NovaPod
GPUs
Melhores GPUsImplantarModelosPreços
Empresa
MCPDocs
BlackwellOfertas ao vivo do mercado disponíveis

NVIDIA B200

B200 is the Blackwell datacenter GPU most teams mean when they say Blackwell: 192 GB HBM3e, 8 TB/s, NVLink 5, and a Transformer Engine with FP4. It is the step up from H200 when you need more than 141 GB or you want FP4. A common node shape is about 30 vCPUs and 184 GB of host memory beside the 192 GB of VRAM.

Ver o mercado ao vivo →Toda a frota de hardware
Modos flexíveis de computação

Implante NVIDIA B200 do seu jeito: pods Docker ou VPS KVM root

A NovaPod deixa você escolher como usar a GPU. Um contêiner já em cache, uma máquina virtual root com IP público ou um cluster NVLink de vários nós: você paga só os segundos em que roda, em cripto.

Docker GPU PodBoot a frio < 15 s

Contêineres de IA leves e rápidos

Computação de GPU sem o peso do hipervisor. Entre direto no PyTorch, vLLM, ComfyUI ou JupyterLab.

  • Medição com precisão de 1 segundo
  • Mais de 37 modelos de framework de ML já em cache
  • Sem taxa ociosa quando a instância está parada
Implantar como pod Docker
Root KVM VPSNVMe persistente

Máquina virtual root completa

Sistema dedicado (Ubuntu 24.04 LTS / Windows GPU Server) com SSH root e sudo desde o primeiro minuto.

  • IPv4 público dedicado (todas as portas abertas)
  • Pause quando quiser; pague só US$ 0,00014/GB/h pelo disco
  • Rode Docker-in-Docker, k3s e drivers próprios
Implantar como VPS root
Multi-GPU Cluster900 GB/s NVLink

Supercomputação e treino distribuído

Junte até 8× NVIDIA B200 em um nó com InfiniBand Quantum-2 RDMA para escalar a vários nós.

  • Troca de gradiente all-reduce sem perda
  • Pronto para PyTorch FSDP e DeepSpeed Zero-3
  • GPUDirect Storage nativo e malha RDMA
Configurar cluster

Especificações de hardware

Architecture
NVIDIA Blackwell
VRAM
192 GB HBM3e
Bandwidth
8.0 TB/s
Tensor Cores
5th generation
CUDA cores
20,480
FP64
40 TFLOPS
FP32
80 TFLOPS
TF32 dense
1,125 TFLOPS
FP8 dense
4,500 TFLOPS
FP4 dense
9,000 TFLOPS
NVLink
1.8 TB/s
Reference host
30 vCPUs, 184 GB RAM
Local disk
Often a few hundred GB of NVMe, listed on the offer
8-GPU pool
About 1.5 TB
TDP
About 1,000 W
Form factor
SXM6

Blackwell in one GPU

B200 is the Blackwell datacenter GPU: 192 GB of HBM3e at 8 TB/s, NVLink 5 at 1.8 TB/s, and a second-generation Transformer Engine with FP4. Against H100 that is about 2.4 times the memory and bandwidth. Against H200 it is more memory, more bandwidth, and a numerics format Hopper does not have.

The menu host is about 30 vCPUs and 184 GB of RAM beside the 192 GB of VRAM. An 8-GPU HGX node pools about 1.5 TB, which is the box people use to serve a 405B model in FP4 with a long context.

FP4, without the brochure

FP4 is 4-bit floating point on Blackwell. On layers that tolerate it, throughput is roughly double FP8 and the weights take half the memory. The engine mixes FP4, FP8, and FP16. It is an inference tool first. Do not flip a pre-training run to FP4 because a banner said the chip supports it.

Published inference comparisons often show large jumps versus H100 at FP4. Treat those as vendor or MLPerf conditions, then measure your own model. A 13B chat model will not show a 15× gain. A 100B+ serve that was sharding across H100s might.

Where it sits in the stack

Under about 70B parameters, H200 or even H100 is usually the better dollar. Between 70B and the point where 192 GB runs out, B200 is the default Blackwell card. Above 192 GB, move to B300. The rack version of this silicon is GB200 NVL72, which is a different page from a single SXM6.

Software expectation is current PyTorch, TensorFlow, or JAX, CUDA 12.4-class libraries, and NCCL that knows Blackwell. Bring the image you already tested. The deploy flow on this site is the same as any other GPU: pick the offer, pick a template, pay per second.

O que os números significam em produção

VRAM
192 GB

HBM3e at 8 TB/s

Versus H100 memory
2.4×

192 GB versus 80 GB

Versus H200 bandwidth
~1.7×

8 TB/s versus 4.8 TB/s

FP4
Native

About 2× FP8 on layers that tolerate 4-bit

Armazenamento e ferramentas

Armazenamento rápido, portas nativas e trilhos cripto

Cada instância NVIDIA B200 na NovaPod se conecta a uma infraestrutura de alto débito feita para tirar gargalos de E/S e a espera inicial.

Armazenamento de rede compartilhado a 25 Gb/s

Monte checkpoints e conjuntos de treino em uma ou várias instâncias ao mesmo tempo. Os dados duram além de um pod e continuam acessíveis quando as instâncias param.

JupyterLab, VS Code & SSH

Conecte do terminal local com chaves SSH comuns ou abra IDEs na web em abas do navegador, sem configurar proxy reverso.

Trilhos multi-chain sem custódia

Deposite USDT, USDC, SOL, ETH ou BNB no seu endereço pessoal. A confirmação on-chain leva segundos e não há verificação de identidade obrigatória.

Cargas indicadas para NVIDIA B200

Large-model training

192 GB and 8 TB/s are for runs that stall on H100 memory or bandwidth, including mixture-of-experts jobs that want Blackwell numerics.

100B and larger fine-tunesMixture-of-experts trainingFP8 training, FP4 only with a checked recipeMultimodal towers that do not fit Hopper

405B-class serving

An 8-GPU node pools about 1.5 TB, which is the box people use for a 405B model in FP4 with a long context.

Tensor parallel across NVLinkFP4 servingHigh batch inference100B and larger, single-node

Generative models

Video, 3D, and audio models that need both the framebuffer and the Tensor Cores.

High-resolution videoLong video contextCode and agent models above 70BBatch generation

Research past Hopper

Use B200 when the experiment needs FP4 or more than 141 GB. A 13B chat model will not show a dramatic gain.

New precision recipesModels that shard across H100s only for memoryLong-context ablationsStop the pod between runs

Quando escolher NVIDIA B200

01

Pick B200 for FP4 and 192 GB

This is the default Blackwell card when H200 memory is no longer enough, or when you want 4-bit inference on a current stack.

02

Pick H100 under about 70B if 80 GB fits

H100 is mature and usually a lower ask. B200 is wasted capacity if the profile never leaves 80 GB.

03

Pick H200 for 141 GB without Blackwell

Same Hopper software as H100, more memory, no FP4. A common middle step for 70B–180B serving.

04

Pick B300 or a GB200 rack above this

B300 is 288 GB on one GPU. GB200 is the NVL72 built from B200s and Grace. Both are different pages from a single SXM6.

What fits

A 405B model can sit in FP4 on one B200. A 200B-class model in FP8 still has room for cache. An 8-GPU node pools about 1.5 TB.

FP4 in practice

Four-bit math is an inference format. It cuts memory and raises throughput when the stack supports it. Training stays on FP8 or BF16 unless you have a recipe that has been checked.

Perguntas frequentes

How much does a B200 cost here?
Whatever the live ask is when you open the marketplace. We bill that ask per second. An 8-GPU node is eight of those asks. There is no standing rate on this page, and no 20-minute minimum.
What is the cheapest way to run one?
Take the lowest ask you trust, use FP4 only if the model was exported for it, keep the batch high enough to fill 192 GB, and stop the machine between runs. A reclaimable offer is cheaper only when you can lose the machine.
Can I rent it by the hour?
You can run it for an hour. The meter is seconds. A 20-minute benchmark costs about 20 minutes.
Is B200 better than H200?
On memory and on FP4, yes: 192 GB at 8 TB/s versus 141 GB at 4.8 TB/s, and a numerics format Hopper does not have. Under about 70B, H200 is usually the better buy.
What makes it different from H100?
About 2.4× the memory and bandwidth, fifth-generation Tensor Cores, and FP4. Real speedups depend on the model. A small chat model will not look 15× faster.
What is FP4?
4-bit floating point on Blackwell. On layers that tolerate it, throughput is about double FP8 and the weights take half the memory. The engine mixes FP4, FP8, and FP16. It is an inference tool first.
Can I train a trillion-parameter model?
An 8-GPU node has about 1.5 TB of HBM. That is a serious training box, and it is not automatically a 2T-parameter run. Past one node you need a fabric and a framework that already scales: FSDP, DeepSpeed, or Megatron.
Which frameworks?
PyTorch, TensorFlow, and JAX on a CUDA 12.4-class image, plus NCCL that knows Blackwell. Bring the container you already tested. vLLM and TensorRT-LLM are the usual serving paths.
What does NVLink change?
1.8 TB/s between GPUs, so gradient sync and tensor parallel do not sit on PCIe. Eight B200s on NVLink are a different machine from eight B200s on a motherboard.
Is buying smarter?
Street prices for the GPU are in the tens of thousands of dollars, and an 8-GPU server with power and 400G networking is several times that. Rent unless utilization stays high for a year and you already run the datacenter.
Inference only?
Yes, especially at 100B and up. Under 70B, H100 or A100 usually wins on dollars per token.
Dedicated or reclaimable?
Read the offer. A reclaimable card can disappear. Use it for batch work and sweeps with checkpoints. A long training run that cannot resume belongs on a card the host will not take back.

Outras opções de hardware

NVIDIA B300

Blackwell Ultra

NVIDIA H200

Hopper

NVIDIA H100

Hopper

Pronto para implantar NVIDIA B200?

Inicie uma instância no mercado ao vivo em menos de 15 segundos. Abasteça o saldo com criptomoeda e pague por segundo, sem compromisso.

Iniciar no mercado →
NovaPod

Aluguel de GPU na nuvem para H100, A100, H200 e placas RTX. Preços ao vivo, cobrança por segundo, pagamento em cripto.

  • AICPASOCSOC 2 Type II
  • HIPAA
  • ISO27001ISO 27001

Aluguel

GPUsMelhores GPUsImplantarCompararPreçosPrimeiro aluguel

Cargas

Casos de usoServiço vLLMAjuste LoRAEstúdio ComfyUIGPU NOID

Empresa

Blog e testesSobre nósSuporte técnicoRedes cripto

Conta

EntrarDocsTermosPrivacy

© 2026 NovaPod Technologies Ltd. Todos os direitos reservados. Suporte: support@novapod.io · Os preços refletem a oferta do momento.