Klangkerne^™

One tile. Two dies.
Four power domains.

An 18 mm² Klangkerne base die on Intel 14A-E, and a 7.2 mm² SoRT ray-tracing die on Intel 04A, joined by a Foveros 2.0 hybrid bond at 0.4 µm pitch. Total height added to the package: thirty-eight microns.

Klangkerne began as the codec that ended codecs — mixer, DSP grid, Beosonic tuning and analog output on one tile. The second generation stacks a dedicated acoustic ray tracer directly above it. SoRT walks the same BVH the GPU uses for light, carries an 8-octave-band payload per ray, and hands a parametric impulse response across the hybrid bond into Klangkerne's FDN and convolution engines. The output stage — Hallraum — then renders that field faithfully to whatever transducer is actually on your head.

Every subsystem sits in one of four independently gated power domains. Nothing shares a rail it doesn't need, and nothing wakes unless it is addressed.

Rays above.
Waves below.

Geometric acoustics fails below roughly 200 Hz in small rooms — wavelengths grow to room scale, and rays stop being an honest model of sound. Tiefton is the named architectural answer.

Below the crossover, a room stops behaving like a hall of mirrors and starts behaving like an instrument: discrete standing-wave modes dominate, and no number of ray bounces will find them. So Klangkerne doesn't trace there. At bake time, an offline FEM solve runs per reverb zone and reduces each one to a modal LUT — up to 48 modes, each stored as frequency, Q, and per-band coupling gain. 288 bytes per zone. About four seconds of bake per zone on a Core 500 development machine.

At runtime, those modes drive 48 biquad resonators in Klangkerne's DSP grid, drawing 0.9 mW. The crossover to geometric ray tracing sits at 180 Hz, blended at 24 dB/oct across the 140–220 Hz region. Where the ray model fails, Klangkerne doesn't paper over the physics — it names the limit and builds for it: a debt, promoted to a subsystem.

The field is computed.
Hallraum makes it true.

Ray tracing in, DSP out. At ear-canal and pinna scale, sound is wave-dominated — so the output stage is not a second ray tracer. It is measured-transfer-function DSP: transducer inversion, HRTF convolution, and a codec bridge, applied to the field SoRT computed.

Hallraum knows what is on your head. First-party devices — Sony, Bang & Olufsen, OPPO, Asus — ship signed transducer profiles, handshaken over the OPPO networking tile's BT/USB negotiiation and loaded from ZAM in under 0.2 ms. The deeper the profile, the closer the rendered field gets to the computed one.

The HRTF stage ships a 720-direction set — 2 ears × 256-tap FIR at 24-bit, about 1.1 MB per profile — with bilinear interpolation between the nearest four directions. Photogrammetry personalization applies as a delta layer over the base set. And the codec bridge encodes Dolby Atmos and DTS:X objects directly from SoRT's parametric domain: no render-to-channels-then-reanalyze double loss.

The honest limit: a non-individualized HRTF is the realism ceiling. Front/back confusion with a generic profile is a property of human hearing, not of the silicon — the hardware renders its transfer function faithfully; the function worth rendering is per-human. That is why the individualized HRTF slot exists in the first-party tier.

Four ways to trace a room.
One draws 118 milliwatts.

Scene: 64 sources × 32 rays × 3 bounces, 48 kHz / 128-sample blocks. The same acoustic workload run as CPU software, GPU compute, time-sliced on the GPU's RTU4 units, and on SoRT.