Nvidia
 | |
Talɣut ɣef termist  | |
| Isem unṣib | Nvidia Corp. |
|---|---|
| Anaw n teṛmist | Titiknulujit |
| Azemz n usbeddi | 5 yebrir, 1993 |
| Amesbedday | Jensen Huang
Chris Malachowsky Curtis Priem |
| Tasɣimt tagejdant | Santa Klara, Kalifurnya, Marikan |
| Tamurt | Iwunak Yedduklen n Marikan |
| Taɣult n urmud | GPU |
| Ifarisen | Tikarḍiwin n GPU(GeForce • RTX / Quadro)
CUDA CPUs (Grace) DPUs (BlueField) SoCs (Tegra • Drive) DGX |
| Amaduf | 215 938 000 000 Adular amarikani (2025) ↑ +65,5 % (2024 → 2025) |
| Afaṛes n urmud | 130 387 000 000 Adular amarikani (2025) ↑ +60,1 % (2024 → 2025) |
| Tigenfit n taggara | 120 067 000 000 Adular amarikani (2025) ↑ +64,7 % (2024 → 2025) |
| Asmel unṣib | https://www.nvidia.com/ |
Nvidia d tarmiṣt tamirikanit n titiknulujit, tettnadi, tessalay yerna tessuffuɣ-d tiɣeṛɣaṛ n yisekkaren udlifen (GPU), inagrawen n usiḍen arurad, iseɣẓanen d wallalen n tigzi taragmawant.
Tettwasbedd deg 5 yebrir 1993 deg Kalifurnya, deg Marikan. Deg tazwara, iswi-is yella deg usnerni n wudlifen s kraḍ n tsekkiwin (3D), ladɣa i wuraren n yiselkimen. Maca, seg useggas n 2006, tebda tarmiṣt ad tesnerni tiknulujit i yessawaḍen GPU ad yexdem asiḍen ur cudden ara kan ɣer tugniwin.
Ass-a, NVIDIA ur teqqim ara d tarmiṣt n tikarḍiwin tudlifin kan. Tuɣal d yiwet seg tiṛmisin tigejdanin deg usali n inagrawen n tigzi taragmawant, aselmed n tmudmiwin n tutlayt, tusna, arrubutik, timacinin tifulmanin d yisekkaren imeqqranen.
Deg 2026, afaris-is ur yettwasres ara kan deg yiselkimen n yimdanen, maca daɣen deg yimmasen n yisefka, i deg xeddmen yiwalen n GPU s umata akken ad siḍnen isefka iṭṭuqqten s zreb.
3. Amezruy n NVIDIA
[ẓreg | ẓreg aɣbalu]3.1. Tallit n tlalit: 1993–1997
[ẓreg | ẓreg aɣbalu]Deg yebrir 1993, kraḍ n yijenyuṛen, Jensen Huang, Chris Malachowsky d Curtis Priem, sbeddan NVIDIA. Ɣur-sen, tikti tagejdant tella d asnerni n tiknulujit n tugniwin i yiselkimen.
Deg tallit-nni, uranen n yiselkimen bdant ad sɛun azal ameqran. Maca isekkaren n wakud-nni ur zmiren ara ad siḍnen s lemɣawla akk isefka ilaqen i usken n wudlifen 3D.
Deg 1995, NVIDIA tessufeɣ-d NV1, yiwen seg yifarisen-is imezwura. Ur yewwi ara lebɣi n tnezzut acku tasegda-s temgarad ɣef kra n yilugan i yettwasqedcen deg tallit-nni.
Deg 1997, tarmiṣt tessufeɣ-d RIVA 128, asekkar udlif i yesɛan amhaz meqqren deg usefrak n tugniwin 3D. Afaris-a yefka-as i NVIDIA tagnit ad tebnu azadur n tnezzut-is.
3.2. GeForce d tlalit n GPU: 1999
[ẓreg | ẓreg aɣbalu]Deg 1999, NVIDIA tessufeɣ-d GeForce 256, i tessekna am uGPU amezwaru.
GPU d agzul n Graphics Processing Unit, d asekkar yettwaselɣen i usiḍen n tugniwin d wudlifen. GeForce 256 yesdukkel deg yiwen n uceṛṛun kra n yimahilen udlifen i yellan uqbel xeddmen-ten yiferdisen yemgaraden.
Ilaq ad nefhem dakken NVIDIA d nettat i d-yessumren GPU s wudem-a n tnezzut deg 1999, maca isekkaren n tugniwin llan yakan uqbel-as.
S GeForce 256, NVIDIA tebda ad tettwasnen ugar deg tnezzut n uraren d yiselkimen n yimeẓla.
3.3. Tallit n CUDA: 2006
[ẓreg | ẓreg aɣbalu]Deg 2006, NVIDIA tessufeɣ-d CUDA (Compute Unified Device Architecture).
D yiwet seg tneqqiḍin tigejdanin deg umezruy-is.
Uqbel CUDA, GPU yettwasqedc s waṭas i tugniwin. S CUDA, imsihelen zemren ad arun ahilen i yessawaḍen GPU ad yexdem isḍanen n tusna, tusnakt, tusnengama d tasleṭ n yisefka.
S waya, GPU yuɣal d allal n usiḍen arurad, ur yeqqim ara kan d allal n uraren.
3.4. NVIDIA d tigzi taragmawant: 2012
[ẓreg | ẓreg aɣbalu]Deg 2012, tamudemt AlexNet, i d-yessuffɣen Alex Krizhevsky, Ilya Sutskever d Geoffrey Hinton, tewwi-d agmuḍ yelhan deg timsizzelt ImageNet.
Imnadiyen-a sqedcen GPU n NVIDIA i uselmed n uzeṭṭa n yiṛununen.
Agmuḍ-a yessekna dakken almad alqay (deep learning) yezmer ad yili d tarrayt tafessast ugar deg waṭas n tɣawsiwin n ussugen d tasleṭ n tugniwin.
Seg tallit-a, aseqdec n GPU deg tigzi taragmawant ibeddel amek xeddmen imnadiyen d yimsihelen.
3.5. Tallit n RTX: 2018
[ẓreg | ẓreg aɣbalu]Deg 2018, NVIDIA tessufeɣ-d tasegda RTX, i yessawaḍen ad yettwaxdem uskan n yizirigen n tafat (ray tracing) deg wakud ilaw.
Tatiknulujit-a tezmer ad tesɛeddi aselɣu n tafat, tilisa n tili d yimeclalen n tugniwin akken ad d-banent s wudem iqerrben ugar ɣer tilawt.
RTX tessnerna daɣen aseqdec n tigzi taragmawant i usnerni n tbadut d tɣara n tugniwin.
3.6. Seg 2022 ar 2026: tallit n inagrawen imeqqranen n AI
[ẓreg | ẓreg aɣbalu]Deg yiseggasen-a, NVIDIA teffeɣ seg wudem n tarmiṣt i yeznuzun GPU ɣer tarmiṣt i yessalayn anagraw ummid: isekkaren, takatut, azeṭṭa, iseɣẓanen, allalen n usihel d yimmasen n yisefka.
Tiɣeṛɣaṛ Hopper, Blackwell d Rubin ttwasnernant iwakken ad siḍnent timudmiwin timeqqranin n tigzi taragmawant.
Deg 2026, NVIDIA tebda asnerni n tɣessa Vera Rubin, i yesdukklen GPU Rubin, CPU Vera d yiferdisen n uzeṭṭa, akken ad yili usiḍen s waṭas n yisekkaren i d-yettmeslayen gar-asen.
Amezruy-a yeskan-d abeddel agejdan: seg usken n tugniwin 3D, ɣer tɣessa n tigzi taragmawant n umaḍal.
4. D acu i d GPU? Amek i iteddu?
[ẓreg | ẓreg aɣbalu]GPU (Graphics Processing Unit) d asekkar udlif, yettwaselɣen iwakken ad yexdem aṭas n usiḍen s yiwet n tikelt, s tarrayt n umahil amsadaɣ (parallel processing).
Tifrat-a temgarad ɣef CPU (Central Processing Unit), i yellan d asekkar agejdan n uselkim.
4.1. Amgired gar CPU d GPU
[ẓreg | ẓreg aɣbalu]
Unuɣ n usegzi, ur d-yettmeslay ara ɣef umḍan n yiferdisen n CPU neɣ GPU n tidet.
CPU yelha i yimahilen yemgaraden, i usendeh n unagraw, i usefrek n yisefka d tigawin i yesran ad myezwarent.
GPU, s tama-nniḍen, yelha ladɣa i yimahilen i nezmer ad nebḍu ɣef waṭas n yiferdisen ad ten-nsiḍen s wudem amsadaɣ.
Amedya: ad neḥseb timernit n sin yisurta, yal yiwen deg-sen yesɛa 1 000 000 n yimḍanen.
Imi yal timernit ur teḥwaǧ ara agmuḍ n tin nniḍen, GPU yezmer ad yebḍu amahil-a ɣef waṭas n yiḥricen. D tamiḍrant-a i d azadur n usiḍen amsadaɣ.
4.2. Iferdisen igejdanen n GPU
[ẓreg | ẓreg aɣbalu]GPU amiran yesɛa aṭas n yiferdisen, gar-asen:
- CUDA Cores: iferdisen n usiḍen i yimahilen n tusnakt d usiḍen amatu.
- Tensor Cores: iferdisen yettwaselɣen i yimahilen n yisiraw (matrices), i yesɛan azal meqqren deg almad alqay.
- RT Cores: iferdisen n uskan n yizirigen n tafat deg tugniwin 3D.
- VRAM: takatut n GPU, i deg ttwaḥrazen yisefka i yettwasiḍnen.
- Takatut L2: takatut taruradt i yellan qrib ɣer yigrawen n usiḍen.
- NVLink: tarrayt n tuqqna taruradt gar GPU d GPU, neɣ gar yiferdisen n unagraw, ilmend n tasegda.
Amedya, RTX 5090 yesɛa 32 GB n tkatut GDDR7, ma d isekkaren n yimmasen n yisefka zemren ad sɛun takatut HBM s wazal d tɣara yemgaraden.
5. CUDA: azadur n tzemmert n NVIDIA
[ẓreg | ẓreg aɣbalu]CUDA d yiwen n unagraw n usihel d tɣessa n usiḍen amsadaɣ i d-tesnulfa NVIDIA. Tessawaḍ i umsihel ad yesseqdec GPU i yimahilen ur nelli ara d tugniwin kan. Deg CUDA, imahilen ttwasuddsen s wudem-a:
Yal akala afessas(thread) yezmer ad yesɛu amḍiq-is, ad iɣer isefka, ad yexdem asiḍen, syin ad yerr agmuḍ.
D acu kan, ur ilaq ara ad neḥseb dakken GPU dima yif CPU deg yal amahil. Ma yella amahil yesɛa aṭas n tigawin i yemyuqqnen, neɣ yesra abrid amezwaru uqbel win wis sin, CPU yezmer ad yili yif-it.
5.1. Aseqdec n CUDA deg usihel
[ẓreg | ẓreg aɣbalu]CUDA yettwasqedc deg waṭas n tutlayin d yisnasen. Gar-asen C++, Python, PyTorch d TensorFlow.
Amedya amezwaru d agzul n wamek i yezmer umsihel ad yesseqdec GPU i tmernit n yisurta:
__global__ void add(const float *a,
const float *b,
float *c, int n) {
int i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < n) {
c[i] = a[i] + b[i];
}
}
Deg uhil-a, yal akala afessas(thread) yettawi yiwen n umḍiq i, syin yesiḍin a[i] + b[i]. S waya, aṭas n yiḥricen zemren ad xedmen deg wakud-nni kan. Amedya-a yeskan kan iɣes(kernel); akken ad yeddu, yesra asenker-is d usefrek n tkatut.
5.2. Ayɣer CUDA yesɛa azal meqqren?
[ẓreg | ẓreg aɣbalu]Azal n CUDA ur yelli ara kan deg tzemmert n GPU.
NVIDIA tebna-d fell-as aṭas n temkarḍiwin n usihel (software libraries), gar-asent cuDNN, cuBLAS, NCCL d TensorRT.
Asihel n waṭas n yisnasen n tigzi taragmawant teqqen ɣer tɣessa-a. Aya yessawaḍ-d i NVIDIA azal nniḍen deg tnezzut, acku amsihel ur yessutur ara GPU kan: yesra daɣen iseɣẓanen, allalen n usenqed d temkarḍiwin i yettwasenqden.
6. Tiɣeṛɣaṛ n NVIDIA: seg Tesla ɣer Vera Rubin
[ẓreg | ẓreg aɣbalu]NVIDIA tessnerna aṭas n tisegda n GPU deg umezruy-is.
| Tallit | Tasegda | Azal-is agejdan |
|---|---|---|
| 2006 | Tesla | Asnerni n usiḍen amsadaɣ |
| 2010 | Fermi | Asnerni n usiḍen ussnan |
| 2012 | Kepler | Tazmert akked tɣellist n useqdec asrurad |
| 2014 | Maxwell | Aseqdec yelhan n usrurad |
| 2016 | Pascal | Asnerni n tigzi taragmawant d GPU |
| 2017 | Volta | Asekcem n Tensor Cores |
| 2018 | Turing | RTX d ray tracing |
| 2020 | Ampere | A100 d RTX 30 |
| 2022 | Hopper | H100 d usiḍen n tmudmiwin timeqqranin |
| 2024 | Blackwell | B200, GB200 d RTX 50 |
| 2026 | Rubin / Vera Rubin | Inagrawen imeqqranen n AI |
6.1. Hopper d H100
[ẓreg | ẓreg aɣbalu]H100 d asekkar i d-tessufeɣ NVIDIA iwakken ad yexdem almad alqayan d usiḍen n tmudmiwin timeqqranin.
Yesɛa Tensor Cores n tziri tis kuẓ, takatut HBM d tigawin n usiḍen i yettwaselɣen i tigzi taragmawant.
H100 yesɛa azal deg umezruy n tmudmiwin n tutlayt timeqqranin (LLM), acku aṭas n yimmasen n yisefka sqedcen-t deg yiseggasen imezwura n tnerni n tigzi taragmawant turriwt.
6.2. Blackwell
[ẓreg | ẓreg aɣbalu]
Blackwell d tasegda i d-tessufeɣ NVIDIA deg 2024, i d-tefka azadur i yisekkaren B200, GB200 d yifarisen nniḍen.
Yiwen seg usnifel-is agejdan d asdukkel n sin yiḥricen n GPU deg yiwen n uceṛṛun ameqran, s tuqqna taruradt gar-asen.
Blackwell tesnerna daɣen aseqdec n yimḍanen s tiseddi taẓẓayant (low precision), iwakken ad yettwasiḍen ulmad n tmudmiwin d useqdec-nsent s tazzla d taɛkwamt n tazzla tameẓẓyant ugar.
Deg 2025 d 2026, NVIDIA tessnerna talɣa Blackwell Ultra, i yettwaselɣen ladɣa i usnerni n tzemmert n usiḍen n tigzi taragmawant.
6.3. Vera Rubin
[ẓreg | ẓreg aɣbalu]Vera Rubin d tasegda i d-yernan ɣef Blackwell. Ur d yiwen n GPU kan, maca d anagraw yesdukklen aṭas n yiferdisen:
- GPU Rubin: i usiḍen n tigzi taragmawant.
- CPU Vera: i usendeh d usefrek n yisefka.
- NVLink: i tuqqna taruradt gar GPU.
- ConnectX d BlueField: i usefrek n uzeṭṭa d yisefka.
- InfiniBand d Ethernet: i tuqqna gar yimmasen n yisekkaren d yimmasen n yisefka.
Anagraw Vera Rubin NVL72 yesdukkel 72 n GPU Rubin d 36 n CPU Vera deg yiwen n unagraw n yiṛak imeqqranen (rack-scale system).
Deg 2026, NVIDIA teqqar-d dakken tasegda-a tettwasnerna ladɣa i tmudmiwin n tigzi taragmawant i yesseqdecen aseɣẓen d yimeggan ifulmanen.