulator-metal (MIM) capacitors are utilized to reduce short-term voltage drops and improve performance per watt. The CPU also uses an algorithm to detect when the CPU is utilizing its clock boost capabilities in non-optimal fashion by, for example, completing a piece of non-latency sensitive work at maximum frequency and power consumption. In theory, this task could’ve been completed within a slightly longer time frame while consuming less power. AMD’s algorithm is designed to detect these opportunities and improve overall performance per watt by tweaking the CPU’s response to take advantage of them.
Epyc also operates over a fairly large power range and OEMs can choose how to allocate that capability. A CPU with a nominal 180W TDP can be pulled into a 165W envelope or expanded into 200W. Lower-power 120W chips can be pulled down into 105W envelopes. Intel, to be clear, also offers some similar TDP flexibility, but AMD also allows data center customers to choose if they wish to pursue a performance model that is performance-deterministic or power-deterministic.
In power-deterministic mode, the CPU runs as fast as possible within a given TDP. There will be some performance variation from CPU to CPU, but overall this mode maximizes performance within a given power envelope. In performance deterministic mode, performance is identical across all servers, but TDP may vary from server to server based on the specific manufacturing quirks of the CPU in question.
Overall, these two modes allow Epyc customers to choose whether they wish to emphasize constant performance and variable TDP, or to hold TDP constant and allow performance to vary. The specifics of which option is “better” depends on the workload in question.
As far Epyc’s overall performance in the market and Intel’s responses to it, there’s not a ton of data out there that we haven’t previously covered. AMD continues to win back some server customers and market share, but it’s not clear if the company will revamp its Epyc product families before the launch of 7nm chips next year. The company has done an excellent job building servers to field against Intel, but the natural conservatism of the server market and Intel’s determination to hold the space are likely to be challenges going forward. Still, Epyc is building momentum and winning customers — and it wasn’t even 18 months ago that this was far from a sure thing.
Continue reading
Happy 5th Birthday to Pascal and the Ongoing Disaster of GPU Availability
Happy fifth birthday to Pascal, Nvidia's last rasterization-only flagship GPU microarchitecture.
Happy 10th Birthday to the Kepler Space Telescope
Happy 10th birthday to the Kepler space telescope, arguably the most important remote observatory humanity has ever launched.
Happy Birthday, ARPANET: The Internet’s Grandfather First Connec
Happy 50th birthday, ARPANET! The grandfather of the modern internet came online 50 years ago today.