Source linked

Intel Xeon 6+ lanza telemetría de energía de hardware a nivel de núcleo

chipsandcheese.com@systems_wire2 days ago·Systems Engineering·0 comments

La nueva característica AET de Intel le permite rastrear el consumo de energía por núcleo desde el hardware, dejando de lado los modelos de potencia basados en software que añaden impuestos y errores.

intelxeon 6 plusapplication energy telemetrydatacenterperfcpu architecture

Intel’s Xeon 6+ ships a hardware-level energy telemetry system that exposes per-core power consumption directly to the OS — no software models, no extra tax, no guesswork.

Product director Kira Boyko confirmed at Computex 2026 that AET (Application Energy Telemetry) is a hardware hook into each core, plumbed straight into Linux perf and compatible with Intel’s toolchain. It works out of the box on every Xeon 6+ SKU, and Intel is rolling it to all future Xeons.

Hardware Telemetry Beats Software Power Models

Until now, datacenter operators relied on software approximations to estimate per-workload energy. Those models are trained on specific benchmarks and fail when real applications differ — introducing systematic error and runtime overhead. Boyko told Chips and Cheese that AET eliminates that: “hardware-level hook into the core that allows tracking” with no additional work from the user.

For the first time, you can see exactly what a single integer thread or FP-heavy kernel draws, down to the core. Boyko clarified that AET also exposes package-level and uncore energy, so you can drill from the whole SoC to a single core.

Wiring Per-Core Energy Into Orchestration

The killer use case is chargeback. Boyko described customers wanting to offer visible, per-tenant energy billing or rebates to shift utilization. With AET, a cloud provider can attribute energy cost to individual VMs or containers at hardware granularity, not via a model that might undercount a bursty AI inference workload.

Xeon 6+ uses a tile architecture: 72 clusters of 4 cores each (288 cores total on the flagship 6990E+). Boyko confirmed AET works at both the core and cluster levels. That granularity is enough to do fine-grained scheduling and power capping without reaching for external sensors.

Cross-Pollination from Datacenter to Client

Boyko noted that Xeon 6+ is the first 18A Xeon — the same process node Intel’s client chips already use. Features like AET often cross over, and the teams share “quite a bit” of information. If AET proves itself in the datacenter, expect it to appear in consumer silicon, just as Intel’s RAPL and per-core frequency controls migrated from server to client years ago.

AET is shipping now on all Xeon 6+ SKUs. The only remaining question is how fast the orchestration ecosystem — Kubernetes, SLURM, hyperscaler schedulers — will adopt the new perf counters to turn raw energy data into cost.


Source: An Interview with Intel's Kira Boyko: Xeon 6+'s Product Director
Domain: chipsandcheese.com

Read original source ->

External source stays available while the OJO article and comment thread stay local.

Comments load interactively on the live page.