GPU Cloud
GPU cloud: kodesha seriveri za NVIDIA GPU ku isaha
Kodesha seriveri za NVIDIA GPU zo gukoresha LLM inference, fine-tuning no guhanga amashusho. Ibiciro biriho ako kanya, kwishyura ku isaha, template za PyTorch, vLLM na ComfyUI, no kwinjiriramo ukoresheje SSH.
- Ibiboneka ako kanya, igiciro kikemezwa mbere y'uko utumiza
- Template ziteguye neza za PyTorch, vLLM, Ollama na ComfyUI
- Seriveri zishyurwa ku isaha zahagaritswe zishyurirwa gusa disiki
Hamwe na Cloudj GPU Cloud ukodesha seriveri ya GPU mu bihari ako kanya: hitamo ubwoko bwa GPU, VRAM na disiki, tangirira ku rufatiro (template) nka PyTorch, vLLM, Ollama cyangwa ComfyUI, hanyuma wihuze ukoresheje SSH. Seriveri z'isaha zishyurwa mu mafaranga ari kuri konti yawe, kandi seriveri y'isaha yahagaritswe yishyurirwa gusa disiki yayo.
Ibiciro byo gukodesha GPU n'iziri kuboneka aka kanya
Ibihari ako kanya: ibiboneka n'ibiciro bihinduka bitewe n'isoko. Ahagenzurirwa seriveri hemeza igiciro cya nyuma mbere yo gutumiza.
| GPU | VRAM kuri buri GPU | Ubwubatsi | Imibare ya FP32 | Ingufu ntarengwa | Kuva kuri | |
|---|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Max-Q | 96 GB | Blackwell | ≈ 119 TFLOPS | 300 W | – | Ibisobanuro → |
| NVIDIA RTX PRO 6000 WS | 96 GB | Blackwell | ≈ 119 TFLOPS | 600 W | – | Ibisobanuro → |
| NVIDIA RTX 5090 | 32 GB | Blackwell | ≈ 107.6 TFLOPS | 575 W | – | Ibisobanuro → |
| NVIDIA RTX 5090D | 32 GB | Blackwell | ≈ 107.6 TFLOPS | 575 W | – | Ibisobanuro → |
| NVIDIA RTX PRO 6000 S | 96 GB | Blackwell | ≈ 93.6 TFLOPS | 600 W | – | Ibisobanuro → |
| NVIDIA RTX 4090 | 24 GB | Ada Lovelace | ≈ 81.4 TFLOPS | 450 W | – | Ibisobanuro → |
| NVIDIA RTX 6000Ada | 48 GB | Ada Lovelace | ≈ 81.4 TFLOPS | 300 W | – | Ibisobanuro → |
| NVIDIA RTX 6000D | 84 GB | Blackwell | ≈ 77.6 TFLOPS | 600 W | – | Ibisobanuro → |
| NVIDIA L40S datacenter | 45 GB | Ada Lovelace | ≈ 73.3 TFLOPS | 350 W | – | Ibisobanuro → |
| NVIDIA RTX PRO 5000 | 48 GB | Blackwell | ≈ 69.6 TFLOPS | 300 W | – | Ibisobanuro → |
| NVIDIA H100 SXM datacenter | 80 GB | Hopper | ≈ 53.5 TFLOPS | 700 W | – | Ibisobanuro → |
| NVIDIA RTX 5080 | 16 GB | Blackwell | ≈ 53.4 TFLOPS | 360 W | – | Ibisobanuro → |
| NVIDIA RTX 4080S | 16 GB | Ada Lovelace | ≈ 50.9 TFLOPS | 320 W | – | Ibisobanuro → |
| NVIDIA RTX 4080 | 16 GB | Ada Lovelace | ≈ 48.6 TFLOPS | 285 W | – | Ibisobanuro → |
| NVIDIA H100 NVL datacenter | 94 GB | Hopper | ≈ 48.3 TFLOPS | 400 W | – | Ibisobanuro → |
| NVIDIA H200 NVL datacenter | 140 GB | Hopper | ≈ 48.3 TFLOPS | 600 W | – | Ibisobanuro → |
| NVIDIA RTX 5070 Ti | 16 GB | Blackwell | ≈ 44.3 TFLOPS | 300 W | – | Ibisobanuro → |
| NVIDIA RTX PRO 4000 | 24 GB | Blackwell | ≈ 44.3 TFLOPS | 145 W | – | Ibisobanuro → |
| NVIDIA RTX 4070S Ti | 16 GB | Ada Lovelace | ≈ 42.6 TFLOPS | 285 W | – | Ibisobanuro → |
| NVIDIA H100 PCIE datacenter | 80 GB | Hopper | ≈ 41 TFLOPS | 350 W | – | Ibisobanuro → |
| NVIDIA RTX 4500Ada | 24 GB | Ada Lovelace | ≈ 38.2 TFLOPS | 210 W | – | Ibisobanuro → |
| NVIDIA RTX 3090 Ti | 24 GB | Ampere | ≈ 36.4 TFLOPS | 450 W | – | Ibisobanuro → |
| NVIDIA RTX A6000 | 48 GB | Ampere | ≈ 36.1 TFLOPS | 300 W | – | Ibisobanuro → |
| NVIDIA RTX 4070S | 12 GB | Ada Lovelace | ≈ 35.8 TFLOPS | 220 W | – | Ibisobanuro → |
| NVIDIA RTX 3090 | 24 GB | Ampere | ≈ 35.4 TFLOPS | 350 W | – | Ibisobanuro → |
| NVIDIA RTX 3080 Ti | 12 GB | Ampere | ≈ 35.1 TFLOPS | 350 W | – | Ibisobanuro → |
| NVIDIA A40 datacenter | 45 GB | Ampere | ≈ 29.9 TFLOPS | 300 W | – | Ibisobanuro → |
| NVIDIA RTX 3080 | 10 GB | Ampere | ≈ 29.5 TFLOPS | 380 W | – | Ibisobanuro → |
| NVIDIA RTX 4070 | 12 GB | Ada Lovelace | ≈ 29.3 TFLOPS | 200 W | – | Ibisobanuro → |
| NVIDIA L4 datacenter | 22 GB | Ada Lovelace | ≈ 24.2 TFLOPS | 72 W | – | Ibisobanuro → |
| NVIDIA RTX 5060 Ti | 16 GB | Blackwell | ≈ 22.8 TFLOPS | 180 W | – | Ibisobanuro → |
| NVIDIA RTX 4060 Ti | 16 GB | Ada Lovelace | ≈ 21.6 TFLOPS | 165 W | – | Ibisobanuro → |
| NVIDIA RTX A4000 | 16 GB | Ampere | ≈ 20.6 TFLOPS | 140 W | – | Ibisobanuro → |
| NVIDIA RTX 3070 | 8 GB | Ampere | ≈ 19.8 TFLOPS | 220 W | – | Ibisobanuro → |
| NVIDIA RTX 3060 Ti | 8 GB | Ampere | ≈ 15.7 TFLOPS | 200 W | – | Ibisobanuro → |
| NVIDIA A100 PCIE datacenter | 40 GB | Ampere | ≈ 15.6 TFLOPS | 250 W | – | Ibisobanuro → |
| NVIDIA A100 SXM4 datacenter | 40 GB | Ampere | ≈ 15.6 TFLOPS | 400 W | – | Ibisobanuro → |
| NVIDIA A800 PCIE datacenter | 80 GB | Ampere | ≈ 15.6 TFLOPS | 300 W | – | Ibisobanuro → |
| NVIDIA RTX 2080 Ti | 11 GB | Turing | ≈ 14.6 TFLOPS | 250 W | – | Ibisobanuro → |
| NVIDIA RTX 3060 laptop | 12 GB | Ampere | ≈ 12.9 TFLOPS | 80 W | – | Ibisobanuro → |
| NVIDIA Tesla V100 datacenter | 16 GB | Volta | ≈ 12.5 TFLOPS | 300 W | – | Ibisobanuro → |
| NVIDIA CMP 170HX datacenter | 64 GB | Ampere | ≈ 12.1 TFLOPS | 250 W | – | Ibisobanuro → |
| NVIDIA RTX 3060 | 12 GB | Ampere | ≈ 12 TFLOPS | 170 W | – | Ibisobanuro → |
| NVIDIA Titan Xp | 12 GB | Pascal | ≈ 11.7 TFLOPS | 160 W | – | Ibisobanuro → |
| NVIDIA RTX A2000 | 6 GB | Ampere | ≈ 11.2 TFLOPS | 70 W | – | Ibisobanuro → |
| NVIDIA GTX 1080 Ti | 11 GB | Pascal | ≈ 11 TFLOPS | 250 W | – | Ibisobanuro → |
| NVIDIA GTX 1080 | 8 GB | Pascal | ≈ 7.8 TFLOPS | 130 W | – | Ibisobanuro → |
| NVIDIA GTX 1070 Ti | 8 GB | Pascal | ≈ 7.4 TFLOPS | 180 W | – | Ibisobanuro → |
| NVIDIA GTX 1650 | 4 GB | ≈ undefined TFLOPS | W | – | Ibisobanuro → | |
| NVIDIA Q RTX 8000 | 45 GB | Turing | ≈ 11.9 TFLOPS | 250 W | – | Ibisobanuro → |
| NVIDIA Tesla P40 datacenter | 24 GB | Pascal | ≈ 9.4 TFLOPS | 250 W | – | Ibisobanuro → |
| NVIDIA RTX 2070S | 8 GB | Turing | ≈ 8.8 TFLOPS | 210 W | – | Ibisobanuro → |
| NVIDIA Q RTX 4000 | 8 GB | Turing | ≈ 7.7 TFLOPS | 125 W | – | Ibisobanuro → |
| NVIDIA RTX 2070 | 8 GB | Turing | ≈ 7.7 TFLOPS | 150 W | – | Ibisobanuro → |
| NVIDIA RTX 2060S | 8 GB | Turing | ≈ 7.5 TFLOPS | 175 W | – | Ibisobanuro → |
| NVIDIA GTX TITAN X | 12 GB | Maxwell | ≈ 6.8 TFLOPS | 150 W | – | Ibisobanuro → |
| NVIDIA RTX 2060 | 12 GB | Turing | ≈ 6.5 TFLOPS | 130 W | – | Ibisobanuro → |
| NVIDIA GTX 1660 Ti | 6 GB | Turing | ≈ 5.3 TFLOPS | 120 W | – | Ibisobanuro → |
| NVIDIA GTX 1660 S | 6 GB | Turing | ≈ 3.6 TFLOPS | 125 W | – | Ibisobanuro → |
Template za GPU zo gukora imirimo ya AI, LLM no gutunganya amashusho
Gukoresha LLM (LLM inference)
vLLM, SGLang cyangwa Ollama hamwe na Open WebUI mu gutanga no gupima moderi z'indimi.
Gutoza no kunoza moderi (Fine-tuning)
Unsloth Studio yo kunoza moderi z'indimi ukoresheje amakuru yawe bwite.
Gukora amashusho na videwo
ComfyUI ya Flux, SDXL hamwe n'inzira z'imikoreshereze ya videwo.
Iterambere n'ubwubatsi muri PyTorch na CUDA
Amashusho ya PyTorch na Ubuntu 24.04 arimo NVIDIA CUDA, agerwaho binyuze kuri SSH.
Desktop cyangwa VM yuzuye
Desktop ya Ubuntu muri mudasobwa y'ububiko bwa interineti (browser), cyangwa mashini nyayo ya KVM igihe cyose ubikeneye.
Uko wakodesha seriveri ya GPU mu ntambwe 4
- 01
Hitamo urufatiro (template) n'ingano ya disiki
- 02
Hitamo GPU ibereye ibyo ukeneye mu bihari ako kanya
- 03
Emeza igiciro hanyuma ukore seriveri ukoresheje amafaranga ari kuri konti yawe
- 04
Ihuze ukoresheje SSH cyangwa iterminali yo muri mushakisha
Ibiciro n'impushya byo kwishyura GPU cloud
- Kwishyura ku isaha: GPU yishyurwa kuri buri saha seriveri iri gukora. Seriveri y'isaha yahagaritswe yishyurirwa gusa disiki yayo.
- Ibyiciro byo kwishyura: buri kwezi, buri mezi atatu, buri mezi atandatu cyangwa buri mwaka, hamwe n'igabanuka ryose ryerekanwa mbere yo kwishyura.
- Ububiko bwa disiki bwishyurwa kuri buri GB hamwe n'igiciro cya GPU.
- Ibiciro biri muri USD; ushobora kubyerekana muri VND hakurikijwe igipimo cy'ivunjisha cy'ubu.
Ibibazo bikunze kubazwa kuri GPU cloud
Imirimo ikwiriye mu bubiko bwa GPU: LLM inference na fine-tuning, guhanga amashusho na videwo, gutunganya amashusho no gukora porogaramu za CUDA. Template za PyTorch, vLLM, SGLang, Ollama, Unsloth na ComfyUI zikwinjiriramo porogaramu zose mu buryo bwikora.
Imbonerahamwe iri hejuru yerekana imiterere iri kuboneka aka kanya, harimo na NVIDIA RTX 5090 ifite 32 GB za VRAM mu gihe ibonetse mu bubiko. Ibiri mu bubiko bihinduka ku munsi hose.
Ibiciro bikurwa ku bintu biri kuboneka muri GPU aka kanya kandi bishobora guhinduka. Console yerekana igiciro cyemejwe mbere y'uko utumiza, kandi seriveri yawe yishyuzwa kuri icyo giciro cyemejwe.
Yego. Uretse kwishyura ku isaha ushobora guhitamo kwishyura ku kwezi, ku mezi atatu, ku mezi atandatu cyangwa ku mwaka, kandi igabanuka ryose ryerekanwa mbere yo kwishyura. Umwanya wa disiki wishyuzwa kuri buri GB hamwe n'igiciro cya GPU.
Kuri gahunda yo kwishyura ku isaha, seriveri yahagaritswe yishyurirwa gusa umwanya wa disiki yayo. Siba seriveri mu gihe utagikeneye amakuru yayo.
Ukoresheje SSH hamwe n'urufunguzo rwawe, cyangwa ukoresheje terminal iri muri mushakisha yawe ukanze muri console. Template za Desktop na VM zitanga n'uburyo bwo gukoresha amashusho (graphical access).