{"success":true,"data":[{"id":"fc070c6d-c642-4bf2-b60e-95c3e20d7ca9","name":"RTX PRO 6000 WS 96GB","architecture":"Blackwell","vramGb":96,"memBandwidthGbs":1792,"hourlyPrice":0.83,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":1.2,"publicPriceMin":1.5625,"publicPriceMax":2.128437,"liveRentablePrice":1.8716,"offers":[],"savingsPct":31,"demandStatus":"售罄","demandStatusOverride":null,"totalGpus":7,"availableGpus":0,"availability":"BUSY","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":true,"purchaseModelCode":null,"sortOrder":0,"heroImageUrl":"/gpus/rtx-pro-6000.jpg","tagline":"NVIDIA's flagship workstation GPU - 96 GB GDDR7, Blackwell architecture.","taglineZh":"NVIDIA 旗艦工作站 GPU - 96 GB GDDR7、Blackwell 架構。","description":"The NVIDIA RTX PRO 6000 Blackwell Workstation Edition is the most powerful professional desktop GPU ever built. Featuring 96 GB of ECC GDDR7 memory and 24,064 CUDA cores on the full GB202 die, it handles AI inference, 3D rendering, scientific simulation, and video production workflows that no previous workstation GPU could manage in a single card. Its 1.79 TB/s memory bandwidth makes it particularly effective for large-model inference and fine-tuning tasks that exceed the capacity of consumer GPUs.","descriptionZh":"NVIDIA RTX PRO 6000 Blackwell Workstation Edition 是頂級專業桌面 GPU，搭載 96 GB ECC GDDR7 記憶體與 24,064 個 CUDA 核心，適合大型模型推論、LoRA 或完整微調、3D/VFX 渲染、科學模擬與 AI 影片生成。1.79 TB/s 記憶體頻寬與大容量 VRAM 讓它能承載許多消費級 GPU 無法一次放入的模型與資料。若您的工作負載需要單卡高記憶體容量、可靠性與穩定長時間運算，這是 GPUtw 上最完整的工作站級選擇。","chipName":"GB202","processNode":"5 nm (TSMC 4NP)","transistorsBillion":92.2,"releaseDate":"Mar 18, 2025","launchPriceUsd":null,"cudaCores":24064,"tensorCoreGen":5,"tensorCores":752,"rtCoreGen":4,"rtCores":188,"baseClockMhz":1590,"boostClockMhz":2617,"memoryType":"GDDR7 ECC","memoryBusBits":512,"tdpW":500,"fp32Tflops":126,"fp16Tflops":126,"aiTops":4000,"cardCount":1,"perCardMemoryGb":null,"perCardCudaCores":null,"interconnect":null,"idealFor":["Large model inference (70 B+ params)","LLM fine-tuning (LoRA / full)","3D & VFX rendering","Scientific simulation","AI video generation"]},{"id":"78382b9a-476c-47fe-b60b-6023f1b65d98","name":"DGX Spark x3 384GB Link","architecture":"GB10 Blackwell","vramGb":384,"memBandwidthGbs":819,"hourlyPrice":2.578125,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":null,"publicPriceMin":null,"publicPriceMax":null,"liveRentablePrice":null,"offers":[],"savingsPct":null,"demandStatus":"售罄","demandStatusOverride":null,"totalGpus":0,"availableGpus":0,"availability":"NO_HARDWARE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":false,"purchaseModelCode":null,"sortOrder":1,"heroImageUrl":"/gpus/dgx-spark.jpg","tagline":"Three DGX Spark systems linked over ConnectX-7 - 384 GB of combined unified memory.","taglineZh":"三台 DGX Spark 透過 ConnectX-7 互連 - 384 GB 統一記憶體。","description":"This configuration links three NVIDIA DGX Spark systems through their ConnectX-7 200 Gb/s ports into a three-node cluster with 384 GB of total unified memory, 18,432 CUDA cores and about 3 PFLOPS of FP4 AI compute. It fits models and context windows that do not fit on one or two Sparks, and distributed fine-tuning across three GB10 nodes, while keeping every byte on-premise in Taiwan.","descriptionZh":"此配置將三台 NVIDIA DGX Spark 透過 ConnectX-7 200 Gb/s 連結組成三節點叢集，提供 384 GB 總統一記憶體、18,432 個 CUDA 核心與約 3 PFLOPS 的 FP4 AI 算力。適合單台或雙台 Spark 放不下的模型與超長上下文，以及跨三個 GB10 節點的分散式微調，且所有資料都留在台灣機房。","chipName":"GB10 Superchip x 3","processNode":"4 nm (TSMC 4NP)","transistorsBillion":null,"releaseDate":"Mar 2025 (available Oct 2025)","launchPriceUsd":11997,"cudaCores":18432,"tensorCoreGen":5,"tensorCores":null,"rtCoreGen":4,"rtCores":null,"baseClockMhz":null,"boostClockMhz":null,"memoryType":"LPDDR5X (unified CPU+GPU)","memoryBusBits":768,"tdpW":720,"fp32Tflops":15,"fp16Tflops":null,"aiTops":3000,"cardCount":3,"perCardMemoryGb":128,"perCardCudaCores":6144,"interconnect":"ConnectX-7 200 Gb/s node link (QSFP)","idealFor":["300 B-class LLM inference at FP4 across three nodes","Multi-node fine-tuning with quantization","Three independent 128 GB Spark instances","Agentic AI with very large context windows","Privacy-sensitive on-premise inference"]},{"id":"7d76fe1f-d81d-47e8-a8b9-726f101d2165","name":"DGX Spark x2 256GB Link","architecture":"GB10 Blackwell","vramGb":256,"memBandwidthGbs":546,"hourlyPrice":1.71875,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":null,"publicPriceMin":null,"publicPriceMax":null,"liveRentablePrice":null,"offers":[],"savingsPct":null,"demandStatus":"售罄","demandStatusOverride":null,"totalGpus":0,"availableGpus":0,"availability":"NO_HARDWARE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":true,"purchaseModelCode":null,"sortOrder":1,"heroImageUrl":"/gpus/dgx-spark.jpg","tagline":"Two DGX Spark systems linked over ConnectX-7 - 256 GB of combined unified memory.","taglineZh":"兩台 DGX Spark 透過 ConnectX-7 互連 - 256 GB 統一記憶體。","description":"This configuration pairs two NVIDIA DGX Spark systems over their built-in ConnectX-7 200 Gb/s link, giving 256 GB of aggregate LPDDR5X unified memory, two GB10 Superchips (40 Arm cores, 12,288 CUDA cores) and up to 2,000 AI TOPS. It is the NVIDIA-supported two-node Spark cluster: models that do not fit in 128 GB (for example 200 B-class LLMs at FP4) can be served with tensor or pipeline parallelism across the pair, and fine-tuning or agentic workloads gain a second GPU without leaving the low-power desktop form factor. The pair can also be rented as two independent 128 GB DGX Spark instances.","descriptionZh":"此配置將兩台 NVIDIA DGX Spark 透過內建的 ConnectX-7 200 Gb/s 連結組成 NVIDIA 官方支援的雙節點叢集，提供 256 GB 總統一記憶體、兩顆 GB10 Superchip（40 核 Arm CPU、12,288 CUDA 核心）與最高 2,000 AI TOPS。單機 128 GB 放不下的模型（例如 FP4 的 200B 級 LLM）可用張量或管線平行跨兩台推論，微調與 agentic 工作負載也能多一顆 GPU。這一對機器也可以拆成兩個獨立的 128 GB DGX Spark 執行個體租用。","chipName":"GB10 Superchip x 2","processNode":"4 nm (TSMC 4NP)","transistorsBillion":null,"releaseDate":"Mar 2025 (available Oct 2025)","launchPriceUsd":7998,"cudaCores":12288,"tensorCoreGen":5,"tensorCores":null,"rtCoreGen":4,"rtCores":null,"baseClockMhz":null,"boostClockMhz":null,"memoryType":"LPDDR5X (unified CPU+GPU)","memoryBusBits":512,"tdpW":480,"fp32Tflops":10,"fp16Tflops":null,"aiTops":2000,"cardCount":2,"perCardMemoryGb":128,"perCardCudaCores":6144,"interconnect":"ConnectX-7 200 Gb/s node link (QSFP)","idealFor":["200 B-class LLM inference at FP4 across two nodes","Multi-node fine-tuning with quantization","Two independent 128 GB Spark instances","Agentic AI with large context windows","Privacy-sensitive on-premise inference"]},{"id":"0e28d9c6-3156-4ba2-bb99-d5f6679e858e","name":"DGX Spark 128GB (GX10)","architecture":"GB10 Blackwell","vramGb":128,"memBandwidthGbs":273,"hourlyPrice":0.690625,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":null,"publicPriceMin":null,"publicPriceMax":null,"liveRentablePrice":0.6906,"offers":[{"gpuCount":1,"hourlyRate":0.6906,"isSlice":false}],"savingsPct":null,"demandStatus":"可使用","demandStatusOverride":null,"totalGpus":1,"availableGpus":1,"availability":"AVAILABLE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":true,"purchaseModelCode":null,"sortOrder":1,"heroImageUrl":"/gpus/dgx-spark.jpg","tagline":"NVIDIA's personal AI supercomputer - GB10 Superchip with 128 GB unified memory.","taglineZh":"NVIDIA 個人 AI 超級電腦 - GB10 Superchip 與 128 GB 統一記憶體。","description":"The NVIDIA DGX Spark is a compact desktop AI supercomputer powered by the GB10 Superchip, combining a 20-core Arm CPU with a Blackwell-generation GPU in a small desktop chassis. Its 128 GB of unified LPDDR5X memory is shared seamlessly between CPU and GPU, enabling models too large for any single-card GPU to run locally. With up to 1,000 AI TOPS and 1 PFLOP at FP4 precision, it is purpose-built for local LLM inference, fine-tuning with quantized weights, and agentic AI workflows that require low-latency GPU access.","descriptionZh":"NVIDIA DGX Spark 是小型桌面 AI 超級電腦，採用 GB10 Superchip，將 20 核 Arm CPU 與 Blackwell 世代 GPU 整合在同一系統中。128 GB LPDDR5X 統一記憶體可由 CPU 與 GPU 共用，適合在本地執行單卡 GPU 放不下的大型模型、長上下文推論、量化微調與 agentic AI 工作流。它的重點不是傳統顯卡式峰值效能，而是以低延遲、低功耗與大記憶體池支援隱私敏感或原型開發場景。","chipName":"GB10 Superchip","processNode":"4 nm (TSMC 4NP)","transistorsBillion":null,"releaseDate":"Mar 2025 (available Oct 2025)","launchPriceUsd":3999,"cudaCores":6144,"tensorCoreGen":5,"tensorCores":null,"rtCoreGen":4,"rtCores":null,"baseClockMhz":0,"boostClockMhz":0,"memoryType":"LPDDR5X (unified CPU+GPU)","memoryBusBits":256,"tdpW":240,"fp32Tflops":5,"fp16Tflops":null,"aiTops":1000,"cardCount":1,"perCardMemoryGb":null,"perCardCudaCores":null,"interconnect":null,"idealFor":["Large model inference (70 B+ at int4)","Local agentic AI workflows","LLM fine-tuning with quantization","Prototyping on large context windows","Privacy-sensitive inference (on-premise)"]},{"id":"c0c092fc-c55f-47be-aa9e-b11d627f73e3","name":"RTX 5090 32GB","architecture":"Blackwell","vramGb":32,"memBandwidthGbs":1792,"hourlyPrice":0.32,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":0.4,"publicPriceMin":0.75,"publicPriceMax":0.949063,"liveRentablePrice":0.9038,"offers":[{"gpuCount":1,"hourlyRate":0.9038,"isSlice":false}],"savingsPct":20,"demandStatus":"中度用量","demandStatusOverride":null,"totalGpus":11,"availableGpus":3,"availability":"AVAILABLE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":true,"purchaseModelCode":null,"sortOrder":2,"heroImageUrl":"/gpus/rtx-5090.jpg","tagline":"NVIDIA's fastest consumer GPU - Blackwell architecture with 32 GB GDDR7.","taglineZh":"NVIDIA Blackwell 世代最快消費級 GPU - 32 GB GDDR7。","description":"The GeForce RTX 5090 is the flagship consumer GPU of NVIDIA's Blackwell generation. Built on the same GB202 die as the RTX PRO 6000 WS with 21,760 active CUDA cores and 32 GB of GDDR7 on a 512-bit bus, it delivers the same 1.79 TB/s memory bandwidth at a fraction of the cost. The RTX 5090 offers exceptional performance for AI inference workloads up to 32 GB, diffusion model generation, and gaming, making it the most versatile Blackwell option for general GPU compute.","descriptionZh":"GeForce RTX 5090 是 Blackwell 世代旗艦消費級 GPU，擁有 21,760 個 CUDA 核心、32 GB GDDR7 與 512-bit 記憶體匯流排，頻寬最高達 1.79 TB/s。它非常適合 32 GB VRAM 以內的 AI 推論、影像與影片生成、LoRA 微調、電腦視覺訓練與即時渲染。相較專業卡，RTX 5090 以更低成本提供極高吞吐量，是一般 AI 開發與創作工作流的高效選擇。","chipName":"GB202","processNode":"5 nm (TSMC 4NP)","transistorsBillion":92.2,"releaseDate":"Jan 30, 2025","launchPriceUsd":null,"cudaCores":21760,"tensorCoreGen":5,"tensorCores":680,"rtCoreGen":4,"rtCores":170,"baseClockMhz":2010,"boostClockMhz":2407,"memoryType":"GDDR7","memoryBusBits":512,"tdpW":500,"fp32Tflops":104.8,"fp16Tflops":209.6,"aiTops":3352,"cardCount":1,"perCardMemoryGb":null,"perCardCudaCores":null,"interconnect":null,"idealFor":["AI inference (up to 32 GB models)","Image & video generation","LLM fine-tuning (LoRA)","Computer vision training","Gaming & real-time rendering"]},{"id":"d6170519-d623-47b2-9760-a20ed4b64db6","name":"RTX 3090 24GB","architecture":"Ampere","vramGb":24,"memBandwidthGbs":936,"hourlyPrice":0.1375,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":0.15,"publicPriceMin":0.25,"publicPriceMax":0.428438,"liveRentablePrice":0.3466,"offers":[{"gpuCount":1,"hourlyRate":0.3466,"isSlice":false}],"savingsPct":8,"demandStatus":"中度用量","demandStatusOverride":null,"totalGpus":12,"availableGpus":4,"availability":"AVAILABLE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":true,"purchaseModelCode":null,"sortOrder":3,"heroImageUrl":"/gpus/rtx-3090.jpg","tagline":"Flagship Ampere consumer GPU - 24 GB GDDR6X, 10,496 CUDA cores.","taglineZh":"Ampere 旗艦消費級 GPU - 24 GB GDDR6X、10,496 CUDA 核心。","description":"The GeForce RTX 3090 was NVIDIA's highest-end consumer GPU of the Ampere generation. With 24 GB of GDDR6X on a 384-bit bus delivering 936 GB/s of bandwidth, it remains a capable workhorse for AI training and inference tasks that fit in 24 GB. Its 10,496 CUDA cores and third-generation Tensor Cores provide strong FP16 and TF32 throughput, making it a cost-effective option for fine-tuning models up to about 13 B parameters with quantization, stable diffusion image generation, and traditional deep learning workloads.","descriptionZh":"GeForce RTX 3090 具備 24 GB GDDR6X、936 GB/s 記憶體頻寬與第三代 Tensor Cores，至今仍是 AI 訓練與推論的實用工作馬。它適合 24 GB VRAM 內的模型微調、小批次 LLM 推論、Stable Diffusion / ComfyUI、電腦視覺訓練與影片處理。對需要比入門 GPU 更大記憶體、但不需要最新 Blackwell 成本的使用者，RTX 3090 是很有性價比的選擇。","chipName":"GA102","processNode":"8 nm (Samsung)","transistorsBillion":28.3,"releaseDate":"Sep 24, 2020","launchPriceUsd":1499,"cudaCores":10496,"tensorCoreGen":3,"tensorCores":328,"rtCoreGen":2,"rtCores":82,"baseClockMhz":1395,"boostClockMhz":1695,"memoryType":"GDDR6X","memoryBusBits":384,"tdpW":300,"fp32Tflops":35.6,"fp16Tflops":71,"aiTops":null,"cardCount":1,"perCardMemoryGb":null,"perCardCudaCores":null,"interconnect":null,"idealFor":["Fine-tuning up to 13 B (int4)","Stable Diffusion & ComfyUI","Small-batch LLM inference","Computer vision training","Video processing"]},{"id":"750748eb-8229-48f5-8d25-a56899cddd52","name":"RTX 3090 x2 48GB NVLink","architecture":"Ampere","vramGb":48,"memBandwidthGbs":1872,"hourlyPrice":0.1375,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":0.3,"publicPriceMin":0.625,"publicPriceMax":0.82,"liveRentablePrice":0.785,"offers":[{"gpuCount":2,"hourlyRate":0.785,"isSlice":false}],"savingsPct":54,"demandStatus":"中度用量","demandStatusOverride":null,"totalGpus":4,"availableGpus":2,"availability":"AVAILABLE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":true,"purchaseModelCode":null,"sortOrder":4,"heroImageUrl":"/gpus/rtx-3090.jpg","tagline":"Two RTX 3090 GPUs bridged by NVLink - 48 GB of combined Ampere VRAM.","taglineZh":"兩張 RTX 3090 透過 NVLink 連接 - 48 GB Ampere VRAM。","description":"This configuration pairs two GeForce RTX 3090 cards via NVIDIA NVLink, presenting 48 GB of aggregate GDDR6X VRAM as a unified address space. NVLink provides up to 112.5 GB/s of bidirectional bandwidth between the cards, dramatically outpacing PCIe for inter-GPU communication. The combined 20,992 CUDA cores and 1.87 TB/s of total memory bandwidth make this config well-suited for model-parallel training of LLMs that fit in 48 GB, as well as batch inference workloads that benefit from high aggregate throughput.","descriptionZh":"此配置使用兩張 GeForce RTX 3090 並透過 NVIDIA NVLink 連接，提供 48 GB 總 GDDR6X VRAM 與更快的卡間通訊。NVLink 對模型平行、批次推論與需要跨 GPU 傳輸的工作負載特別有利，比只靠 PCIe 更適合多卡協作。對於可切分到 48 GB 記憶體內的 LLM 訓練、3D 渲染、資料科學管線與高吞吐推論，這是成本與容量平衡良好的 Ampere 多卡方案。","chipName":"GA102 x 2","processNode":"8 nm (Samsung)","transistorsBillion":56.6,"releaseDate":"Sep 24, 2020","launchPriceUsd":null,"cudaCores":20992,"tensorCoreGen":3,"tensorCores":656,"rtCoreGen":2,"rtCores":164,"baseClockMhz":1395,"boostClockMhz":1695,"memoryType":"GDDR6X","memoryBusBits":768,"tdpW":700,"fp32Tflops":71.2,"fp16Tflops":142.4,"aiTops":null,"cardCount":2,"perCardMemoryGb":24,"perCardCudaCores":10496,"interconnect":"NVLink 3.0 (112.5 GB/s bidirectional)","idealFor":["Model-parallel LLM training (<= 48 GB)","Batch inference with large context","3D rendering farms","Data science pipelines"]},{"id":"4420e1e5-2c1f-4ac3-a0c3-05c84820d779","name":"H100 x 8","architecture":"Hopper","vramGb":80,"memBandwidthGbs":3350,"hourlyPrice":2.03125,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":2.39,"publicPriceMin":1.9375,"publicPriceMax":2.34375,"liveRentablePrice":null,"offers":[],"savingsPct":15,"demandStatus":"售罄","demandStatusOverride":null,"totalGpus":0,"availableGpus":0,"availability":"NO_HARDWARE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":false,"purchaseModelCode":null,"sortOrder":5,"heroImageUrl":"/gpus/h200.jpg","tagline":"Hopper data center GPU - 80 GB HBM3 for LLM inference, training, and HPC.","taglineZh":"Hopper 資料中心 GPU - 80 GB HBM3，適合 LLM 推論、訓練與 HPC。","description":"The NVIDIA H100 Tensor Core GPU is a Hopper-generation data center accelerator built for large language models, generative AI, recommender systems, and high-performance computing. The SXM 80 GB variant provides 80 GB of HBM3 memory, 3.35 TB/s of memory bandwidth, fourth-generation Tensor Cores, FP8 acceleration, and MIG partitioning for multi-tenant workloads. It remains one of the most widely supported accelerators for production LLM inference, fine-tuning, distributed training, and frameworks that are already tuned for Hopper.","descriptionZh":"NVIDIA H100 Tensor Core GPU 是 Hopper 世代資料中心加速器，面向大型語言模型、生成式 AI、推薦系統與高效能運算。SXM 80 GB 版本提供 80 GB HBM3 記憶體、3.35 TB/s 記憶體頻寬、第四代 Tensor Cores、FP8 加速與 MIG 分割能力。它仍是生產級 LLM 推論、模型微調、分散式訓練，以及已針對 Hopper 最佳化框架中最常見且支援成熟的 GPU 之一。","chipName":"GH100","processNode":"4 nm (TSMC 4N)","transistorsBillion":null,"releaseDate":"Mar 2022","launchPriceUsd":30000,"cudaCores":16896,"tensorCoreGen":4,"tensorCores":528,"rtCoreGen":null,"rtCores":null,"baseClockMhz":0,"boostClockMhz":0,"memoryType":"HBM3","memoryBusBits":5120,"tdpW":700,"fp32Tflops":67,"fp16Tflops":1979,"aiTops":3958,"cardCount":1,"perCardMemoryGb":null,"perCardCudaCores":null,"interconnect":"NVLink 4 (platform dependent)","idealFor":["Production LLM inference","LLM fine-tuning","Distributed deep learning","RAG and embedding workloads","HPC and simulation"]},{"id":"71eba9e3-be6c-4631-a9ea-3aa079ceb5d4","name":"RTX 4090 24GB","architecture":"Ada Lovelace","vramGb":24,"memBandwidthGbs":1008,"hourlyPrice":0.28,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":0.35,"publicPriceMin":0.5625,"publicPriceMax":0.697812,"liveRentablePrice":0.6978,"offers":[],"savingsPct":20,"demandStatus":"售罄","demandStatusOverride":null,"totalGpus":1,"availableGpus":0,"availability":"BUSY","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":true,"purchaseModelCode":null,"sortOrder":9,"heroImageUrl":"/gpus/rtx-4090.jpg","tagline":"Ada Lovelace flagship consumer GPU - 24 GB GDDR6X, 16,384 CUDA cores.","taglineZh":"Ada Lovelace 旗艦消費級 GPU - 24 GB GDDR6X、16,384 CUDA 核心。","description":"The GeForce RTX 4090 is NVIDIA's flagship Ada Lovelace consumer GPU. With 24 GB of GDDR6X memory, 16,384 CUDA cores, fourth-generation Tensor Cores, third-generation RT Cores, and 1,008 GB/s of memory bandwidth, it remains a strong option for AI inference, image generation, video workflows, rendering, and deep learning workloads that fit in 24 GB of VRAM.","descriptionZh":"GeForce RTX 4090 是 NVIDIA Ada Lovelace 世代旗艦消費級 GPU，搭載 24 GB GDDR6X、16,384 個 CUDA 核心、第四代 Tensor Cores、第三代 RT Cores 與 1,008 GB/s 記憶體頻寬。它適合 24 GB VRAM 以內的 AI 推論、Stable Diffusion 和影像生成、量化微調、3D 渲染、影片工作流與電腦視覺訓練。對需要高效能但不需要 32 GB 以上 VRAM 的使用者，RTX 4090 仍是非常強的選擇。","chipName":"AD102","processNode":"4 nm (TSMC 4N)","transistorsBillion":76.3,"releaseDate":"Oct 12, 2022","launchPriceUsd":1599,"cudaCores":16384,"tensorCoreGen":4,"tensorCores":512,"rtCoreGen":3,"rtCores":128,"baseClockMhz":2235,"boostClockMhz":2520,"memoryType":"GDDR6X","memoryBusBits":384,"tdpW":350,"fp32Tflops":82.6,"fp16Tflops":165.2,"aiTops":1321,"cardCount":1,"perCardMemoryGb":null,"perCardCudaCores":null,"interconnect":null,"idealFor":["AI inference up to 24 GB models","Stable Diffusion and image generation","LLM fine-tuning with quantization","3D rendering and video workflows","Computer vision training"]},{"id":"5506385c-b98e-4d48-8b2b-621859a15169","name":"HGX B300 XE9780","architecture":"Blackwell Ultra","vramGb":270,"memBandwidthGbs":8000,"hourlyPrice":5.6,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":null,"publicPriceMin":5.6,"publicPriceMax":6.9,"liveRentablePrice":null,"offers":[],"savingsPct":null,"demandStatus":"售罄","demandStatusOverride":null,"totalGpus":0,"availableGpus":0,"availability":"NO_HARDWARE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":true,"purchaseModelCode":"WS-B300","sortOrder":11,"heroImageUrl":"/api/gpu-images/614f22b43e755a0347b0ee11.jpg","tagline":"Blackwell Ultra data center GPU - 270 GB HBM3e per GPU in HGX B300. Rent it hourly, or buy the Dell server.","taglineZh":"Blackwell Ultra 資料中心 GPU - HGX B300 每卡 270 GB HBM3e。可租用，也可購買 Dell 整機。","description":"NVIDIA B300 pairs Blackwell Ultra compute with 270 GB of HBM3e per GPU at 8 TB/s (2.16 TB across an HGX B300 board). Rent it by the hour on GPUtw, or own it outright: we deliver and support Dell PowerEdge XE9785 B300 servers in Taiwan.","descriptionZh":"NVIDIA B300 搭載 Blackwell Ultra 架構，每卡 270 GB HBM3e（8 TB/s 頻寬，HGX B300 整板共 2.16 TB）。可在 GPUtw 按小時租用，也可直接向我們購買 Dell PowerEdge XE9785 整機 — 台灣本地交付與保固支援。","chipName":"B300","processNode":null,"transistorsBillion":null,"releaseDate":"2025","launchPriceUsd":null,"cudaCores":null,"tensorCoreGen":null,"tensorCores":null,"rtCoreGen":null,"rtCores":null,"baseClockMhz":null,"boostClockMhz":null,"memoryType":"HBM3e","memoryBusBits":null,"tdpW":1400,"fp32Tflops":null,"fp16Tflops":null,"aiTops":null,"cardCount":1,"perCardMemoryGb":null,"perCardCudaCores":null,"interconnect":null,"idealFor":[]},{"id":"30581c0d-8bd9-49f3-b0bc-4f64d46a0615","name":"V100 16GB","architecture":"Volta","vramGb":16,"memBandwidthGbs":900,"hourlyPrice":0.12,"cpuPerCoreHr":0.005,"diskPerGbHr":0.0002,"vastAiPrice":0.15,"publicPriceMin":0.125,"publicPriceMax":0.1875,"liveRentablePrice":null,"offers":[],"savingsPct":20,"demandStatus":"售罄","demandStatusOverride":null,"totalGpus":0,"availableGpus":0,"availability":"NO_HARDWARE","rentalMode":"SELF_SERVE","contactEmail":"support.gputw.ai@gmail.com","contactRentalNote":"Email-based long-session rental only.","isActive":true,"pinnedHome":false,"purchaseModelCode":null,"sortOrder":14,"heroImageUrl":"/api/gpu-images/9001f96e175e208aba611162.webp","tagline":"The GPU that launched the Tensor Core era - 16 GB HBM2, first-gen Tensor Cores.","taglineZh":"開啟 Tensor Core 時代的經典 GPU - 16 GB HBM2。","description":"The NVIDIA Tesla V100 SXM2 introduced first-generation Tensor Cores, enabling mixed-precision training that became the standard technique for large-scale deep learning. With 5,120 CUDA cores, 640 Tensor Cores, and 16 GB of HBM2 delivering 900 GB/s of bandwidth, it remains a reliable platform for training and inference tasks up to 16 GB, particularly workloads with mature codebases tuned for V100 compatibility. Its high-bandwidth HBM2 memory makes it efficient for bandwidth-bound operations despite its age.","descriptionZh":"NVIDIA Tesla V100 SXM2 是第一代 Tensor Core GPU，支援讓深度學習普及的混合精度訓練。它具備 5,120 個 CUDA 核心、640 個 Tensor Cores、16 GB HBM2 與 900 GB/s 記憶體頻寬，適合 BERT、ResNet、傳統 Transformer、科學計算、資料前處理與 16 GB 以內的推論任務。V100 的優勢在於成熟框架相容性與高頻寬記憶體，對舊版研究程式碼尤其友善。","chipName":"GV100","processNode":"12 nm (TSMC)","transistorsBillion":21.1,"releaseDate":"Jun 2017","launchPriceUsd":10000,"cudaCores":5120,"tensorCoreGen":1,"tensorCores":640,"rtCoreGen":null,"rtCores":null,"baseClockMhz":1245,"boostClockMhz":1530,"memoryType":"HBM2","memoryBusBits":4096,"tdpW":300,"fp32Tflops":15.7,"fp16Tflops":125.4,"aiTops":null,"cardCount":1,"perCardMemoryGb":null,"perCardCudaCores":null,"interconnect":null,"idealFor":["Classic deep learning (BERT, ResNet, Transformer)","LLM inference (<= 16 GB models)","Scientific computing","Data preprocessing pipelines"]}],"error":null}