What is memory yield analysis?

It estimates what share of manufactured memories will meet specification despite process variation. Because a memory repeats the same bit-cell across the whole array, even a failure rate of one in a billion matters: CEA-Leti’s project set out to check whether such rates guarantee acceptable yield for SRAMs above the megabyte range.

Why is analyzing a memory slice not enough?

The usual approach extracts memory slices or critical paths, simulates them with limited Monte Carlo analysis and extrapolates the results, which can introduce significant error in the yield estimate. Fast SPICE simulation may compromise accuracy, and combining individual block sigma-corners may give unrealistic results. XMA instead rebuilds the entire memory and simulates it with the golden SPICE simulator.

Is VarMan XMA a separate product or an option?

It is an option, not a separate product. VarMan XMA is an option to VarMan Memory, and it uses VarMan’s Fast Monte Carlo and high-sigma kernels. With the option, VarMan Memory gains three analyses that cover the entire memory: XMA Monte Carlo, XFail analysis and yield verification.

Which memory technologies has XMA been verified on?

Silvaco says XMA has been verified on leading planar CMOS, FDSOI and FinFET foundry technologies. In a project announced in 2019, CEA-Leti set out to use it on a 256 Kb ultra-low-voltage SRAM, and to assess its predictions against silicon from STMicroelectronics’ 28 nm UT-FDSOI process.

Does XMA work with third-party SPICE simulators?

Yes. XMA works with industry golden SPICE circuit simulators, so the whole memory is simulated with the reference simulator your team already trusts rather than a faster approximation. Its published materials name no particular simulator, so confirm yours with our technical experts before you plan an evaluation.

How long does it take to build a full memory analysis?

Building the analysis for an entire memory takes only minutes, according to the product page and brief; Silvaco publishes no more precise figure. The simulations then run on top of that. In the brief’s SRAM example, 10,000 simulations covered 100,000 instances of an 8K by 80-bit memory.

VarMan XMA: full-chip memory yield analysis

Estimate yield for the whole memory, not a slice of it. VarMan XMA rebuilds the full array and runs Monte Carlo across it, so the yield you report comes from simulation rather than extrapolation.

Talk to our Technical Experts Read the VarMan XMA brief

What VarMan XMA does

VarMan XMA runs memory yield analysis on the entire memory, with your golden SPICE simulator doing the work. At full size that's normally impractical. So the usual workaround simulates a slice or a critical path and projects the result. The projection is where the error comes from. Fast SPICE trades accuracy for runtime, and combining per-block sigma corners can produce a yield the silicon doesn't match.

Capabilities

The whole memory, held to the accuracy you'd expect from a single-cell run.

Whole-memory analysis

Banks, word-line drivers, local I/O, logic and the controller all go into the run. XMA assembles the memory from those blocks in minutes, then simulates all of it.
Method

VarMan's Fast Monte Carlo and high-sigma kernels, applied to the whole memory for fast-fail detection and yield estimation. Failures surface up to 100X faster than with classical Monte Carlo analysis.
Validation

Verified on FinFET, FDSOI and planar CMOS foundry processes. XMA runs alongside the golden SPICE simulator in your flow.
Interface

Set up and run the analysis from a graphical interface. Results show the statistical spread behind each yield estimate. You can read them without a statistician on the team.

Works with

You keep your design and your simulator. XMA changes how much of the memory they cover.

Takes in

A memory design and a golden SPICE simulator. The reference-accuracy simulator your team already trusts.

Extends

VarMan, the base product. XMA is an option to VarMan Memory and runs on VarMan's analysis kernels.

VarMan: variation-aware design for analog, RF, standard cell, I/O and memory teams, with up to 30X the simulation performance of traditional Monte Carlo

Applications

Where a slice-based estimate costs you the most.

SRAM and embedded memory yield estimationSRAMs and large embedded memories, where one rare bit-cell failure can decide whether a whole instance passes
Full-chip memory yield verificationA yield figure checked across the complete memory before you commit it to silicon
Fast-fail detection across a whole memoryRare failures located anywhere in the array, with one simulation standing in for many instances

Proof and resources

Check the numbers yourself.

Press releaseJune 2019

CEA-Leti and Silvaco team up on yield prediction project for ultra-low-power static memories

CEA-Leti chose VarMan XMA for fail detection and yield estimation on a 256 Kb ultra-low-voltage, ultra-low-leakage SRAM. The project planned to test its predictions against silicon from STMicroelectronics' 28 nm UT-FDSOI process.
Product brief · PDF

VarMan XMA Option – Full-chip Memory Yield Analyzer

The SRAM case in numbers. An 8K × 80-bit memory, 100K instances (65,536M equivalent runs), 10,000 simulations, two failures found.
Product brief · PDF

VarMan – Statistical Variation and Yield Analyzer

What each VarMan edition analyzes, including the three analyses the XMA option brings to VarMan Memory.
WebinarSeptember 2018

Memory Statistical Characterization Solution with VarMan

Dr. Jean Baptiste Duluc shows how VarMan's tools, XMA among them, cut the simulation count for high-sigma memory characterization.

Evaluate and engage

Send us a memory design and the yield target it has to hit. An application engineer runs XMA on it with you and walks you through the result for the full memory.

Talk to our Technical Experts Read the documentation Support and Downloads

Questions about memory yield analysis with VarMan XMA

Memory yield analysis predicts how many manufactured memories will meet spec once process variation is counted in. Because the same bit-cell repeats across the whole array, even a one-in-a-billion failure rate matters. CEA-Leti's XMA project set out to test whether rates that low still guarantee acceptable yield for SRAMs above a megabyte.

A slice or critical path leaves most of the memory unsimulated, so the yield figure is extrapolated, and the error can be significant. Fast SPICE cuts runtime at the cost of accuracy, and block-by-block sigma corners can combine into an unrealistic result. XMA simulates the whole rebuilt memory with golden SPICE instead.

An option. VarMan XMA gives VarMan Memory three analyses that span the entire memory: XMA Monte Carlo, XFail analysis and yield verification. They run on the same analysis kernels as base VarMan, which remains the tool for variation work on individual cells and blocks.

Three foundry technology families: planar CMOS, FinFET and FDSOI. In 2019, CEA-Leti announced a project applying XMA to a 256 Kb ultra-low-power SRAM, with measured 28 nm UT-FDSOI silicon from STMicroelectronics as the check on its predictions. For a node outside those families, our technical experts can tell you where it stands.

Yes. XMA works with the golden SPICE circuit simulators used across the industry, so the full memory runs at reference accuracy rather than through a Fast SPICE approximation. Name the simulator your flow depends on, and we'll confirm the setup before you scope an evaluation.

The build takes minutes, and XMA then keeps the simulation count small. In one SRAM run, 10,000 simulations covered 100,000 instances of an 8K × 80-bit memory, equal to 65,536M individual runs, and found two failures. You get a full-memory yield figure without simulating every instance one by one.

VarMan XMA: full-chip memory yield analysis

What VarMan XMA does

Capabilities

Whole-memory analysis

Method

Validation

Interface