Me ya sa za a zaɓi H100 SXM maimakon H100 PCIE?

An jera H100 SXM a 53.5 TFLOPS FP32 da 2869 GB/s idan aka kwatanta da 41 TFLOPS da 1760 GB/s na katin PCIE, tare da iyakar ƙarfi ta 700 W maimakon 350 W. A kowane aiki yawanci yana gamawa da sauri.

Shin H100 SXM ko H200 NVL ne ya fi kyau ga LLMs?

H200 NVL yana da 140 GB da 3884 GB/s, don haka ya fi kyau ga manyan samfura sosai da dogon mahallin rubutu. H100 SXM yana da ƙarin ƙarfin compute na FP32 (53.5 idan aka kwatanta da 48.3 TFLOPS) kuma ya dace da ayyukan horaswa da ke iya shiga cikin 80 GB.

Zan iya yin fine-tuning na samfurin 70B a kan H100 SXM guda ɗaya?

Tare da QLoRA, e: nauyin 4-bit na 70B yana ɗaukar kusan 35–40 GB, yana barin sarari don adapters, kunnawa (activations) da yanayin optimizer. Cikakken fine-tuning na 16-bit na 70B yana buƙatar GPUs da yawa.

Menene zan biya lokacin da aka dakatar da sabar H100 SXM a kan Cloudj?

A tsarin biyan kuɗi na kowace sa'a, sabar da aka dakatar ana cajin ta ne kawai don diski (disk). Goge sabar da zarar ka kwashe checkpoints ɗinka kuma za ka daina biya gaba ɗaya.

GPU Cloud · NVIDIA · Hopper

Yi Hayar GPU ta NVIDIA H100 SXM Mai 80 GB VRAM

Yi hayar sabar GPU ta girgije ta H100 SXM a kowace sa'a a Cloudj: 80 GB VRAM, NVIDIA Hopper, kusan 53.5 TFLOPS FP32 da 2869 GB/s na bandwidth na ƙwaƙwalwa. Tare da iyakar ƙarfi ta 700 W, wannan shine H100 ga ƙungiyoyin da ke horaswa, daidaitawa (fine-tune) da tura manyan samfura kuma suke son samfurin Hopper mafi sauri a cikin jerin don aikin sarrafa lissafi (compute). Fara tsarin PyTorch, vLLM ko Unsloth sannan ka haɗu ta SSH a matakai kaɗan.

Yi hayar H100 SXM Duk GPUs

Bayanan fasaha na H100 SXM

VRAM a kowace GPU

80 GB

Ginin fasaha

NVIDIA Hopper

Ikon lissafi (Compute capability)

Lissafin FP32

≈ 53.5 TFLOPS

Gudun ƙwaƙwalwa (Bandwidth)

≈ 2869 GB/s

Ƙayyadaddun wutar lantarki

700 W

Ƙimomin da mai samar da kayan aiki ya bayar don na'urorin da aka jera a yanzu; ainihin aikin ya bambanta dangane da na'ura.

Tsare-tsare da farashin H100 SXM na yanzu

Kayan da ke akwai da farashin kai tsaye, ana sabunta su akai-akai. Ana tabbatar da farashin ƙarshe kafin ka yi oda.

Yi haya yanzu

Abubuwan da za ka iya gudanarwa a kan cloud GPU na H100 SXM

Horaswa mai buƙatar ƙarfin sarrafa lissafi mai yawa

53.5 TFLOPS FP32 da iyakar ƙarfin wuta ta 700 W suna ba H100 SXM mafi yawan ƙarfin sarrafa lissafi a cikin duk wani katin Hopper da aka lissafa a nan.

Saurin inference na LLM

2869 GB/s na bandwidth yana hanzarta fitar da kalmomi (token generation) don samfuran 13B–34B a cikin 16-bit ko samfuran 70B da aka yi wa quantized.

Cikakken fine-tuning na samfuran 7B–13B

80 GB yana ba da sarari don cikakken fine-tuning na kowane fanni (full-parameter) na ƙananan LLMs, ba kawai adapters ba, tare da daidaitattun saitin optimizer.

Horas da samfuran diffusion da na bidiyo

Horas ko yi wa samfuran hoto da na bidiyo fine-tuning inda ƙwaƙwalwa da ƙarfin sarrafa lissafi ke tantance tsawon lokacin da aiki zai ɗauka.

Shin H100 SXM ita ce GPU da ta dace da aikinku?

Zaɓi H100 SXM lokacin da ayyukan horaswa ko na inference ɗinku ke buƙatar 80 GB na babban bandwidth na ƙwaƙwalwar cibiyar bayanai da siffofin tsarin Hopper. GPU na workstation ko na masu amfani na yau da kullun zaɓi ne mafi dacewa don ayyukan ƙirƙira marasa nauyi ko ƙananan samfura, yayin da sabuwar GPU ko mai ƙarfin aiki ta cibiyar bayanai ta fi dacewa idan sabar guda ɗaya ba za ta iya ɗaukar duka samfurinku a cikin 80 GB na VRAM ba.

H100 SXM idan aka kwatanta da sauran cloud GPU

GPU	VRAM a kowace GPU	Lissafin FP32	Ƙayyadaddun wutar lantarki
NVIDIA H100 SXM	80 GB	≈ 53.5 TFLOPS	700 W
NVIDIA H100 PCIE	80 GB	≈ 41 TFLOPS	350 W	Cikakken bayani →
NVIDIA A800 PCIE	80 GB	≈ 15.6 TFLOPS	300 W	Cikakken bayani →
NVIDIA H100 NVL	94 GB	≈ 48.3 TFLOPS	400 W	Cikakken bayani →
NVIDIA CMP 170HX	64 GB	≈ 12.1 TFLOPS	250 W	Cikakken bayani →

Cloud GPU na H100 SXM: tambayoyin da aka fi yi

Yi hayar cloud GPU na H100 SXM a kowace sa'a

Zaɓi samfuri, zaɓi tsari na kai tsaye, sannan ka haɗa ta SSH ko terminal na browser. Caji na tsayawa da zarar ka goge GPU ɗin.

Yi hayar H100 SXM Duk GPUs