NVIDIA B300
B300 is the Blackwell Ultra GPU: 288 GB of HBM3e at 8 TB/s, fifth-generation Tensor Cores, and NVLink 5 at 1.8 TB/s. The extra 96 GB over B200 is what matters when a 200B-class model or a very long context does not fit in 192 GB. A typical node quote pairs the GPU with about 32 vCPUs and 184 GB of host RAM.
Implante NVIDIA B300 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.
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
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
Supercomputação e treino distribuído
Junte até 8× NVIDIA B300 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
Especificações de hardware
- Architecture
- NVIDIA Blackwell Ultra
- VRAM
- 288 GB HBM3e
- Bandwidth
- 8.0 TB/s
- Tensor Cores
- 5th generation
- CUDA cores
- 20,480 and up
- FP4 dense
- About 15 PFLOPS
- FP8 dense
- About 7,500 TFLOPS
- FP32
- About 120 TFLOPS
- FP64
- About 60 TFLOPS
- NVLink
- 1.8 TB/s
- Reference host
- 32 vCPUs, 184 GB RAM
- Local disk
- On the order of 250 GB NVMe, varies by offer
- 8-GPU pool
- About 2.3 TB HBM3e
- TDP
- About 1,200 W
- Form factor
- SXM6
The GPU itself
B300 is the Blackwell Ultra chip: 288 GB of HBM3e at 8 TB/s, fifth-generation Tensor Cores, NVLink 5 at 1.8 TB/s, and a board power around 1,200 W. CUDA core counts published for the part sit above 20,000. Dense FP4 is about 15 PFLOPS and dense FP8 about 7,500 TFLOPS. FP64 is about 60 TFLOPS and FP32 about 120 TFLOPS, which tells you this is an AI card, not a double-precision HPC card.
The menu row assumes a single-GPU host of about 32 vCPUs and 184 GB of RAM, with on the order of 250 GB of local NVMe. An 8-GPU HGX node pools about 2.3 TB of HBM3e. That node is the unit most training jobs actually buy.
What 288 GB changes
Bandwidth matches B200 at 8 TB/s. The reason to pay for B300 is capacity: 50 percent more memory than 192 GB. A 200B-class dense model in FP8 can sit on one GPU. A trillion-parameter mixture-of-experts can be planned across one 8-GPU node instead of two nodes that exist only to hold weights. Context windows past several hundred thousand tokens stop evicting the KV cache as early.
If the model and the cache already fit in 192 GB, the extra 96 GB is idle silicon. Take B200. If you do not need FP4 at all, H200 at 141 GB is the Hopper card with a lower ask.
How people run it
Training containers are the current NVIDIA PyTorch and NeMo images with Transformer Engine new enough for FP8, and FP4 only if you have checked that recipe. Serving is vLLM or TensorRT-LLM with the quantization the model was exported for. Eight-way jobs use tensor and pipeline parallel inside the node before they touch the network.
NVLink 5 at 1.8 TB/s is the in-node fabric: full NVSwitch across eight GPUs, GPU-to-GPU traffic that does not bounce through the CPU, and the usual NCCL collectives. That is what makes an 8-GPU box behave like one machine. Past eight GPUs you are on the GB300 rack or on InfiniBand between nodes.
Billing
This marketplace does not publish a standing B300 rate. When a host lists one, the card shows the hourly ask and we bill that ask per second. A dedicated offer stays until you stop it. A reclaimable offer is only for work you can checkpoint.
Buying the card is a different problem. List prices for a single B300 are in the tens of thousands of dollars, and an 8-way node with networking and cooling is an order of magnitude more. Renting wins unless the GPU would be busy most of the year and you already operate the rack.
O que os números significam em produção
- Memory vs H100
- 3.6×
- Memory vs B200
- +96 GB
- FP4 dense
- ~15 PFLOPS
- 8-GPU pool
- ~2.3 TB
288 GB versus 80 GB
288 GB versus 192 GB, bandwidth still 8 TB/s
Second-generation Transformer Engine
Eight B300s on one NVLink node
Armazenamento rápido, portas nativas e trilhos cripto
Cada instância NVIDIA B300 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 B300
Frontier training
288 GB per GPU lets a larger dense or mixture-of-experts model sit on fewer cards than a 192 GB plan.
Large-model serving
FP4 plus the extra 96 GB cuts tensor-parallel splits for models that spill off a B200.
Research that is memory-bound
Use it when the profile says the run died on capacity. If a cheaper Hopper card already fits, this GPU is idle silicon.
Inside one NVLink node
Eight B300s with NVLink 5 behave like one machine. Past eight GPUs you are on the GB300 rack or on a network between nodes.
Quando escolher NVIDIA B300
01
Pick B300 when 192 GB is the wall
One B300 can replace two B200s that existed only to hold weights.
02
Pick B200 when the model fits
Bandwidth stays near 8 TB/s. You would be paying for 96 GB you do not touch.
03
Pick H200 when you do not need FP4
141 GB of HBM3e on Hopper is the cheaper long-context card for jobs that already fit.
04
Pick GB300 when eight GPUs are not enough
The rack is the same silicon with Grace and a shared NVLink domain.
When 288 GB changes the job
Use B300 when weights, optimizer state, or a 1M-token KV cache do not fit in 192 GB. If they fit, B200 is the cheaper Blackwell card. The bandwidth does not jump versus B200; the capacity does.
Software
Current PyTorch, TensorFlow, and JAX containers with CUDA 12.5-class libraries, cuDNN, and TensorRT are what you want. FP4 needs a recent Transformer Engine. Bring that image on deploy.
Perguntas frequentes
What changed versus B200?
Does 288 GB change inference?
Is one B300 cheaper than two B200s?
Can I use it for inference only?
What host does the menu assume?
Which software?
What is NVLink 5 doing?
Should I buy the card?
How is it billed here?
Outras opções de hardware
Pronto para implantar NVIDIA B300?
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 →