AMD Doubles Down On High-Performance Computing With Milan-X

Every time I write about AMD, I start with “it’s been a great few weeks for the company….” This piece will begin the same way. The company had a great quarter; its 2022 outlook is strong, the Xilinx acquisition closed, and EPYC continues to gain share across all market segments.

Last week, AMD launched Milan-X – a CPU aimed at delivering significantly better performance for high-performance computing (HPC) workloads that require a bit more processing power. Does Milan-X deliver? I will attempt to answer this question in the following few paragraphs.

HPC – it’s a big deal for a couple of reasons

To be clear, Milan-X is designed explicitly for high-performance workloads. General-purpose workloads will not realize any measurable benefit. But that doesn’t matter. Because HPC is strategically important for two reasons:

  1. It’s a lucrative market. Think the HPC market comprises giant clusters that populate the SC500 list? You’d be correct. But it’s also a lot more. It includes the applications used to design and model products; applications used to model weather; applications that model financial markets and help investors make intelligent and timely trades. The HPC market is massive – and the systems that support this market come at a premium.
  2. It’s the future. I take this back a little bit. HPC is not necessarily the future, but workloads with various computational needs beyond what we are accustomed to supporting in the data center are the new norm. We have seen hints of this with analytics. And more directly, we see it with the broad appeal of machine learning (ML) to help organizations with business processes like decision trees. ML is not an outlier in terms of its needs – instead, it’s a sign of what’s to come. 

Suppose you have any questions about this opportunity. In that case, I suggest you read about the recent announcements from NVIDIA regarding its next-generation CPU and compute platform – Grace and Hopper. The company is placing a large bet on HPC in the present and the future.

Milan-X – AMD wants a more significant piece of the high-performance pie

For those who haven’t watched, EPYC Milan-X (formal name is 3rd Generation AMD EPYC processor with AMD 3D V-Cache Technology) is the company’s latest design iteration of AMD’s third-generation EPYC processor (codenamed Milan). Milan is a performance leader across the range of workloads a typical enterprise would deploy – from virtualized infrastructure to cloud-native to database and beyond. In fact, since its launch, Milan-based servers have topped over 250 benchmarks.

Based on the success of Milan, it’s fair to say that the design for Milan-X assumed a position of strength. The challenge for AMD was to improve upon the performance of a chip that already wears the benchmarking crown without changing the real estate footprint on a motherboard. The answer is to build up, not out. And from this came AMD’s implementation of 3D die staking – called 3D V-Cache. The result? The largest L3 cache (by far) available for high-performance and analytics workloads, ready to be deployed without any lifting. No refactoring of applications. No rearchitecting. No new server model to acquire.

Milan-X SoC architecture with 3D V-Cache AMD

3D technology is not a new concept. But AMD’s implementation is. The company employs a hybrid bonding method to stack cache – utilizing a chemical bond instead of soldering. This bonding method results in considerably less power and substantially better performance. Some of the specific claims made around performance include:

  • Electronic Design Automation (EDA) – AMD claims the16-core Milan-X 7373X CPU can perform simulations in Synopsys VCS 66% faster than the traditional 16-core EPYC 73F3 CPU.
  • Computational Fluid Dynamics (CFD) – AMD claims the 32-core Milan-X 7573X CPU can solve 88% more problems per day using Ansys CFX than a similarly configured competitive CPU.
  • Finite Element Analysis (FEA) – AMD claims the 64-core Milan-X 7773X CPU can deliver 44% more performance running Altair Radioss simulation applications than a “top bin” (richest configured) CPU from its competition.
Milan-X – workload affinity AMD

While these performance numbers alone look impressive, it is more compelling because these gains require nothing more than swapping out the CPU and flashing BIOS. No new server. No compiler updates. No application refactoring. This significant magnitude of performance improvement sans corresponding work from IT or software architects.

Isn’t compute becoming commoditized?

One of the common themes in the market is the commoditization of computing. It’s not uncommon to hear this from IT executives with whom I speak. I think we, as a market, are confusing “simple” with “commoditized.” The cloud and increasing levels of automation define the modern enterprise. The business needs immediate access to infrastructure and software environments in real-time. And IT must be able to deliver these.

The above is very real and a challenge IT faces every day. However, the software environments, workloads, and applications IT must provide access to are increasingly diverse. That environment is not just a virtualized database or access to a Kubernetes cluster. That environment spans the computational range – and this is why compute still matters. While the cloud has abstracted these infrastructure platforms from the users, the underlying infrastructure’s needs are still critical. 

The SKU lineup for Milan-X maps nicely to the compute demands of the workloads supported. Each Milan-X CPU uses all eight core complex dies (CCD), which allows for the richest L3 cache across all offerings (768 MB).

Milan-X SKU stack spans 16 to 64 cores AMD

For those workloads that are highly parallelized or require a lot of computational power, Milan-X makes all 64 cores available. But for those workloads that are less dependent on cores – need lots of memory – SKUs are available down the stack, so the customer doesn’t have to pay a performance tax. The earlier claims outlined above show Milan-X supporting applications that span this range.

What Milan-X means for enterprise IT

As previously mentioned, Milan-X has particular applicability in the enterprise. It is a CPU that is tailor-made for scientific applications and highly performant workloads and requires large L3 cache footprints to hold as much data as possible. However, Milan-X represents the AMD EPYC portfolio having the breadth to support virtually any use case in the enterprise data center. Traditional EPYC for everyday needs such as virtualized infrastructure, EPYC FX series (72F3, 73F3, 74F3, 75F3) for typical enterprise workloads such as database and analytics; and now Milan-X for highly performant, cache hungry workloads. This lineup enables AMD to compete directly with its competitors across any use case or deployment scenario in the modern data center.

Closing thoughts

I will finish this analysis the way I started – it’s been a good few weeks for AMD. EPYC has reached cruising altitude in the cloud data center and is a prominent part of every major CSP’s offerings. The CPU has also gained momentum in the enterprise data center as IT organizations have started to adopt EPYC for those workloads that power the modern business. And now, with Milan-X, the company is looking to build on the strong position it has established in the high-performance computing market.

Now, about that Xilinx integration…

Note: Moor Insights & Strategy writers and editors may have contributed to this article.