HY

HYT DDI 0001 · r3p0

Personal datasheet

The one document genre I read every day, applied to its author. Values are typical unless measured; errata are permanent.

1Title block

Table 1 — Device identification
FieldValue
Part numberHYT-2026
DocumentHYT DDI 0001
Revisionr3p0
Issue2026-10
StatusActive. No drop-in replacement.
ConfidentialityNon-confidential. Distribution encouraged with attribution.

2Features

  • GEMM and attention kernels on ARM SME/SVE, in multiple precisions.
  • Core GEMM kernels sustain above 95% of theoretical peak.
  • SME2 sbgemm in upstream OpenBLAS — bf16 × bf16 → fp32, Apple M4.
  • HPC performance work across MPI and UCX; KBLAS and Eigen maintenance.
  • AI agents: tool use, remote execution, and agent workflows.
  • Model deployment on ARM: local inference, stability monitoring.
  • Bilingual documentation; both languages are load-bearing (see E002).

3Absolute maximum ratings

Table 2 — Absolute maximum ratings
Symbol Parameter Min Max Unit
Tsocial Continuous social engagement —3h
Cctx Concurrent projects —3—
Bunbench Tolerance for unverified performance claims —0—
Trec Recovery time after an interrupt during single-threaded thought —15min
Vcuda Systolic pressure during GPU comparisons —120mmHg

Stressing the device beyond these limits may cause permanent personality changes. Prolonged exposure at the limits may affect reliability.

4Recommended operating conditions

Table 3 — Recommended operating conditions
Symbol Parameter Min Typ Max Unit
Tdeep Deep-work block length 1.534h
Nctx Concurrent projects, recommended —12—
Sbench Claims backed by measurements 100——%
Treply Reply latency on written, asynchronous channels —1224h
Nmanual Architecture references within arm's reach 11—vols

Operation outside these conditions is supported, but its behavior is outside the documented envelope.

5Performance characteristics

Table 4 — Performance characteristics
Parameter Min Typ Max Unit Conditions
Core GEMM kernel, share of theoretical peak 95——% multi-precision asm · SME/SVE
SME2 sbgemm micro-kernel shape —64 × 16—— BFMOPA · bf16 × bf16 → fp32
ZA32 tiles accumulated per micro-kernel —4—tiles Apple M4 · OpenBLAS #6142
Users served by a self-hosted community service 3,000——users WeChat group bot · > 1 yr uptime

All values are as already published on this site; none were added for this document. Benchmarks beat adjectives — see Errata E001.

6Functional description

The device converts reference manuals into running kernels. Inputs: an architecture reference, a workload, a target. Output: assembly whose numbers hold up under re-measurement.

Inputs

Architecture references
Workloads and performance targets
Benchmarks that argue back

HYT-2026

GEMM and attention on SME/SVE

Tool use, remote execution, and agent workflows

Local inference, ARM platforms, and performance checks

Outputs

Systems that run
Numbers that repeat
Figure 1 — Block diagram

Education

  • 2019–2024Harbin Institute of Technology, Shenzhen · B.Eng. in Electronics and Information Engineering
  • 2021–2023KU Leuven · B.Sc. in Engineering Technology, joint program
  • 2023–2024KU Leuven · M.Sc. in Engineering Technology, Software Systems · Cum Laude

Experience

  • 2025–presentGEMM and attention on ARM SME/SVE: multi-precision assembly generation, core GEMM kernels above 95% of theoretical peak, and memory-access and compute optimizations across several attention variants.
  • Same periodMPI and UCX performance work in HPC settings, and maintenance of KBLAS and Eigen.
  • 2023–2024Stability after model deployment: monitoring input distributions and output anomalies to reduce concept drift.
  • 2022On-site research intern at Bloomberg, New York: data augmentation for event-driven stock-prediction sentiment analysis, later published in AEMR.
  • 2022Independent full-stack project: a WeChat group service bot for a voice-social community, serving 3,000+ users and running steadily for over a year.
  • 2021Research assistant at Peking University's Institute of Advanced Information Technology: a survey of low-cost phased-array antennas, plus microwave-anechoic-chamber experiments.
  • 2019–2020Student team project at HIT Shenzhen: a blockchain medical-records system on Ethereum, registered as software copyright (second author).

7Timing diagrams

One clock in years, four signals: education, work, open source, water. 1 2 3 4 5 6 7 8 CLK EDU WORK OSS WATER 2019 2020 2021 2022 2023 2024 2025 2026

WATER activity is real but its timing was never characterized; pulses are illustrative, the last one is not. The log is under beyond/.

Table 5 — Clock events
Edge Year Event
12019Enrolls at HIT Shenzhen — Electronics and Information Engineering.
22020First team project ships: a blockchain medical-records system.
32021Joint program with KU Leuven begins; research assistant at PKU.
42022Bloomberg internship; the WeChat bot goes live, 3,000+ users.
52023M.Sc. year at KU Leuven; model-deployment stability work.
62024Graduates — Cum Laude.
72025Current stepping: GEMM and attention on ARM SME/SVE; MPI/UCX; KBLAS, Eigen.
82026SME2 sbgemm reaches OpenBLAS upstream; six dive days in the Maldives.

8Pin description

Table 6 — Pin description
Pin Type Description
EMAIL I/O · async tomhoyt88 [at] gmail.com. Preferred channel for anything longer than two sentences.
GITHUB Output · open-drain github.com/0tomhoyt. Emits code, notes, and the occasional upstream patch.
LANG[1:0] Config 00 = English, 01 = Chinese, 1x = bilingual with English terms leaking through (see E002).
Tlocal Clock Local clock, UTC+8 typ. Drifts during deep work; see Recommended operating conditions.

9Ordering information

HYT-2026-r3p0

Table 7 — Part number decode
Field Value Meaning
Manufacturer HYT Manufacturer code (Hou Yuteng).
Model year 2026 Year of this document's issue.
Revision r3p0 Silicon stepping. Field updates arrive through experience, not firmware.

Minimum order quantity: one (1) conversation. This part is not sold through distribution — supported interfaces are listed in Pin description.

10Errata

Known defects, in the standard format. Everything not listed here is a feature.

E001Affects: all revisions

Description
Performance claims without a benchmark are instinctively rejected, even when the source is an authoritative document.
Workaround
Provide the measurement, the machine configuration, and N.
Status
WONTFIX (by design)

E002Affects: r2p0 onward

Description
Chinese technical writing carries a high density of English terms.
Workaround
None.
Status
WONTFIX

E003Affects: meetings > 3 h, consecutive

Description
Output quality degrades, without the device noticing.
Workaround
Split the meeting, or move it to a written asynchronous channel.
Status
Under investigation

E004Affects: r1p0–r2p0

Description
Any topic could be steered toward the roofline model.
Workaround
None required.
Status
Fixed in r3p0 (mostly)

11Revision history

Table 8 — Change history
Rev Date Changes
r1p0 2019 Initial release: enrolls at HIT Shenzhen, B.Eng. Electronics and Information Engineering (to 2024).
r2p0 2021 New subsystem: joint program with KU Leuven; B.Sc. 2023, M.Sc. Software Systems 2024, Cum Laude.
r3p0 2025 Current stepping: GEMM and attention on ARM SME/SVE; MPI/UCX; KBLAS and Eigen maintenance; OpenBLAS upstream.

12Legal notice

Arm, SME, SVE, and ZA are trademarks of Arm Ltd. HYT DDI 0001 is a personal page written in the style of an architecture reference manual; it is not affiliated with Arm, and no silicon was harmed in its production.

All performance figures are as published on this site and are not guarantees. Returns are not accepted; read the Errata before deployment.