Kungani ukhetha i-H100 SXM kunokukhetha i-H100 PCIE?

I-H100 SXM ibalwe ku-53.5 TFLOPS FP32 kanye no-2869 GB/s uma kuqhathaniswa no-41 TFLOPS kanye no-1760 GB/s ekhadini le-PCIE, enomkhawulo ka-700 W esikhundleni sika-350 W. Emsebenzini ngamunye, ivamise ukuqeda ngokushesha.

Ingabe i-H100 SXM noma i-H200 NVL ingcono kuma-LLM?

I-H200 NVL ine-140 GB kanye no-3884 GB/s, ngakho-ke ihamba phambili kumamodeli amakhulu kakhulu kanye nezimo ezinde. I-H100 SXM ine-compute ye-FP32 eyengeziwe (53.5 vs 48.3 TFLOPS) futhi ifanele imisebenzi yokuqeqesha engena ku-80 GB.

Ngingakwazi yini ukushuna kahle imodeli ye-70B ku-H100 SXM eyodwa?

Nge-QLoRA, yebo: ama-weight e-4-bit 70B athatha cishe 35–40 GB, kusele indawo yama-adapter, ama-activation kanye nesimo se-optimizer. Ukushuna okugcwele kwe-16-bit kwe-70B kudinga ama-GPU amaningi.

Ngikhokha malini uma iseva ye-H100 SXM imisiwe ku-Cloudj?

Ekukhokhweni ngehora, iseva emisiwe ikhokhiswa kuphela idiski yayo. Sula iseva uma usukopishile ama-checkpoint akho futhi uzoyeka ukukhokha ngokuphelele.

GPU Cloud · NVIDIA · Hopper

Qasha i-NVIDIA H100 SXM GPU ene-80 GB VRAM

Qasha amaseva e-cloud GPU e-H100 SXM ngehora ku-Cloudj: 80 GB VRAM, NVIDIA Hopper, cishe 53.5 TFLOPS FP32 kanye ne-bandwidth yememori engu-2869 GB/s. Ngomkhawulo wamandla ongu-700 W, lena yi-H100 yamaqembu aqeqesha, ashuna kahle (fine-tune) futhi aphaka amamodeli amakhulu adinga imojuli ye-Hopper eshesha kakhulu ohlwini lwe-compute. Qalisa ithempulethi ye-PyTorch, vLLM noma Unsloth bese uxhuma nge-SSH ngezinyathelo ezimbalwa.

Qasha i-H100 SXM Wonke ama-GPU

Imininingwane ye-H100 SXM

VRAM nge-GPU ngayinye

80 GB

Architecture

NVIDIA Hopper

Compute capability

Ukubala kwe-FP32

≈ 53.5 TFLOPS

I-bandwidth yememori

≈ 2869 GB/s

Umkhawulo wamandla

700 W

Amanani abikwa umhlinzeki wengqalasizinda yemishini ebalwe ohlwini njengamanje; ukusebenza kwangempela kuyehluka ngomshini ngamunye.

Izilungiselelo zamanje nezintengo ze-H100 SXM

Isitoko esibukhoma nezintengo, kubuyekezwa ngokuqhubekayo. Intengo yokugcina iqinisekiswa ngaphambi kokuba u-ode.

Qasha manje

Ongakusebenzisa ku-cloud GPU ye-H100 SXM

Ukuqeqesha okudinga i-compute enkulu

I-53.5 TFLOPS FP32 kanye nomkhawulo wamandla ongu-700 W kunikeza i-H100 SXM i-compute enkulu kunanoma yiliphi ikhadi le-Hopper elisohlwini lapha.

I-inference esheshayo ye-LLM

I-bandwidth engu-2869 GB/s isheshisa ukukhiqizwa kwama-token kumamodeli we-13B–34B ku-16-bit noma amamodeli we-70B quantized.

Ukushuna kahle okugcwele kwamamodeli we-7B–13B

I-80 GB ishiya indawo yokushuna kahle kwamapharamitha agcwele kuma-LLM amancane, hhayi ama-adapter kuphela, ngezilungiselelo ezifanele ze-optimizer.

Ukuqeqeshwa kwamamodeli e-diffusion nawevidiyo

Qeqesha noma ushune kahle amamodeli ezithombe nawevidiyo lapho kokubili imemori ne-compute kunquma ukuthi ukusebenza kuthatha isikhathi esingakanani.

Ingabe i-H100 SXM iyi-GPU efanele umsebenzi wakho?

Khetha i-H100 SXM lapho imisebenzi yakho yokuqeqesha noma ye-inference idinga imemori ye-datacenter enomkhawulo ophezulu engu-80 GB nezici ze-architecture ye-Hopper. I-GPU encane ye-workstation noma yabathengi iyisinqumo esingcono sokuthuthukisa okulula noma amamodeli amancane, kanti i-GPU ye-datacenter entsha noma enamandla kakhulu iyona ekhethekayo uma isibonelo esisodwa singakwazi ukugcina imodeli yakho yonke ku-80 GB ye-VRAM.

H100 SXM uma iqhathaniswa nama-cloud GPU afanayo

GPU	VRAM nge-GPU ngayinye	Ukubala kwe-FP32	Umkhawulo wamandla
NVIDIA H100 SXM	80 GB	≈ 53.5 TFLOPS	700 W
NVIDIA H100 PCIE	80 GB	≈ 41 TFLOPS	350 W	Imininingwane →
NVIDIA A800 PCIE	80 GB	≈ 15.6 TFLOPS	300 W	Imininingwane →
NVIDIA H100 NVL	94 GB	≈ 48.3 TFLOPS	400 W	Imininingwane →
NVIDIA CMP 170HX	64 GB	≈ 12.1 TFLOPS	250 W	Imininingwane →

I-cloud GPU ye-H100 SXM: imibuzo evame ukubuzwa

Qasha i-cloud GPU ye-H100 SXM ngehora

Khetha ithempulethi, khetha izilungiselelo zamanje bese uxhuma nge-SSH noma itheminali yesiphequluli. Ukukhokhisa kuyama uma ususa i-GPU.

Qasha i-H100 SXM Wonke ama-GPU