Benchmarks Available at Launch

[jcork-intel]

SUITE: Publish benchmarks with the releases they describe

What we are asking for

Publish validated performance benchmarks at the same time as the releases they describe, rather than afterwards as an ongoing gate to call the work done.

Specifically for Q4 2026 that means Order Accuracy, Storewide Loss Prevention and Voice-Enabled Interactions, on current target hardware. It is critical that we have benchmarks for Nova Lake results at product launch for Voice-Enabled Interactions ,Loss Prevention, and Order Accuracy.

Why it matters

Someone evaluating one of our repos needs performance numbers at the moment they are evaluating it. When the numbers arrive a quarter after the release, the people who most needed them have already moved on, or have gone and asked a competitor instead.

Publishable also has to mean reproducible. A figure a partner cannot regenerate on their own hardware does not do the job these repositories exist to do. Every number we publish should ship with the command that produces it and the configuration it ran on.

Two dates make this urgent. NRF is in January 2027, and the Nova Lake launch follows its own schedule. Results need to exist before those, not during them.

What is in scope

Repo	Panther Lake	Arrow Lake	Wildcat Lake	Nova Lake
Order Accuracy	yes	yes	yes	yes, by end of year
Storewide Loss Prevention	yes	yes	yes	no
Voice-Enabled Interactions	yes	yes	yes	yes, by end of year
Loss Prevention	already covered	already covered	yes	by end of year

This is an ongoing request rather than a one-off. Existing and new pipelines are expected to be tested across Arrow Lake, Wildcat Lake, Panther Lake and Nova Lake as they come. SWLP will eventually want to have benchmarks for NVL but trying to prioritized so I haven't scoped that into the end of the year deliverable.

What is not in scope

• Which figures marketing chooses to use, and how they are presented.
• Benchmarking anything other than Intel hardware.
• Hardware procurement scheduling, which is tracked separately.

---

Acceptance Criteria

1. Benchmarks are published at the same time as their Q4 releases, in the repo and reproducible by an end user.
2. Specific SKUs listed in our benchmarking hardware matrix are targeted
3. Performance Metrics are generated using our standard methodology that we've built.
4. Nova Lake results are available for Voice-Enabled Interactions & Loss Prevention (including the ASC use cases).
5. Results have PDT approval.
6. Everything is published and referenceable ahead of NRF.

---

[jcork-intel]

@FarhaanMohideen --> Benchmarks request draft take a look at the tabled above please and let me know if you are getting any signal that says differently or want a specific prioritization.

[avinash-palleti]

Hi @jcork-intel ,
For kiosk usecase we dont have stream density logic, we display runtime benchmark values on application UI for understanding platform capabilities.

What is the expected publishable numbers for this usecase.

[jcork-intel]

Hi @jcork-intel , For kiosk usecase we dont have stream density logic, we display runtime benchmark values on application UI for understanding platform capabilities.

What is the expected publishable numbers for this usecase.

@avinash-palleti

I have the detailed out in the epic: intel-retail/voice-enabled-interactions#28

Latency under load: Time to First Audio is the primary.

This is the gist of what is in the epic:
Time to First Audio under 500 ms at p95. This is the primary gate.
LLM first token under 200 ms, LLM tokens per second, ASR latency, TTS first audio.
Word Error Rate at 55, 65, 75 and 80 dB.
Standard resource utilization: CPU / GPU / NPU / memory / power reported alongside those.