Shin CMP 170HX yana da kyau don horaswa (training)?

Ya fi dacewa da inference da kuma ayyukan da ke buƙatar ƙwaƙwalwa mai faɗi. Tare da 12.1 TFLOPS FP32, yana da jinkiri wajen horaswa mai nauyi idan aka kwatanta da katuna kamar A40 ko RTX 3090, don haka ka fara gwada aikinka kafin tsara gudanarwa mai tsawo.

Me ya sa ya kamata in fara gwada aikina tukunna?

CMP 170HX katin sarrafa bayanai ne zalla (compute-only) kuma wasu hanyoyin software suna aiki daban a kansa fiye da na kan manyan GPUs na yau da kullum. Gwaji na ɗan gajeren lokaci na sa'a guda zai nuna maka ko tsarin aikinka (framework), drivers da kuma model ɗinka suna aiki kamar yadda aka zata.

Wane siffa ne mafi ƙarfi a CMP 170HX?

Memory bandwidth: kusan 1516 GB/s, wanda ya fi 1313 GB/s da aka lissafa na A100 PCIE girma, wanda ke taimaka wa duk wani aiki da ke buƙatar jiran ƙwaƙwalwar ajiya fiye da jiran lissafi (math).

Har tsawon wane lokaci zan iya riƙe sabar CMP 170HX a kan Cloudj?

Har tsawon lokacin da kake buƙata. Lissafin kuɗi na kowace sa'a yana karɓar kuɗi a kowace sa'a muddin sabar tana aiki, kuma zaka iya goge ta da zarar gwajin ko aikin ya kammala.

GPU Cloud · NVIDIA · Ampere

Yi Hayar GPU ta NVIDIA CMP 170HX Mai 64 GB VRAM

Yi hayar sabar cloud GPU ta CMP 170HX a duk sa'a a kan Cloudj: VRAM 64 GB da aka lissafa, fasahar Ampere da kuma kusan 1516 GB/s memory bandwidth domin sarrafa bayanan da suka dogara kan ƙwaƙwalwar ajiya. Katin sarrafa bayanai ne maras madubi (headless) na masu haɓaka software da ke buƙatar bandwidth da ƙarfin ajiya fiye da gudun FP32 (12.1 TFLOPS, 250 W, compute capability 8.0). Fara gwaji na ɗan lokaci a kan samfurin PyTorch ko Ollama, sannan ka faɗaɗa aikin.

Yi hayar CMP 170HX Duk GPUs

Bayanan fasaha na CMP 170HX

VRAM a kowace GPU

64 GB

Ginin fasaha

NVIDIA Ampere

Ikon lissafi (Compute capability)

Lissafin FP32

≈ 12.1 TFLOPS

Gudun ƙwaƙwalwa (Bandwidth)

≈ 1516 GB/s

Ƙayyadaddun wutar lantarki

250 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 CMP 170HX 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 CMP 170HX

Inference da ke buƙatar bandwidth mai yawa

Ƙirƙirar token ta LLM ta fi dogaro ne kan bandwidth na ƙwaƙwalwar ajiya, kuma 1516 GB/s yana daidai da kewayon katunan rukunin A100.

Ɗaukar nauyin quantized LLM

Loda quantized models a Ollama ko llama.cpp sannan ka bayar da su ta API don yin gwaji ko kayan aikin cikin gida.

Gwajin CUDA kernel

Bincika da auna kernels masu buƙatar ƙwaƙwalwa sosai a katin Ampere mai compute capability 8.0 kafin ka mayar da su zuwa manyan na'urori masu ƙarfi.

Sarrafa tarin bayanai (Batch data processing)

Gudanar da manyan arrays ta cikin ƙwaƙwalwar GPU don ayyukan canza bayanai, hashing, ko binciken bayanai (analytics) da ke buƙatar shigar ƙwaƙwalwa da sauri.

Shin CMP 170HX ita ce GPU da ta dace da aikinku?

Zaɓi CMP 170HX idan aikinka na fuskantar cikas saboda ƙarancin girman ƙwaƙwalwa da bandwidth maimakon raw single-precision compute. Zaɓi GPU mai ƙarfin sarrafa bayanai mafi girma idan kana buƙatar hanzarta lokutan horaswa (training iterations), ko ka zaɓi ƙaramin katin amfanin yau da kullum idan models ɗinka ba sa buƙatar 64 GB na VRAM.

CMP 170HX idan aka kwatanta da sauran cloud GPU

GPU	VRAM a kowace GPU	Lissafin FP32	Ƙayyadaddun wutar lantarki
NVIDIA CMP 170HX	64 GB	≈ 12.1 TFLOPS	250 W
NVIDIA A800 PCIE	80 GB	≈ 15.6 TFLOPS	300 W	Cikakken bayani →
NVIDIA H100 PCIE	80 GB	≈ 41 TFLOPS	350 W	Cikakken bayani →
NVIDIA H100 SXM	80 GB	≈ 53.5 TFLOPS	700 W	Cikakken bayani →
NVIDIA A40	45 GB	≈ 29.9 TFLOPS	300 W	Cikakken bayani →

Cloud GPU na CMP 170HX: tambayoyin da aka fi yi

Yi hayar cloud GPU na CMP 170HX 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 CMP 170HX Duk GPUs