NVIDIA GB300 NVL72
GB300 is the NVL72 rack built from B300 GPUs and Grace CPUs. A full rack is 72 B300 GPUs and 36 Grace CPUs on one NVLink 5 domain, with about 20 TB of HBM3e. Smaller slices still sit in that architecture. It is the rack you ask for when a single 8-GPU node is not enough memory.
Implante NVIDIA GB300 NVL72 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 GB300 NVL72 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
- What it is
- NVL72 rack, not one GPU
- GPUs
- 72 × B300
- CPUs
- 36 × Grace, 72-core Neoverse V2
- GPU memory
- 288 GB HBM3e each
- Rack HBM
- About 20 TB
- Host memory
- About 17 TB LPDDR5X
- FP4 sparse
- About 1.44 EFLOPS
- FP4 dense
- About 1.08 EFLOPS
- NVLink 5
- About 130 TB/s in the rack
- GPU to CPU
- 900 GB/s NVLink-C2C
- Networking
- ConnectX-8, 800 Gb/s
- Cooling
- Liquid
- Smallest slice
- 8 GPUs
- Larger steps
- 36 GPUs, then 72
Rack, not a single GPU
GB300 is the NVL72 system: 72 B300 GPUs and 36 Grace CPUs on one NVLink 5 domain. Each GPU is a Blackwell Ultra part with 288 GB of HBM3e. The rack total is about 20 TB of HBM3e, plus on the order of 17 TB of LPDDR5X on the Grace CPUs. Sparse FP4 for the rack is quoted near 1.44 EFLOPS and dense FP4 near 1.08 EFLOPS.
NVLink 5 inside the rack is about 130 TB/s. GPU-to-CPU traffic uses NVLink-C2C at about 900 GB/s, which is the point of Grace: tool calls, retrieval, and orchestration do not cross a slow PCIe hop on every step. Networking on the reference design is ConnectX-8 at 800 Gb/s. The rack is liquid cooled.
How a rental is sliced
You do not have to take all 72 GPUs. The practical minimum is one 8-GPU node. From there the steps are a half rack (36) and a full rack (72), then more than one rack if the run needs it. A smaller slice still lives in the NVL72 fabric, so you keep the same interconnect at a lower GPU count.
A lone B300 on an x86 host is a different product. Same silicon, no Grace domain, no 20 TB of unified HBM. Use that when the model fits in one node and you want per-second billing without a rack conversation.
Who it is for
The clear full-rack jobs are models that do not fit in the roughly 2.3 TB of an 8× B300 node: trillion-parameter dense and mixture-of-experts training, frontier inference at high concurrency, and long-context serving with a large KV cache. Smaller GB300 allocations still make sense when you specifically want Blackwell Ultra and Grace, not when any 8-GPU Hopper box would do.
GB200 is the previous NVL72 layout with B200 GPUs, about 192 GB each inside the rack instead of 288 GB, and ConnectX-7 rather than ConnectX-8. If the extra memory per GPU is unused, GB200 is the rack to price first.
Price and timing
A rack slice is not a single number on a card. Quantity, region, and how long you need the fabric change the ask. On this site the charge is still that ask, counted per second, once the machines are running. There is no separate monthly GB300 plan.
Vera Rubin is the following rack, with much higher per-GPU bandwidth if the preliminary spec holds. It is not a reason to idle a job that needs Blackwell Ultra this year.
Comparação entre gerações
Rack figures follow the NVL72 reference. An 8-GPU H100 node is listed only so the scale is obvious.
| Spec | GB300 NVL72 | GB200 NVL72 | 8× H100 |
|---|---|---|---|
| GPUs | 72 × B300 | 72 × B200 | 8 × H100 |
| GPU memory | 288 GB HBM3e each | 192 GB HBM3e each | 80 GB HBM3 each |
| Rack HBM | About 20 TB | About 13 TB | 640 GB |
| CPU | 36 × Grace | 36 × Grace | x86 host |
| NVLink domain | About 130 TB/s | About 130 TB/s | Node NVLink only |
| NIC | ConnectX-8 · 800 Gb/s | ConnectX-7 · 400 Gb/s | ConnectX-7 class |
Armazenamento rápido, portas nativas e trilhos cripto
Cada instância NVIDIA GB300 NVL72 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 GB300 NVL72
Trillion-parameter training
One rack holds far more HBM than an 8-GPU node, so a very large dense or mixture-of-experts run can stay inside one NVLink domain.
Frontier FP4 inference
B300's dense FP4 rate plus the rack fabric is for high-concurrency serving of models that do not fit one HGX node.
Agent systems
Grace CPUs share a fast link with the GPUs, which helps loops that bounce between host tools and device weights.
A slice, not the whole rack
Eight GPUs is the small unit. You still want this architecture when you specifically need Blackwell Ultra and Grace, not when any 8-GPU Hopper box would do.
Quando escolher NVIDIA GB300 NVL72
01
Pick GB300 for a rack
Ask for a slice from 8 GPUs up when the model does not fit one HGX node, or when you need the Grace link.
02
Pick a lone B300 for a short job
Same GPU, x86 host, no 20 TB domain. That is the per-second path.
03
Pick GB200 when 192 GB per GPU is enough
The fabric is similar, memory per GPU is lower, and the NIC is ConnectX-7 rather than ConnectX-8.
04
Do not confuse the name with B300
B300 is the chip. GB300 is 72 of those chips with 36 Grace CPUs.
B300 versus GB300
B300 is one GPU in an 8-way node on an x86 host. GB300 is the same silicon in an NVL72 rack: 72 GPUs, 36 Grace CPUs, one NVLink domain. Rent the GPU when the job fits a node. Ask for the rack when it does not.
How much to take
The small slice is 8 GPUs. Half a rack is 36. A full rack is 72. Multi-rack jobs are quoted, not clicked from a card.
Perguntas frequentes
How is GB300 different from B300?
What is the smallest order?
How do we price a rack?
What is sparse FP4 versus dense FP4?
Why Grace?
How does it compare with eight H100s?
Is Vera Rubin the upgrade?
Outras opções de hardware
Pronto para implantar NVIDIA GB300 NVL72?
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 →