This post has been republished via RSS; it originally appeared at: Azure Compute Blog articles.
Small chips, Big problems
Ansys has developed simulation software tools for over 50 years, helping their customers virtually prototype their products within a simulation before they ever build a physical prototype and test in the real world. This simulation process is critical in designing high performance electronics from cell phones to computers to complex radar systems. The integrated circuits (ICs) in these systems have steadily increased in density and speed to handle the growing needs of computation, data handling, and communication. The problems are exacerbated in RF (radio-frequency) IC’s, which operate at incredibly high frequencies and require very precise noise isolation between digital and analog blocks. With the advent of 5G+ communication products, these problems are not going away.
Ansys provides a portfolio of products that enable electrical engineers to calculate the electrical properties of these ICs and determine if they will meet their stringent performance requirements. Whether it is extracting the inductance of a resonant circuit or understanding the unwanted and unavoidable coupling between two adjacent blocks, engineers need to be able to simulate these effects to approve a “tape-out” of a chip. Historically, engineers have tackled this in two ways:
- Simulate a full IC using a tool with a medium level of accuracy which is sufficient for early-stage design or lower-frequency applications.
- Simulate a partial IC using a tool with a high level of accuracy, ideal for late-stage validation studies or higher-frequency applications (like 5G).
Figure 1: Top and Angled view of Integrated Circuit in HFSS.
Today, more and more companies are pushing for the ability to simulate a full RF IC using the highest-level accuracy tool possible. Until recently, this was impossible due to the computational complexity required and limitations of the meshing technology.
It was impossible, until now.
It’s true – a Full Chip Solved in HFSS!
I am excited to announce that Ansys HFSS has completely solved an adaptively converged mesh (15 passes) for an entire RFIC (5.5 x 5.5mm) (see figure 1) at 5GHz in under 30 hours. This model has 64 ports distributed across the entire IC, which represents the most challenging conditions for the adaptive refinement process to solve. We were able to do this by bringing together two new significant technologies:
This work represents a huge leap forward in raw power and capability for IC designers, and anyone who has solved computational electromagnetic problems will share my astonishment at the numbers in this blog (for the rest of you, bear with us as we dip into some statistics here). Here are a few:
- Compute cores used: 704 cores (Intel Xeon Platinum 8168, Azure “HC44” VM)
- Initial Mesh solve time: 1h54m
- Final RAM used: 2.6TB (yes, TERA bytes)
- Final mesh size: 23.5M Tetrahedron and 93M unknowns
- Final mesh point solution time (5GHz frequency point solved after adaptive pass 15) = 2h19m
- Total Adaptive Mesh Time: 29h47m
HC44 VM features 100 Gb/sec EDR InfiniBand, the same HPC interconnect technology featured in the world’s most powerful supercomputer. Ansys is leveraging Azure’s unique InfiniBand fabrics of our H-series VMs to host this HFSS workload.
Azure InfiniBand fabrics is based on non-blocking fat trees with a low-diameter design for consistent low latencies - together with SRIOV - it allows partners such as Ansys to use standard Mellanox/OFED drivers just like they would on a bare metal HPC cluster.
The first technology required was the meshing. Advanced meshing and geometry preparation capability in HFSS Layout enables the adaptive mesh algorithms in HFSS to efficiently handle the layered 3D structures in the IC. Approaching the problem from a layout-centric viewpoint enables an initial mesh to be efficient and sufficient, and thus a final mesh that accurately capture the physics.
The second technology is the ability to easily solve on the cloud. Even with this improved mesh, solving this kind of problem is constrained by the amount of available RAM on a compute node. By leveraging HFSS distributed direct matrix solve, we were able to harness all the required RAM across 16 HC compute nodes. The Ansys Cloud R&D teams partnered with Microsoft Azure engineers to create a robust and user-friendly way for customers to request and use this compute power.
“It is so rewarding to see a problem of this size and complexity solved on Azure, putting this level of HPC power in the hands of engineers when they need it the most.”, says Merrie Williamson, Microsoft VP Azure Apps and Infrastructure. “This is a great example of two partners blending their strengths to serve their joint customers, and I’m eager to see the impact this will make on companies tackling the new wave of IC-design challenges posed by 5G and beyond.”
Ansys Cloud’s simple GUI gives the user the ability to choose the machines and region, providing supercomputing-like capability on-demand to companies of all sizes. The Intel Xeon Platinum 8168 CPU’s on the Azure HC-series machines provide the performance (speed) boost, and Ansys developers regularly work with Intel to benchmark their simulations on upcoming Intel chips to provide feedback for future architectural decisions. Ansys Cloud provides flexible on-demand HPC access for most mainstream Ansys solvers – HFSS, Maxwell, SIwave, Mechanical, Fluent, LSDyna, Discovery - and more on the way.
Figure 2: Ansys Cloud, available on Azure.
Sleep better at the 11th hour
So, what does all this mean?
If you design RFICs and you are approaching crunch-time for a tape-out, you need to ensure that you have no surprises when you get that first chip back. Mistakes or missed problems can hit companies with millions of dollars in unplanned NRE cost and, more importantly, weeks or months of product delay to market.
A full-IC HFSS simulation can provide that 11th hour peace of mind.
Until now, this kind of simulation was not possible. It is not like we are saying, “look, we can now run this normal simulation 25% faster.” Companies had to make approximations with their simulation technology or only look at sub-sections of an RFIC. But now, thanks to advancements in HFSS meshing and breakthroughs in ease of cloud computing via Azure, it is possible to envision the full-wave electromagnetic activity, and extract the coupled models, for an entire RFIC. Imagine the possibilities.
And we have the feeling that a few of you reading this will be eager to test this out as soon as possible…
For more info, visit www.ansys.com or email firstname.lastname@example.org