
Every interface on a chip needs something on the other side of it in simulation: a model that drives traffic, a monitor that watches it, checkers that catch protocol errors and coverage that shows what was exercised. That bundle is verification IP, and a team can get it in three ways. It can buy it from a vendor, build it in house, or borrow it from an open-source project or a partner.
The choice is usually made on licence price alone, which is the wrong place to start. This post sets out a decision method that weighs labour, licence, schedule and the risk of a protocol bug reaching silicon. It includes a worked cost comparison. The inputs in that comparison are illustrative, so the ranking it produces depends on them and is not a general verdict. On those inputs, buying costs about $825,000 over three years against $910,000 for borrowing and $1.44 million for building, once risk is priced in.
Key takeaways
- The deciding cost is usually not the licence. It is the engineering time to build and maintain the IP, and the expected cost of a protocol bug it fails to catch.
- In the illustrative example, borrowing has the lowest direct cost ($460,000), but buying wins once respin risk is added. Borrowing comes out ahead if an escape would cost less than about $3.1 million.
- Building makes most sense for interfaces nobody sells: proprietary links, custom accelerators and draft standards.
- Whatever the source, the team still owns the verification plan, integration and coverage closure.
Why the choice matters now
Verification is where projects succeed or fail, and the industry's own survey data shows how hard it has become. The 2024 Wilson Research Group IC/ASIC functional verification trend report, written by Harry Foster of Siemens EDA and dated 17 February 2025, found that only 14% of projects achieved first-silicon success, the lowest level in two decades of the study. It drew on 597 eligible participants.
The next edition looks worse. In a Siemens Verification Horizons blog post dated 28 October 2026, Foster wrote that reported first-silicon success for IC/ASIC projects had fallen to 5% in the 2026 study, and that about three in four IC/ASIC and FPGA projects were missing their original schedules. The two figures come from different survey years and different respondent groups, so the drop should be read as a direction rather than a precise change. The same post makes clear that the study does not establish a single cause.
Standards also keep moving. The UCIe Consortium's 3.0 specification, described on its site on 3 September 2025, added 48 GT/s and 64 GT/s data rates beyond 32 GT/s while keeping backward compatibility. That is good for designers and more work for verification: every older mode still has to be checked.
So the real question is what it costs to be sure an interface works, on this project and the next two.
What buy, build and borrow really mean
The three options are less tidy than the labels suggest. Each is a bundle of choices about who writes the code, who keeps it current and who answers when it is wrong.
| Buy | Build | Borrow | |
|---|---|---|---|
| Source | Commercial VIP from an EDA or IP vendor | Written by the project or a central methodology team | Open-source projects, a partner's testbench, or models shipped with design IP |
| Upfront cost | Licence, often annual | Large labour cost before first use | Low, plus the effort to fill gaps |
| Who tracks the spec | Vendor, subject to the contract | Your team | The project's maintainers, or nobody |
| Typical gaps | Fit with your flow, licence limits on seats or emulation | Spec interpretation shared with the design team, so errors can match | Protocol checking, coverage models and documentation |
| Best fit | Published standards with large, complex test suites | Proprietary or unreleased interfaces | Simpler buses, processors with open ecosystems, early exploration |
Two points in that table deserve more space.
The first is shared interpretation. If one organisation reads the specification to write both the RTL and the checker, a misreading can end up in both, and the bug passes. Independent IP gives a second reading of the standard. This is the main technical argument against building, separate from cost.
The second is that "borrow" covers very different things. Two examples show the range. The CHIPS Alliance riscv-dv project is a SystemVerilog and UVM instruction generator for RISC-V processor verification under the Apache-2.0 licence. Its README states that it has been verified with Synopsys VCS, Cadence Incisive and Xcelium, Mentor Questa and Aldec Riviera-PRO, and it needs a simulator with UVM 1.2 support. At the other end, cocotbext-axi, an MIT-licensed Python package, provides AXI, AXI-Lite, AXI stream and APB models for the cocotb framework. Its repository page describes simulation models and does not describe protocol compliance checking. Neither should be treated as a full VIP until the team has checked what it checks.
Our earlier post on hardware verification languages and verification IP covers how VIP grew out of constrained-random testbenches.
A simple cost and risk model
The model has four parts, and each one is easy to estimate roughly.
- Labour. Engineer-months to create, maintain and integrate the IP, multiplied by a loaded monthly cost. Creation happens once. Maintenance recurs every year the IP is in use. Integration recurs on every project.
- Licence. Cash paid to a vendor. Check how it scales: per year, per seat, per project, and whether emulation use costs extra.
- Risk. The probability that a protocol bug escapes to silicon because the IP missed it, multiplied by the cost of that escape. A respin, a delayed launch or a field fix all count.
- Schedule. The calendar time before the IP is usable. This is easy to forget because it does not show up as cash, but a testbench that is not ready delays everything that depends on it.
Expected cost is labour plus licence plus risk. Schedule is kept separate because it rarely converts cleanly into money, and teams tend to argue about the conversion rate instead of the decision.
The weakest input is the escape probability, which nobody knows precisely. The model still forces the team to say whether it believes a home-grown or open-source checker is as likely to catch a bug as a commercial one. If yes, the risk terms cancel. If no, the risk term often dominates.
Worked example on illustrative inputs
Here is the model applied to one protocol interface used on three projects over three years. Every input below is an illustrative assumption, not a vendor price, a survey result or data from any real programme. Replace them with your own figures. All results were computed with a short script.
| Input (illustrative) | Buy | Build | Borrow |
|---|---|---|---|
| Loaded cost per engineer-month | $20,000 | $20,000 | $20,000 |
| Creation or gap-filling effort | 0 | 18 engineer-months | 8 engineer-months |
| Maintenance per year | 1 engineer-month | 6 engineer-months | 3 engineer-months |
| Integration per project | 1.5 engineer-months | 2 engineer-months | 2 engineer-months |
| Licence per year | $150,000 | $0 | $0 |
| Chance of a protocol escape per project | 1.5% | 4% | 3% |
| Cost of one escape | $5 million | $5 million | $5 million |
Buying still carries some maintenance, for tracking vendor releases. The escape probabilities reflect the shared-reading risk of building and the thinner checking of borrowed code.
Over three years and three projects, the totals work out as follows.
| Result (illustrative) | Buy | Build | Borrow |
|---|---|---|---|
| Total engineer-months | 7.5 | 42 | 23 |
| Labour cost | $150,000 | $840,000 | $460,000 |
| Licence cost | $450,000 | $0 | $0 |
| Direct cost (labour plus licence) | $600,000 | $840,000 | $460,000 |
| Expected escape cost | $225,000 | $600,000 | $450,000 |
| Expected total | $825,000 | $1,440,000 | $910,000 |
To check one line: building takes 18 engineer-months to create, plus 6 a year for three years, plus 2 for each of three projects, which is 42. At $20,000 each, that is $840,000. Its expected escape cost is 4% of $5 million on each of three projects, or $600,000.
Three things stand out.
Borrowing is cheapest until risk is counted. On direct cost alone, borrowing beats buying by $140,000. The higher escape probability then adds $225,000 more expected cost than buying, which reverses the order. The break-even is easy to find: borrowing and buying cost the same when one escape costs about $3.1 million. Below that, on these inputs, borrowing wins. A small controller on a mature node with a cheap mask set might sit below that line. A large SoC on an advanced node almost certainly sits above it.
The licence can rise a long way before buying loses. With all other inputs fixed, buying matches building on expected total at a licence of about $355,000 a year, and matches borrowing at about $178,000 a year. On direct cost alone, ignoring risk, the break-even licences fall to $230,000 a year against building and about $103,000 a year against borrowing. A licence quote is therefore best read against these thresholds rather than against zero.
Building also costs calendar time. Eighteen engineer-months spread across three engineers means about six months before the IP is ready. If the RTL is ready sooner, the interface is verified late or with a partial model, and both raise the risk term.
The model has limits. It treats the licence as flat, while many licences scale with seats or emulation use. It ignores the in-house knowledge a built VIP creates. It uses one escape cost, when a bug that firmware can work around costs far less than one it cannot. Each can be added to the model.
A decision method in six questions
The worked example gives the arithmetic. These questions give the order in which to ask them.
- Is the interface a published standard? If not, there is nothing to buy. Choose between building and asking the design IP supplier for its test environment, then skip to question five.
- How large and fast-moving is the standard? Frequent revisions, many speed modes and backward compatibility are where vendor upkeep earns its fee. A small, stable bus suits building or borrowing.
- What does one escape cost on this product? Put a rough number on it. It is the input most likely to flip the result.
- What does the open-source option actually check? Read the code, list the rules it asserts and the coverage it collects, and price the gaps honestly.
- Who will maintain it in two years? Built and borrowed VIP depend on named people staying.
- Will it run where you need it? If the project plans emulation, check for transactors and whether that use is licensed separately. Our post on acceleration and emulation for hardware and software integration explains why many teams need both platforms.
Then run the cost model with your own inputs. If two options land within 10% to 15% of each other, that is inside the error of the estimates, and questions five and six should decide.
If you buy, what to check
Vendors publish their own selection criteria, and they are a useful starting checklist as long as the source is kept in mind. A Siemens EDA Verification Horizons post by Gordon Allan, dated 18 May 2026, argues that teams usually do better buying than building, mainly because of ongoing maintenance and the opportunity cost of scarce verification engineers. It lists breadth of standards coverage, how early the VIP tracks new specification versions, reuse across simulation and emulation, UVM support, memory model quality and support for specialised requirements. That is a vendor's view of its own market, and it should be read as such.
A buyer's checklist, in our own words, adds a few items that a vendor list is less likely to stress:
- Checker visibility. Ask for the list of protocol rules the IP checks, with references to specification sections. Unchecked rules are where escapes come from.
- Coverage model. Confirm the coverage groups map to the specification features in your verification plan, not only to the vendor's own test list.
- Debug. Run a short trial on your own design and judge the error messages and transaction views.
- Licence terms. Check seats, emulation use, regression limits and what happens when a new revision ships.
- Exit plan. Tests written against one vendor's sequence library may not carry across.
If you build or borrow, how to keep the risk down
The main danger with building is the shared reading of the specification. Three habits reduce it:
- Have someone outside the design team write the checker from the specification alone.
- Write protocol rules as assertions that a formal tool can also run, to reach behaviour simulation may miss. Our earlier post on formal verification in semiconductor engineering explains how it works.
- Keep a traceability table from each specification clause to the check or coverage point that covers it, and review it at each milestone.
For borrowed IP, treat the code as something the team now owns. Pin the version, record the licence, test the IP before trusting it, and budget the gap-filling work. Code from a partner needs a written agreement on reuse.
The SoC lifecycle on The Semiconductor Hub shows where verification planning and IP selection sit in a programme, which helps when deciding how early this choice has to be made.
FAQ
Is a bus-functional model the same as verification IP?
No. A bus-functional model drives and responds to transactions. Verification IP adds monitors, protocol checkers, coverage and often a test suite. Comparing a model's cost with a full VIP's price makes building look cheaper than it is.
Can we mix the options?
Yes. Run the cost model separately for each interface. Large published standards, proprietary links and simple buses may each land on a different answer.
Does open-source licensing matter for verification code that never ships?
It can. Permissive licences such as Apache-2.0 and MIT have conditions on notices and redistribution, and testbench code is sometimes shared with partners or customers. This is not legal advice. Check the licence terms with whoever handles licensing in your organisation.
Sources
- Harry Foster, Siemens EDA and Wilson Research Group, "IC/ASIC Functional Verification Trend Report, 2024", Verification Academy, 17 February 2025.
- Harry Foster, "Are We Looking Beyond the Numbers in the 2026 Functional Verification Study?", Siemens Verification Horizons blog, 28 October 2026.
- Gordon Allan, "Verification IP: Why design teams buy instead of build", Siemens Verification Horizons blog, 18 May 2026.
- UCIe Consortium, "UCIe 3.0 Specification: Redefining Chiplet Interconnects", uciexpress.org, 3 September 2025.
- CHIPS Alliance, riscv-dv repository README, GitHub, read 10 November 2026.
- Alex Forencich, cocotbext-axi repository README, GitHub, read 10 November 2026.
The worked example uses illustrative inputs, and its ranking of the three options follows from those inputs. Vendor statements are the companies' own views, as published on the dates shown. Survey figures are as reported by the study author on the dates shown.


Your comments will be moderated before it appears here.