# Umhlahlandlela Wezinsimbi Ze-AI Zendawo (2026)

Published: 2026-09-14

Uma bakha umshini wama-ejenti we-AI wendawo nama-LLM anesisindo esivulekile, abathuthukisi benza iphutha lokuthenga izidingo ze-PC yemidlalo: ama-CPU angu-5.8 GHz, ama-liquid cooling loops enziwe ngokwezifiso, nama-GPU wemidlalo asheshayo ane-VRAM engu-8GB kuphela. Kodwa ukukhiqizwa kwamathokheni okuzenzakalelayo kukhawulelwe umkhawulo wokudluliswa kwedatha, hhayi ukubalwa: ukuze kukhiqizwe ithokheni elilodwa, isisindo ngasinye kumodeli kufanele siwele ibhasi lememori. Uma izisindo zakho noma umongo we-KV cache udlula i-VRAM yangempela, isikhathi sokusebenza sidlulisa izingqimba ku-system DDR5 phezu kwe-PCIe, futhi ushaya umkhawulo wokudluliswa kwedatha ophindwe ka-20: isivinini siwe ngokuphuma kumathokheni angu-40 ngesekhondi kuya ku-1.5. Kulesi sivivinyo senkundla, uNiko uhlaziya imishini yezinsimbi, imithetho yokulinganisa imodeli ye-4-bit quantization, amazinga amathathu angempela esabelomali kusuka ku-$1,200 kuya ku-$5,000, ukuthi kungani i-RTX 3090 24GB esetshenzisiwe iyona ntengo engcono kakhulu ye-VRAM ngedola ekubaleni, ugibe lwe-Raspberry Pi, kanye nesu le-home gym lokusebenzisa okwakho ngokumelene ne-cloud inference. Isinqumo: SHIP IT.

Canonical: https://thedailydiff.dev/zu/video/2026-09-14-local-ai-hardware/

## Lokho okuhlanganiswa yile vidiyo

- Umkhawulo wokudluliswa kwedatha ophindwe ka-20: 936 GB/s GDDR6X vs 50 GB/s DDR5 & 31.5 GB/s PCIe
- Usayizi wemodeli @ 4-bit: 8B (5GB), 14B (10GB), 32B (20GB), 70B (40GB)
- Izinga 1 ($1,200–$1,500): RTX 4060 Ti 16GB (ungathengi 8GB) noma Mac Mini 16GB
- Izinga 2 ($1,500–$2,000): Okuhle kakhulu i-RTX 3090 24GB esetshenzisiwe noma Mac Mini M4 Pro 64GB
- Izinga 3 ($3,500+): RTX 4090 24GB / ama-3090 amabili noma Mac Studio Ultra

## Umbhalo obhaliwe ohunyushwe

Kuhunyushwe kusuka ekulandiseni kokuqala kwesiNgisi. Umsindo namazwibela atholakalayo alawulwa yi-YouTube.

0:00 Uma uhlela ukwakha i-PC yokusebenzisa ama-ejenti we-AI wendawo, yeka ukwakha i-rig yemidlalo. Awudingi i-Core i9 engu-5.8 gigahertz, i-liquid cooling loop, noma ikhadi lezithombe elingu-8 gigabyte elimbozwe i-RGB. Ekusebenziseni i-AI yendawo, izidingo zemidlalo cishe azilutho. Inani elilodwa kuphela elinqumayo ukuthi i-ejenti yakho iyasebenza noma iyancipha i-VRAM capacity kanye nebhasi lememori eline-bandwidth. Imithetho yokuhlolwa kwezinsimbi: khohlwa nge-CPU clocks namabhentshimakhi we-rasterization.

0:24 Sikhathazeka ngamanani amathathu: ububanzi bebhasi lememori, i-bandwidth ngama-gigabytes nge sekondi, kanye ne-VRAM headroom eluhlaza ye-KV cache bloat. Kulesi sivivinyo senkundla, ngizohlaziya umkhawulo we-bandwidth ophindwe ka-20, ngichaze imithetho yokulinganisa imodeli, ngibhentshimakhi amazinga amathathu angempela kusuka ku-$1,200 kuya ku-$5,000, futhi ngichaze ukuthi kungani i-3090 esetshenzisiwe iseyikhodi yokukhwabanisa enkulu kunazo zonke ekubaleni komhlaba. Lokhu yi-The Daily Diff, isivivinyo senkundla. Ukuze uqonde ukuthi kungani ama-rigs wemidlalo ehluleka, bheka ukuthi ukukhiqizwa kwamathokheni kusebenza kanjani ngempela. Emdlalweni, i-CPU yakho inikeza ama-draw calls kanti i-GPU yakho ikhiqiza amafreyimu.

0:54 Kodwa i-LLM decoding yokuzenzakalelayo cishe i-bandwidth yememori kuphela ekhawulelwe. Ukuze kukhiqizwe ithokheni elilodwa, i-GPU kufanele idlulise yonke ipharamitha yesisindo se modweli kusuka kumemori ngezinhloko zayo zokubala. Uma imodeli yakho ingu-10 gigabytes, ukukhiqiza ithokheni elilodwa kusho ukufunda ama-gigabytes angu-10 edatha. Ku-Nvidia RTX 3090 enebhasi lememori elingu-384-bit, uthola ama-gigabytes angu-936 ngesekhondi we-GDDR6X bandwidth, kukhiqiza amathokheni angu-80 ngesekhondi. Kodwa bheka okwenzekayo ngesekhondi imodeli yakho idlula i-VRAM yangempela.

1:22 Uma i-VRAM igcwele, isikhathi sokusebenza sidlulisa izingqimba ezisele phezu kwebhasi le-PCIe kuya kumemori ye-DDR5 yesistimu. Izinhloko ze-GPU yakho zilindele ama-gigabytes ayi-1,000 ngesekhondi. Esikhundleni salokho, i-dual-channel DDR5 ibapha ama-50, kanti i-PCIe 4 icishe kuma-31. Lokho ukuwohloka kwe-bandwidth okuphindwe ka-20. Ipayipi lokukhiqiza amathokheni limisa masinyane lilindele ibhasi, futhi isivinini siwe ngokuphuma kumathokheni angu-40 ngesekhondi kuya ku-1.5. Uphendule i-rig eyi-$2,000 yaba isithando somoya.

1:47 Futhi kuba kubi nakakhulu ngesikhathi sokusebenza kwe-ejenti okuningi, ngoba izisindo zinguhhafu kuphela wempi. Yonke i-prompt, ukubiza ithuluzi, kanye ne-file chunk kugcinwa ku-Key-Value cache. Ifasitela lomongo elingu-32K lingahlafuna ama-gigabytes angu-4 kuya kwangu-8 we-VRAM ngaphezu kwezisindo. Uma ikhadi lakho linama-gigabytes angu-16 kuphela, i-ejenti yakho ijikeleza kabili, ishaya udonga lomongo, bese iyashayisa ngephutha lokuphela kwememori noma ichitheka ku-DDR5. Nansi i-cheat sheet yokulinganisa ye-4-bit. Imodeli engu-7 noma engu-8 billion parameter njenge-Llama 3.1 noma i-DeepSeek Distill

2:16 ithatha cishe ama-gigabytes angu-5. Imodeli engu-14 billion ithatha ama-gigabytes angu-10. Imodeli engu-32 billion, indawo enhle yobuhlakani bama-ejenti okubhala amakhodi, idinga ama-gigabytes angu-20. Futhi imodeli ye-frontier engu-70 billion idinga ama-gigabytes angu-40 ngaphambi kokuthi abelule umongo. Okusiletha ekwakhiweni kwezinsimbi zangempela. Izinga 1 yi-starter rig, $1,200 kuya ku-$1,500. Ku-PC, i-anchor yakho yi-RTX 4060 Ti 16-gigabyte.

2:39 Isixwayiso esibucayi: ungalokothi uthenge inguqulo eyi-8-gigabyte ukuze wonge ama-$70. Ama-gigabytes angu-8 e-VRAM ngo-2026 yisigwebo sokufa sokuphela kwememori ngokushesha kunoma yikuphi ukusebenza kwe-ejenti yangempela. Ihlanganise ne-Ryzen 5 enezinhloko eziyisithupha, ama-gigabytes angu-64 we-DDR5, kanye ne-NVMe drive engu-2-terabyte. Ezweni le-Apple, okufanayo yi-Mac Mini noma i-MacBook Pro enama-gigabytes angu-16 wememori ehlanganisiwe.

3:02 Uzobona amathokheni angu-40 kuya kwangu-50 ngesekhondi kumamodeli angu-8 billion. Izinga 2 yindawo enhle yabasebenzisi abaphambili: ukwakha ngqo ukusebenzisa amamodeli angu-32 billion parameter njenge-Qwen 2.5 32B. Indlela A yi-RTX 4070 Ti Super entsha enama-gigabytes angu-16 ngama-$800. Kodwa Indlela B yisiphakamiso sami esiqinile: thenga i-Nvidia RTX 3090 esetshenzisiwe engama-gigabytes angu-24. Ungathola ama-3090 ahlanzekile ku-eBay ngama-$650 kuya ku-$750. Lokho kukunika ama-gigabytes angu-24 e-VRAM kubhasi elikhulu elingu-384-bit eliphusha

3:33 ama-gigabytes angu-936 ngesekhondi. Ngokuphathelene nedola, yintengo engcono kakhulu ye-VRAM ekubaleni. Ku-macOS, okufanayo nezinga 2 yi-Mac Mini M4 Pro enama-gigabytes angu-64 wememori ehlanganisiwe. Ikhiqiza amathokheni angu-11 kuya kwangu-12 ngesekhondi kumodeli engu-32 billion: kancane kune-Nvidia, kodwa ithule ngokuphelele kuma-watts angu-30 nendawo yefasitela lomongo elikhulu. Izinga 3 yibakaki labathanda okuningi lama-multi-agent swarms. Ku-PC, lokho kusho i-RTX 4090 enama-gigabytes angu-24,

3:57 i-Ryzen 9, nama-gigabytes angu-128 e-RAM, noma ama-RTX 3090 amabili anikeza ama-gigabytes angu-48 e-VRAM isiyonke. Ekuqaleni kwe-Apple, lokhu yi-Mac Studio Ultra enama-gigabytes angu-96 kuya kwangu-128 wememori ehlanganisiwe. Amandla amakhulu yi-memory residency: ungagcina imodeli ye-embedding, i-router esheshayo, kanye nemodeli yokubhala amakhodi engu-32 billion elayishiwe ngesikhathi esisodwa ngokulibaziseka okungashintshi. Manje ukwehluleka okuqotho kokuhlolwa kwethu kwenkundla: ugibe lwe-edge computing.

4:24 Njalo ngeviki okuthunyelwe komphakathi kuqamba ukuthi ungasebenzisa ama-ejenti okubhala amakhodi azimele ku Raspberry Pi 5 engama-$80. Ngikuhloliwe. Ukusebenzisa imodeli encane engu-1.5 billion parameter kukunika amathokheni ambalwa nje amathathu ngesekhondi ngenkathi ibhodi lishaya kuma-degrees angu-85 Celsius. Kuyithoyizi lempelasonto elihle, kodwa njengomshayeli wentuthuko wansuku zonke, kuyisijeziso esimsulwa. Esoftware, sebenzisa i-Ollama uma ufuna i-daemon elula yomugqa womyalo exhuma

4:47 ngqo ku-Cursor, Cline, noma i-VS Code. Uma ufuna indawo yokudlala enezithombe enokulanda kwamamodeli kanye nezilayidi zezendlalelo ze-GPU, sebenzisa i-LM Studio. Futhi qondanisa ifomethi yakho ne-silicon yakho. Ku-Apple Silicon, hlala ulanda amamodeli e-GGUF alungiselelwe i-Metal. Ku-Windows noma i-Linux nge-Nvidia, landa amamodeli e-AWQ noma e-EXL2, asebenzisa ama-Nvidia Tensor Cores ukuze asebenze ngokushesha ngama-20 kuya kuma-30%. Manje ukuthi ungakusebenzisa kanjani lokhu ngaphandle kokuhluleka?

5:11 Cabanga ngezinsimbi zendawo njenge-home gym vs. i-commercial gym. Ukuba ne-squat rack ekhaya kusho ukuthi uqeqesha nsuku zonke ngaphandle kokuhamba, izinkokhelo zokubhalisa ezingekho, kanye nobumfihlo obuphelele. Umshini wakho wendawo uphatha amaphesenti angu-80 wokubhala amakhodi kwansuku zonke, ukuhlolwa kweyunithi, kanye nokucubungula imibhalo yangasese ngamadola ayiziro ngethokheni ngalinye. Futhi lapho udinga ukudonsa ama-pounds angu-500 — njenge-repo-wide enkulu ukubuyekeza kwezakhiwo ezifayeleni ezingama-50 — uvakashela i-commercial gym: khokha i-cloud API ye-Claude 3.5 Sonnet noma i-OpenAI o3 imizuzu eyi-15

5:38 bese uqhubeka. Nansi i-bill: ukwakhiwa kwezinga 2 nge-3090 esetshenzisiwe kubiza ama-$1,600 konke kuhlanganisiwe. Uma uchitha ama-$50 kuya kuma-$100 ngenyanga kumathokheni e-API, kuzikhokhela yona ngezinyanga eziyi-18 ngenkathi igcina ikhodi yakho yobunikazi ingaphandle kwamaseva enkampani yangaphandle. Isinqumo sokuhlolwa kwenkundla yanamuhla: SHIP IT. I-VRAM kanye ne-memory bandwidth zishaya i-clock speeds ngaso sonke isikhathi. Yeka ukuthenga ama-CPU angu-5-gigahertz we-AI, bese uthola i-3090 esetshenzisiwe.

6:04 Ngitshele ukuthi yikuphi ukwakhiwa kwezinsimbi okusebenzisayo kumamodeli endawo kumazwana. Futhi uma ungafisa ukufunda lokhu kuhlukaniswa, thola i-newsletter ku-the daily diff dot dev, isixhumanisi esingezansi. Futhi lokho yi-diff yanamuhla. Ngingu-Niko wase-Axrisi. Hlanganisa ngokuzibophezela.

## Imithombo

- [Niko from Axrisi: Local AI Hardware — Stop Building a Gaming PC](https://axrisi.com/) — axrisi.com
- [NVIDIA RTX 3090 Specifications (384-bit bus, 936 GB/s GDDR6X)](https://www.nvidia.com/) — www.nvidia.com
- [llama.cpp Memory Bandwidth & Offloading Architecture](https://github.com/ggerganov/llama.cpp) — github.com
- [Ollama Model Library & Benchmarks](https://ollama.com/) — ollama.com
- [vLLM PagedAttention & KV Cache Memory Management](https://github.com/vllm-project/vllm) — github.com
- [JEDEC DDR5 Memory Standard Specifications](https://www.jedec.org/) — www.jedec.org