Inference Silicon & OS

Built
For One
Thing.

Modern GPUs carry silicon, software, and power overhead built for broader workloads. Inference is bottlenecked by memory bandwidth, not compute. We built the chip and the OS for the one thing that matters: AI inference.

FPGA-Validated · May 2026
353.4mm² · 41.2% UTIL
TLE_ARRAY 128× · 162.4mm² · 18.2W 6×6 TILE / 8b·MX WEIGHT_CACHE_SRAM 96MB · 142.6mm² · 6.8W SLC · 11.2 TB/s ASE 28.4mm² · 1.9W ATTN / SOFTMAX CTRL_FABRIC 19.6mm² · 0.8W NoC / DMA HTU 24.0mm² · 1.5W HOST / TOKEN UNIT HBM4_BOND_AREA · SoIC-X 8× STACK · 512GB · 16 TB/s S0 S1 S2 S3 S4 S5 S6 S7
265tok/s
Sustained Throughput
30W
Passive TDP
512GB
HBM4 Capacity
16TB/s
Bandwidth

Platform

K-CORE

265 tokens/second at 30 watts.

128 Tensor-Linear Engines on TSMC A16 with SoIC-X 3D bonding to 512GB HBM4. No training paths, no graphics legacy. Just the inference dataflow.

In our measured open-stack benchmark configuration: 5.25× to 5.62× the throughput of H100 and 3.71× to 3.98× the throughput of B200, at less than 1/20 the power. Low enough to deploy anywhere. No liquid cooling, no special power circuits, no facility upgrades.

K-OS

5,300 lines. Zero unnecessary code.

Three execution contexts. Fewer than 8 interrupts/sec under load. The kernel paths Linux disables for latency? We never wrote them. The schedulers it papers over? Not present.

0.5s cold boot to first token. Deterministic memory layout. The stack you'd build if you started from the workload instead of POSIX.

VALIDATIONMay 2026

Validated on AWS FPGA hardware.

First Transformer Layer Engine stage synthesized and validated on AWS f2.6xlarge (AMD VU47P FPGA). End-to-end Verilator-to-hardware match, against the same reference vectors the simulator validated.

3 / 3 strict tests PASS · Cycle-exact at the count claimed · Bit-correct within design tolerance
Technical Paper
K-Core & K-OS · Architecture and Measured Performance
Zenodo · v9.1 · May 2026 · DOI 10.5281/zenodo.20194507
Contact
Technical due diligence
Investor inquiries
Technical paper