Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 37 additions & 1 deletion apps/website/messages/de.json
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,8 @@
"navExtra": {
"dashboard": "Dashboard",
"tree": "Forschungslab",
"docs": "Doku"
"docs": "Doku",
"about": "Über den Autor"
},
"theorems": {
"badge": "MATHEMATISCHE GRUNDLAGE",
Expand Down Expand Up @@ -1728,5 +1729,40 @@
"parameters": "Konsolidierte Parameter",
"optimizer": "Optimierer",
"adam": "ADAM"
},
"about": {
"name": "Dmitrii Vasilev",
"role": "Hardware-AI- und FPGA/RTL-Ingenieur · Autor des ternären GF-T-Gleitkommaformats · Spezifikation → RTL → FPGA · ML-Systeme und Edge AI",
"bioTitle": "Über mich",
"bio": [
"Ich entwickle den vollständigen Stack für effiziente KI-Berechnungen — vom Zahlenformat und dem ausführenden RTL bis zur bitgenauen Verifikation auf einem Board. Ich bin Autor von GF-T, einem ternären Gleitkommaformat, und der umfassenderen GoldenFloat-Familie (GF-T8/16/32; φ-abgeleitetes GF4→GF1024), veröffentlicht als Katalog mit 83 Formaten und bitgenauen Konformitätsvektoren.",
"Ich habe ein neues Zahlenformat aus einem arXiv-Paper auf eine echte Xilinx Artix-7 (ALINX AX7203, XC7A200T) gebracht — mit vollständig offenen Werkzeugen: Yosys, nextpnr und prjxray, ohne Hersteller-Tools. Ein SKY130-Design wurde über Tiny Tapeout zur Fertigung eingereicht; Messungen auf einem Chip gibt es nicht. Im Mittelpunkt steht die Kette Spezifikation → RTL → Verifikation → Board: Jeder RTL-Knoten wird bitweise gegen ein unabhängiges Python-Goldenmodell geprüft, sodass Abweichungen zwischen Spezifikation und RTL vor der Synthese auffallen.",
"Vor der Hardware habe ich mehr als zehn Jahre Software entwickelt und vermittelt — KI-Agenten und Multi-Agenten-Systeme, React Native und Web3. Ich bin weltweit (UTC+7) für Remote- und Vertragsarbeit in den Bereichen Hardware-AI, FPGA/RTL/Verifikation, ML-Systeme und Edge AI offen."
],
"achievementsTitle": "Wichtige Ergebnisse (Hardware, 2026)",
"achievements": [
"Neuronales Training auf dem FPGA — Primitive, die direkt auf dem Board trainieren; die gesamte Kette Spezifikation → RTL → Board ist bitweise gegen das Goldenmodell geprüft.",
"GF16-Matrixmultiplikation 4×4 — auf Xilinx-Artix-7-Logik mit 0 DSP48 und 0 Latches: 32.252 LUTs oder 21.223 bei Nutzung der Hard-Multiplizierer.",
"Tiny Tapeout SKY130 — das Design wurde zur Fertigung eingereicht: GDS, Gate-Level-Test und Precheck sind bestanden. Messungen auf einem Chip gibt es noch nicht.",
"Bitgenaue Verifikation pro Knoten — jeder RTL-Knoten wird gegen ein unabhängiges Python-Goldenmodell geprüft (iverilog plus KAT-Vektoren); Abweichungen zwischen Spezifikation und RTL werden vor der Synthese erkannt.",
"tri-net — ein vollständiger ternärer Netzwerk-Stack (133 .t27-Spezifikationen): GF16-PHY, ein BPSK-Modem über AD9361, Mesh-Routing und AEAD-Kryptografie — zwischen Geräten über Funk nachgewiesen."
],
"publicationsTitle": "Veröffentlichungen",
"publications": [
{
"title": "GoldenFloat: Eine φ-abgeleitete Gleitkommafamilie (GF4→GF1024) — arXiv:2606.05017",
"url": "https://arxiv.org/abs/2606.05017"
},
{
"title": "Ein numerischer Katalog mit 83 Formaten und bitgenauen Konformitätsvektoren: FP8, BF16, MXFP4 — arXiv:2606.09686",
"url": "https://arxiv.org/abs/2606.09686"
}
],
"links": {
"linkedin": "LinkedIn",
"github": "GitHub",
"arxiv1": "arXiv: GoldenFloat",
"arxiv2": "arXiv: Katalog mit 83 Formaten"
}
}
}
38 changes: 37 additions & 1 deletion apps/website/messages/es.json
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,8 @@
"navExtra": {
"dashboard": "Panel",
"tree": "Lab. Investigacion",
"docs": "Docs"
"docs": "Docs",
"about": "Sobre el autor"
},
"theorems": {
"badge": "FUNDAMENTO MATEMÁTICO",
Expand Down Expand Up @@ -1760,5 +1761,40 @@
"desc": "arXiv:2502.16473"
}
]
},
"about": {
"name": "Dmitrii Vasilev",
"role": "Ingeniero de Hardware-AI y FPGA/RTL · Autor del formato de coma flotante ternario GF-T · especificación → RTL → FPGA · sistemas ML y Edge AI",
"bioTitle": "Sobre mí",
"bio": [
"Construyo la pila completa para un cómputo de IA eficiente: desde el formato numérico y el RTL que lo ejecuta hasta la verificación bit a bit en una placa. Soy el autor de GF-T, un formato de coma flotante ternario, y de la familia más amplia GoldenFloat (GF-T8/16/32; GF4→GF1024 derivada de φ), publicada como un catálogo de 83 formatos con vectores de conformidad bit a bit.",
"Llevé un formato numérico nuevo desde un artículo de arXiv hasta una Xilinx Artix-7 activa (ALINX AX7203, XC7A200T) usando un flujo completamente abierto: Yosys, nextpnr y prjxray, sin herramientas del fabricante. Un diseño SKY130 fue enviado a fabricación mediante Tiny Tapeout; no hay mediciones en un chip. Mi trabajo se centra en la cadena especificación → RTL → verificación → placa: cada nodo RTL se comprueba bit a bit frente a un modelo dorado independiente en Python, para detectar divergencias entre la especificación y el RTL antes de la síntesis.",
"Antes del hardware pasé más de diez años desarrollando y enseñando software: agentes de IA y sistemas multiagente, React Native y Web3. Estoy disponible para trabajo remoto o por contrato en todo el mundo (UTC+7) en hardware-AI, FPGA/RTL/verificación, sistemas ML y edge AI."
],
"achievementsTitle": "Resultados principales (hardware, 2026)",
"achievements": [
"Entrenamiento neuronal en FPGA: primitivas que entrenan en la propia placa, con toda la cadena especificación → RTL → placa comprobada bit a bit frente al modelo dorado.",
"Multiplicación de matrices GF16 4×4: se implementa en la lógica Xilinx Artix-7 con 0 DSP48 y 0 latches: 32.252 LUT, o 21.223 usando multiplicadores dedicados.",
"Tiny Tapeout SKY130: el diseño fue enviado a fabricación; GDS, prueba a nivel de puertas y precheck han pasado. Aún no hay mediciones en un chip.",
"Verificación bit a bit por nodo: cada nodo RTL se comprueba frente a un modelo dorado independiente en Python (iverilog y vectores KAT), detectando divergencias antes de la síntesis.",
"tri-net: una pila de red ternaria completa (133 especificaciones .t27): PHY GF16, módem BPSK sobre AD9361, encaminamiento mesh y criptografía AEAD, demostrados entre dispositivos por radio."
],
"publicationsTitle": "Publicaciones",
"publications": [
{
"title": "GoldenFloat: Una familia de coma flotante derivada de φ (GF4→GF1024) — arXiv:2606.05017",
"url": "https://arxiv.org/abs/2606.05017"
},
{
"title": "Un catálogo numérico de 83 formatos con vectores de conformidad bit a bit: FP8, BF16, MXFP4 — arXiv:2606.09686",
"url": "https://arxiv.org/abs/2606.09686"
}
],
"links": {
"linkedin": "LinkedIn",
"github": "GitHub",
"arxiv1": "arXiv: GoldenFloat",
"arxiv2": "arXiv: Catálogo de 83 formatos"
}
}
}
38 changes: 37 additions & 1 deletion apps/website/messages/zh.json
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,8 @@
"navExtra": {
"dashboard": "仪表板",
"tree": "研究实验室",
"docs": "文档"
"docs": "文档",
"about": "关于作者"
},
"theorems": {
"badge": "数学基础",
Expand Down Expand Up @@ -1760,5 +1761,40 @@
"desc": "arXiv:2502.16473"
}
]
},
"about": {
"name": "Dmitrii Vasilev",
"role": "硬件 AI 与 FPGA/RTL 工程师 · GF-T 三进制浮点格式作者 · 规范 → RTL → FPGA · ML 系统与边缘 AI",
"bioTitle": "关于我",
"bio": [
"我构建高效 AI 计算的完整技术栈:从数值格式和执行它的 RTL,到开发板上的逐位精确验证。我是 GF-T 三进制浮点格式以及更广泛 GoldenFloat 系列(GF-T8/16/32;基于 φ 的 GF4→GF1024)的作者;该系列以包含 83 种格式和逐位精确一致性向量的目录形式发布。",
"我将一个 arXiv 论文中的新数值格式部署到实际的 Xilinx Artix-7(ALINX AX7203,XC7A200T)上,使用完全开放的流程:Yosys、nextpnr 和 prjxray,不依赖厂商工具。一个 SKY130 设计已通过 Tiny Tapeout 送交制造;目前没有芯片上的测量结果。我的工作围绕规范 → RTL → 验证 → 开发板这条链路展开:每个 RTL 节点都与独立的 Python 黄金模型逐位比对,在综合前发现规范与 RTL 的偏差。",
"在硬件工作之前,我用了十多年开发和教授软件:AI 智能体与多智能体系统、React Native 以及 Web3。我接受全球远程或合同工作(UTC+7),方向包括硬件 AI、FPGA/RTL/验证、ML 系统和边缘 AI。"
],
"achievementsTitle": "主要成果(硬件,2026)",
"achievements": [
"FPGA 上的神经网络训练:原语可直接在开发板上训练,规范 → RTL → 开发板的完整链路均与黄金模型逐位核对。",
"GF16 4×4 矩阵乘法:映射到 Xilinx Artix-7 逻辑,使用 0 个 DSP48 和 0 个锁存器:32,252 个 LUT;允许使用硬乘法器时为 21,223 个。",
"Tiny Tapeout SKY130:设计已送交制造;GDS、门级测试和预检查均已通过。目前还没有芯片测量结果。",
"逐节点逐位精确验证:每个 RTL 节点都与独立的 Python 黄金模型核对(iverilog 加 KAT 向量),在综合前发现规范与 RTL 的偏差。",
"tri-net:完整的三进制网络栈(133 个 .t27 规范),包括 GF16 PHY、基于 AD9361 的 BPSK 调制解调器、网格路由和 AEAD 加密;已完成设备间无线验证。"
],
"publicationsTitle": "公开发表",
"publications": [
{
"title": "GoldenFloat:基于 φ 的浮点格式系列(GF4→GF1024)— arXiv:2606.05017",
"url": "https://arxiv.org/abs/2606.05017"
},
{
"title": "包含逐位精确一致性向量的 83 种格式数值目录:FP8、BF16、MXFP4 — arXiv:2606.09686",
"url": "https://arxiv.org/abs/2606.09686"
}
],
"links": {
"linkedin": "LinkedIn",
"github": "GitHub",
"arxiv1": "arXiv:GoldenFloat",
"arxiv2": "arXiv:83 种格式目录"
}
}
}
Loading