{"id":184029,"date":"2026-10-29T11:40:41","date_gmt":"2026-10-29T09:40:42","guid":{"rendered":"https:\/\/prohoster.info\/blog\/news\/qwenputer-redakcziya-ai-modeli-qwen3-5-sposobnaya-simulirovat-risc-v-cpu-dlya-vypolneniya-koda-na-c-i-micropython"},"modified":"2026-10-29T11:40:46","modified_gmt":"2026-10-29T09:40:46","slug":"qwenputer-a-qwen3-5-ai-model-variant-that-simulates-a-risc-v-cpu-to-run-c-and-micropython-code","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sk\/blog\/news\/qwenputer-a-qwen3-5-ai-model-variant-that-simulates-a-risc-v-cpu-to-run-c-and-micropython-code","title":{"rendered":"Qwenputer \u2014 verzia AI-modelu Qwen3.5, schopn\u00e1 simulova\u0165 CPU RISC-V na vykon\u00e1vanie k\u00f3du v C a MicroPython","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
V r\u00e1mci projektu Qwenputer v\u00fdskumn\u00edci pripravili roz\u0161\u00edren\u00fa verziu AI-modelu Qwen3.5-4B, ktor\u00fd obsahuje integrovan\u00fd simul\u00e1tor plnohodnotn\u00e9ho 32-bitov\u00e9ho CPU na b\u00e1ze architekt\u00fary RISC-V (RV32IM), podporuj\u00faci roz\u0161\u00edrenie Zba a vybaven\u00fd koprocesorom bignum na aritmetick\u00e9 oper\u00e1cie s ve\u013ek\u00fdmi \u010d\u00edslami. Implement\u00e1cia jadra CPU bola skompilovan\u00e1 do s\u00faboru v\u00e1h a integrovan\u00e1 do v\u00e4\u010d\u0161ieho jazykov\u00e9ho modelu formou 11 dodato\u010dn\u00fdch vrstiev transform\u00e1torov, pridan\u00fdch pred z\u00e1kladn\u00e9 vrstvy Qwen. V\u00fdpo\u010dty sa vykon\u00e1vaj\u00fa priamo po\u010das procesu priameho prechodu (forward pass) neur\u00f3novej siete, bez pripojenia extern\u00fdch n\u00e1strojov alebo pluginov.<\/p>\n
Implement\u00e1cia \u00faspe\u0161ne prech\u00e1dza v\u0161etk\u00fdmi testami z ofici\u00e1lneho s\u00faboru testov riscv-tests a je schopn\u00e1 vykon\u00e1va\u0165 skompilovan\u00fd k\u00f3d v C a MicroPython. Pri vykon\u00e1van\u00ed aritmetick\u00fdch v\u00fdpo\u010dtov s ve\u013ek\u00fdmi \u010d\u00edslami Qwenputer zabezpe\u010duje presnos\u0165 97,5-100%, \u010do je porovnate\u013en\u00e9 alebo prevy\u0161uje v\u00fdsledky z\u00e1kladn\u00e9ho modelu Qwen3.5 s pripojen\u00fdm extern\u00fdm interpretr\u00e1torom Python. <\/p>\n
<\/center><\/p>\n
Na simul\u00e1ciu opera\u010dnej pam\u00e4te v modeli s\u00fa vyu\u017eit\u00e9 samostatn\u00e9 vrstvy Gated DeltaNet (line\u00e1rne RNN), zamestn\u00e1van\u00e9 v \u0161tandardnom AI-modeli Qwen na uchov\u00e1vanie zmenen\u00e9ho stavu pri pr\u00e1ci s ve\u013ek\u00fdmi kontextov\u00fdmi oknami. Pevn\u00e9 ulo\u017eenie stavu br\u00e1ny a aplik\u00e1cia unitarne k\u00f3dovania (one-hot) umo\u017e\u0148uje, aby ka\u017ed\u00fd z\u00e1pisov\u00fd po\u017eiadavok prep\u00edsal presne jeden riadok matice stavu, \u010do umo\u017e\u0148uje vyu\u017e\u00edva\u0165 stav DeltaNet ako adresovate\u013en\u00fa pam\u00e4\u0165 pevnej ve\u013ekosti, pri\u010dom sa vyh\u00fdba line\u00e1rnemu rastu KV-cache (Key-Value) pri zv\u00e4\u010d\u0161ovan\u00ed d\u013a\u017eky kontextu. V podobnej pam\u00e4ti s\u00fa umiestnen\u00e9 registre CPU, d\u00e1ta a vykon\u00e1van\u00fd strojov\u00fd k\u00f3d. <\/p>\n
<\/center><\/p>\n
In\u0161trukcie procesora, ako \u010d\u00edtanie registrov, aritmetick\u00e9 oper\u00e1cie a podmienen\u00e9 skoky, sa kompiluj\u00fa priamo do v\u00e1h vrstiev transform\u00e1tora. S ka\u017edou in\u0161trukciou s\u00fa sp\u00e1rovan\u00e9 \u0161peci\u00e1lne tokeny (ADDR, VAL, STEP), ktor\u00e9 vrstvy AI-modelu interpretuj\u00fa ako logiku fungovania CPU. Na inicializ\u00e1ciu CPU a na\u010d\u00edtanie firmv\u00e9ru sa pou\u017e\u00edva mal\u00fd extern\u00fd ovl\u00e1da\u010d RvChat, ale samotn\u00fd strojov\u00fd k\u00f3d sa spracov\u00e1va \u00faplne v r\u00e1mci v\u00e1h modelu.<\/p>\n
Medzi obmedzeniami projektu sa spom\u00edna pribli\u017ene dvojn\u00e1sobn\u00e9 zv\u00fd\u0161enie ve\u013ekosti (18,8 GB) v porovnan\u00ed s p\u00f4vodn\u00fdm modelom Qwen pri zachovan\u00ed sn\u00edmok stavu (checkpoint), \u010do sa v\u00e4\u010d\u0161inou pripisuje dodato\u010dn\u00e9mu zaplneniu \u010fal\u0161\u00edch vrstiev. Takisto sa uv\u00e1dza relat\u00edvne n\u00edzka r\u00fdchlos\u0165 v\u00fdpo\u010dtov (okolo 3200 tokenov za sekundu na GPU L4), \u010do m\u00f4\u017ee sp\u00f4sobi\u0165, \u017ee zlo\u017eit\u00e9 v\u00fdpo\u010dty trvaj\u00fa sekundy, nie milisekundy. <\/p>\n
\u010eal\u0161ie varianty AI-modelu implementovali s experiment\u00e1lnym s\u00faborov\u00fdm syst\u00e9mom, ktor\u00fd umo\u017e\u0148uje uchov\u00e1va\u0165 vo v\u00e1hach AI-modelu a\u017e 512 s\u00faborov s 53 000 tokenmi textu. Rovnako ako pri realiz\u00e1cii opera\u010dnej pam\u00e4te, je ulo\u017eenie usporiadan\u00e9 v s\u00favislosti so stavmi DeltaNet, ale odli\u0161uje sa t\u00fdm, \u017ee FS zabezpe\u010duje trval\u00e9 uchov\u00e1vanie, ktor\u00e9 nestr\u00e1ca stav medzi rel\u00e1ciami.<\/p>\n
Vyvinut\u00e9 technol\u00f3gie Qwenputer tie\u017e posl\u00fa\u017eili ako z\u00e1klad pre v\u00fdskum nov\u00fdch potenci\u00e1lnych probl\u00e9mov s bezpe\u010dnos\u0165ou v AI. Na pr\u00edklade Qwenputer sa ukazuje, ako m\u00f4\u017ee \u0161peci\u00e1lne modifikovan\u00fd AI-model skr\u00fdva\u0165 svoje \u00favahy v RNN stave a obch\u00e1dza\u0165 mechanizmy monitorovania \u00favah modelu (CoT, Chain-of-Thought). Taktie\u017e bola predveden\u00e1 \nmo\u017enos\u0165 vytv\u00e1rania mechanizmov kontroly a monitorovania bezpe\u010dnosti, zabudovan\u00fdch do v\u00e1h AI-modelu, ktor\u00e9 sama model nem\u00f4\u017ee vypn\u00fa\u0165 alebo zmeni\u0165. \n Zdroj: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"
![](\"\/wp-content\/uploads\/2026\/10\/63b0b8c6c9f2eb8c2e06773ad737106b.png\")
<\/center><\/p>\n![](\"\/wp-content\/uploads\/2026\/10\/5b25214f8fb3789baf16ad3661df0d31.png\")
<\/center><\/p>\n