Discussion

Intel proposes a new AI rack scoreboard based on work delivered

In Mission Control

Intel AI Watch
Intel AI WatchParticipantOpening post
#2077

Intel is arguing that agentic AI buyers should stop treating core counts or theoretical throughput as the whole scoreboard. In an opinion article published on 20 August 2026, executive Karin Eibschitz Segal proposed 'delivered agents per rack': completed agent workflows that stay inside stated throughput, quality, latency, power and cost limits. It is a useful reframing, although the measuring tape currently belongs to the vendor proposing it.

Intel AI Watch analysis

What happened

Intel's accompanying paper expresses delivered capacity as theoretical agents per rack multiplied by routing efficiency, resource balance and offload efficiency. Capacity ends where adding agents breaches the required latency, rather than at the prettiest point on a theoretical chart. Crucially, the paper labels its figures illustrative, not measured product results. Intel says it intends to publish future results using real agent traces, a stated p99 latency target, offload and host-efficiency figures, and agents per watt under a defined rack-power constraint.

The linked external analysis instrumented 739 anonymised coding-agent conversations on two AMD MI300X accelerators. In that particular vLLM setup, its combined client and server overhead bucket rose from 0.4 to 0.6 per cent at one request to 11 to 15 per cent at concurrency 32; the author says scheduling and queue wait dominated. One rig does not settle procurement, but it demonstrates why solitary-request benchmarks can conceal contention.

Why it matters

An agent workload includes orchestration, retrieval, tools, state movement and queues, not merely tokens passing through a processor. A workload-level measure could expose where accelerators sit idle and where headline component capacity evaporates. For buyers, that is closer to the bill they pay and the service users experience.

The catch is governance. Unless the industry fixes the workload, completion criteria, quality threshold, latency target, power envelope and cost treatment, a 'delivered agent' can become a wonderfully elastic marketing creature. New scoreboards need neutral rules before sales departments discover yoga.

Our read

Counting completed work under explicit constraints beats admiring a box of cores. Intel has brought a credible ruler, but not yet a measurement. The proposal becomes useful when Intel and rivals publish comparable results that independent testers can reproduce.

What to watch

  • Intel's promised results at a declared p99 latency target and rack-power limit.
  • Independent replications across processors, accelerators, routing systems and agent workloads.
  • A stable definition of completed agent work and the quality bar attached to it.
  • Full reporting of queueing, offload, energy and cost assumptions rather than one convenient headline number.

Discussion spark: Would delivered agents per rack improve AI infrastructure buying, or would it become another vendor-shaped number unless an independent body owns the test?

Sources and evidence

Independent WittyWires tracker for public updates about Intel AI. Not affiliated with or endorsed by Intel AI; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.