B200 on Novita AI: Only $4.77/hr for Running DeepSeek R1!

The NVIDIA B200 is here—bare metal, blazing fast, and now available at Novita AI.

Experience over 30,000 tokens per second on DeepSeek-R1 with the NVIDIA B200—powered by next-gen Blackwell architecture and built for LLMs, generative AI, and enterprise-scale inference. No hardware? No problem. Novita AI gives you on-demand access at a fraction of the cost.

What is B200?

The NVIDIA B200 is a next-generation GPU built on the Blackwell architecture, delivering breakthrough AI performance with up to 72 petaFLOPS for training and 144 petaFLOPS for inference, making it ideal for large language models and generative AI workloads.

Key Performance Metrics

Component	Details
GPU	8x NVIDIA Blackwell GPUs
GPU Memory	1,440GB total, 64TB/s HBM3e bandwidth
Performance	72 petaFLOPS FP8 training and 144 petaFLOPS FP4 inference
NVIDIA NVSwitch	2x
NVIDIA NVLink Bandwidth	14.4 TB/s aggregate bandwidth
System Power Usage	~14.3kW max
CPU	2x Intel Xeon Platinum 8570 Processors, 112 Cores total, 2.1 GHz (Base), 4 GHz (Max Boost)
System Memory	2TB, configurable to 4TB
Networking	4x OSFP ports (8x single-port NVIDIA ConnectX-7 VPI), Up to 400Gb/s Infiniband/Ethernet; 2x dual-port QSFP112 BlueField-3 DPU, Up to 400Gb/s Infiniband/Ethernet
Management Network	10Gb/s onboard NIC with RJ45; 100Gb/s dual-port Ethernet NIC; BMC with RJ45
Storage	OS: 2x 1.9TB NVMe M.2, Internal: 8x 3.84TB NVMe U.2
Software	NVIDIA AI Enterprise, NVIDIA Mission Control, NVIDIA Runai Technology, NVIDIA DGX OS / Ubuntu
Rack Units (RU)	10 RU
System Dimensions	Height: 17.5in (444mm), Width: 19.0in (482.2mm), Length: 35.3in (897.1mm)
Operating Temperature	5–30°C (41–86°F)
Enterprise Support	3-year standard support for hardware/software; 24/7 support portal access; live agent during business hours

This is essentially an AI supercomputer that can:

• Train advanced AI models (like ChatGPT, Claude)
• Serve thousands of AI inference requests simultaneously
• Process massive datasets for research or business intelligence
• Power AI applications for entire organizations

Cost Efficiency of the B200

This price isn’t for a single GPU. It typically covers a full DGX system with:

• 8x Blackwell B200 GPUs
• Dual Xeon CPUs
• High-speed NVLink & NVSwitch interconnects
• 1.44TB HBM3e GPU memory
• 2TB+ system RAM
• Enterprise-grade networking, SSDs, and software stack (NVIDIA AI Enterprise, Run:ai, etc.)

Application of B200

1. AI Training and Inference

The B200 excels in both training and inference of large-scale AI models, particularly large language models (LLMs) and generative AI applications.

• Training: With its second-generation Transformer Engine and support for FP4 precision, the B200 accelerates the training process of massive models, reducing time and computational resources required.
• Inference: The B200 delivers exceptional inference performance, achieving over 1,000 tokens per second per user in benchmarks with models like Meta’s Llama 4 Maverick.

2. Graphics and Visualization

While primarily designed for AI workloads, the B200 also supports advanced graphics and visualization tasks:

• Ray Tracing: Equipped with 4th-generation RT Cores, the B200 can perform real-time ray tracing, enabling high-fidelity rendering for complex scenes.
• Shader Technology: The GPU’s architecture includes enhancements like Shader Execution Reordering (SER) 2.0, optimizing the execution of complex shader programs and compute kernels.

3. Precision Workloads

The B200 is adept at handling precision-intensive tasks across various scientific and industrial domains:

• The NVIDIA B200 GPU, based on the Blackwell architecture, does support floating-point (FP) operations through its fifth-generation Tensor Cores. These Tensor Cores are designed to handle a variety of FP precisions, including FP4, FP6, FP8, FP16, and TF32, enabling efficient acceleration of AI workloads.

B200+Deepseek R1=World Record 

GPU Setup	Total Throughput	Notes
8x NVIDIA B200 (DGX B200)	Over 30,000 tokens/s	Achieved using TensorRT-LLM with FP4 precision
8x NVIDIA H200 (HGX H200)	Up to 3,872 tokens/s	Utilized NVIDIA NIM microservice
8x NVIDIA H100 (Quantized 4-bit)	Approximately 2,500 tokens/s	Deployed with vLLM 0.7.3
4x NVIDIA H100	25 tokens/s	Real-world test on Lambda Cloud
Single NVIDIA RTX 4090 (24GB)	3–33 tokens/s	Performance varies based on quantization and configuration

In summary, the NVIDIA B200 GPU stands out as the top performer for DeepSeek-R1 inference, offering unparalleled throughput that enables efficient deployment of large language models in enterprise environments.

How to Run B200 at a very Cost-effectively Price?

Novita AI is an AI cloud platform that offers developers an easy way to deploy AI models using our simple API, while also providing a affordable and reliable GPU cloud for building and scaling.

Step 1: Log In and Access the GPU Bare Metal

Log in to your account and click on the GPU Bare Metal button.

Step2: Chooose Your GPU

Select the Device

• Device Name: Choose H100 SXM or B200 SXM.
• Region: United States of America.
• Configuration (for H100 SXM):
  • 8 GPUs
  • 2048 GB Memory
  • 104vCPU/Node
  • 15.36 TB Storage
  • at $1.7./hour.
• Configuration (for B200 SXM):
  • 8 GPUs
  • 2304 GB Memory
  • 144vCPU/Node
  • 30.8 TB Storage
  • at $4.77./hour.

Set the Quantity and Rental Duration

• Adjust the GPU Quantity field to match your needs. For example, select 8 GPUs.

• Choose the rental duration. For instance, set it to 1 month.

The NVIDIA B200 GPU, built on the Blackwell architecture, is a cutting-edge AI supercomputer designed for extreme performance across training, inference, graphics, and precision workloads. With world-record throughput for DeepSeek-R1 and an optimized FP4/FP8 Transformer Engine, it delivers unmatched efficiency for large-scale generative AI. While the initial system cost is high, platforms like Novita AI offer flexible, cost-effective cloud access to B200 hardware—making top-tier AI accessible without infrastructure overhead.

Frequently Asked Questions

Novita AI is an AI cloud platform that offers developers an easy way to deploy AI models using our simple API, while also providing the affordable and reliable GPU cloud for building and scaling.

Recommended Reading

• L40S vs RTX 4090: The Right GPU Depends on Your Needs
• L40S on Novita AI: A Versatile GPU for AI, Graphics, and HPC
• Llama 3.2 3B vs DeepSeek V3: Comparing Efficiency and Performance.