---
格式版本: 2
标题: "NVIDIA GB200 NVL72: Grace Blackwell Specs & Rental | Spheron"
原文链接: "https://www.spheron.network/gpu-rental/gb200/"
发布日期: "未标注"
发现时间: "2026-06-01T16:19:30+08:00"
入库时间: "2026-06-01T14:52:56.811Z"
来源平台: "SerpApi Google"
搜索渠道: "serpapi"
搜索词: "NVL72"
匹配关键词:
  - "NVL72"
  - "GPU"
  - "NVIDIA"
  - "Nvlink"
  - "Liquid Cooling"
  - "HBM"
相关厂家:
  - "NVIDIA"
相关专家:
  []
内容类型: "网页"
抓取工具: "XCrawl Scrape"
清洗工具: "XCrawl Markdown + LLM 正文裁剪"
原始附件:
  []
AI优质: "否"
AI打分: 74
AI分档: "召回候选"
AI质检状态: "不通过"
AI打分理由: "内容聚焦GB200 NVL72机柜级架构与规格，技术细节完整，但来源为第三方算力租赁平台，非原厂或权威研究机构；商业信号侧重租赁报价与预约，缺乏供应链量产、客户落地或生态合作等一手部署信息，综合评为召回候选。"
AI质检模型: "qwen3.6-plus"
AI质检时间: "2026-06-02T23:15:59+08:00"
AI主题相关性: 18
AI来源权威性: 7
AI新颖性: 15
AI技术细节: 16
AI商业部署信号: 9
AI完整性: 9
采集批次: "2026年6月1日16点18分53秒"
采集批次ID: "20260601-161853-741"
去重键: "https://www.spheron.network/gpu-rental/gb200"
---

# NVIDIA GB200 NVL72: Grace Blackwell Specs & Rental

# NVIDIA GB200 NVL72 Grace Blackwell Rack-Scale 72x B200 · 13.5 TB HBM3e · 1.4 EFLOPS FP4

NVIDIA's Blackwell rack-scale system. 72 B200 GPUs and 36 Grace CPUs share one NVLink 5 domain with 13.5 TB of HBM3e and 1.4 EFLOPS of FP4. Built for trillion-parameter training and 100B+ inference where rack-scale memory and NVLink fabric decide the run. Reserve any count on Spheron, from a single 8-GPU node to multi-rack. Tell us how many GPUs you need. We confirm availability within one business day. 

| | | | |
|---|---|---|---|
| Per GPU | 192 GB HBM3e | | |
| Per Rack | 1.44 EFLOPS FP4 | | |
| NVLink | 130 TB/s | | |
| Starts at | 8 GPUs | | |

[GB200 architecture guide →](https://www.spheron.network/blog/nvidia-gb200-nvl72-guide/)
[GB300 Blackwell Ultra →](https://www.spheron.network/gpu-rental/gb300/)
[R100 (Rubin) pre-order →](https://www.spheron.network/gpu-rental/r100/)
[B200 on-demand rental →](https://www.spheron.network/gpu-rental/b200/)

Other GPU options
| | | |
|---|---|---|
| B200 | Per GPU | From$2.14/hr |
| B300 | Per GPU | From$2.63/hr |
| H200 | Per GPU | From$1.40/hr |

[View all GPU pricing →](https://www.spheron.network/pricing/)

## Where GB200 sits in the stack

GB200 is the broadly-deployed Blackwell rack-scale system. GB300 is the Blackwell Ultra upgrade with 50% more memory. Rubin R100 is the next generation, available H2 2026.

| | | | |
|---|---|---|---|
| Hopper H100 | 80 GB | Available | |
| Hopper H200 | 141 GB | Available | |
| Blackwell | Reserve GB200 192 GB HBM3e | This page | |
| Blackwell Ultra B300 | 288 GB | Available | |
| Blackwell Ultra GB300 | 288 GB HBM3e | Available | |
| Rubin H2 2026 R100 | 288 GB HBM4 | Pre-Order | |

## GB200 NVL72 specifications

| | |
|---|---|
| 01 GPUs per rack | 72 × B200 Blackwell |
| 02 Grace CPUs | 36 × 72-core Neoverse V2 |
| 03 Total HBM3e | 13.5 TB unified |
| 04 System Memory | 17 TB LPDDR5X |
| 05 FP4 Compute | 1.44 EFLOPS |
| 06 FP8 Compute | 720 PFLOPS |
| 07 NVLink 5 Fabric | 130 TB/s aggregate |
| 08 GPU-to-CPU | 900 GB/s NVLink-C2C |
| 09 Networking | ConnectX-7 · 400 Gb/s |
| 10 Rack power | ~120 kW liquid-cooled |

Per-rack specifications. Smaller reservations inherit the same architecture at the slice they're allocated. All systems run in NVL72 reference configuration with liquid cooling, 2:1 InfiniBand fat-tree fabric across racks, and persistent NVMe per chassis.

## GB200 NVL72 vs GB300 NVL72 vs HGX H100 Spec (per rack)

| Spec (per rack) | GB200 NVL72 Reserve | GB300 NVL72 | HGX H100 (8x) |
|---|---|---|---|
| Architecture | Blackwell | Blackwell Ultra | Hopper |
| GPUs / rack | 72 × B200 | 72 × B300 | 8 × H100 |
| Total HBM | 13.5 TB HBM3e | 20.7 TB HBM3e | 640 GB HBM3 |
| FP4 sparse | 1.44 EFLOPS | 2.16 EFLOPS | N/A |
| FP4 dense per GPU | 10 PFLOPS | 15 PFLOPS | N/A |
| FP8 throughput | 720 PFLOPS | 720 PFLOPS | 16 PFLOPS |
| NVLink fabric | 130 TB/s | 130 TB/s | 7.2 TB/s (8 GPUs) |
| CPU | 36 × Grace | 36 × Grace | x86 host |
| Networking | ConnectX-7 · 400 Gb/s | ConnectX-8 · 800 Gb/s | ConnectX-7 · 400 Gb/s |
| Spheron access | By reservation | By reservation | Available |

Per-rack specifications for NVL72 systems. HGX H100 figures are for a standard 8-GPU node for relative scale, since H100 does not come in NVL72 configuration. GB200 specs match the NVIDIA GB200 NVL72 reference design.

## Workloads built for GB200

### Large-Scale LLM Pre-Training

Each B200 GPU has 192 GB HBM3e. A single 8-GPU node holds 1.5 TB; a full NVL72 rack holds 13.5 TB inside one NVLink domain. 1.44 EFLOPS FP4 per rack is a 4x lift over equivalent H100 capacity for 200B to 1T parameter pre-training. Reserve the count that fits your run.

- 200B to 1T dense and MoE pre-training
- Multi-modal foundation models
- RLHF and post-training at scale
- Long-sequence transformer training

### High-Throughput FP4 Inference

B200 doubles FP4 throughput over H100 and delivers 8 TB/s HBM bandwidth per GPU. 130 TB/s of NVLink fabric inside an NVL72 rack supports tensor and pipeline parallelism for large models without leaving the rack. Smaller reservations fit 70B to 200B parameter inference on 8 to 32 GPUs.

- MoE serving at high concurrency
- Long context windows (256K to 1M tokens)
- Disaggregated prefill and decode
- Real-time agentic inference for 200B+ models

### Reasoning and Agentic Workloads

Grace CPUs handle orchestration, retrieval, and tool calls in the same coherent address space as the GPUs over 900 GB/s NVLink-C2C, removing PCIe round-trips that slow down agent loops on x86 nodes. Reserve from a few GPUs for prototyping up to full racks for production serving.

- Test-time compute and chain-of-thought serving
- Tool-using agents with shared memory
- Code generation with 200K+ context
- Multi-step planning and self-reflection

### Fine-Tuning and Post-Training

Reserve the GPU count that matches your run. DPO and GRPO on 70B models fit on a single 8-GPU node. Large-scale RL with rollouts, value heads, and reference models benefits from full-rack co-location inside one NVLink domain. Persistent NVMe per chassis handles checkpoint streaming at any scale.

- DPO, GRPO, and PPO fine-tuning
- Reward modeling for production LLMs
- Distillation pipelines into smaller models
- Curriculum learning with millions of rollouts

## When to pick GB200

### Pick GB200 if

You want Blackwell rack-scale at the best price-to-performance. Reserve any quantity, from a single 8-GPU node to multi-rack deployments. Models that need rack-scale memory and fit in 13.5 TB of HBM3e (most workloads under 1T parameters) get the same NVLink 5 fabric and Grace CPU coupling as GB300 at lower cost.

### Pick GB300 instead if

Your workload needs 288 GB per GPU or is bottlenecked on attention compute. GB300 has 50% more HBM3e per GPU and 2x attention throughput over GB200, with ConnectX-8 (800 Gb/s) replacing ConnectX-7 (400 Gb/s). For 1T+ parameter training and reasoning workloads, GB300 is the upgrade.

### Pick B200 (per GPU) instead if

You want per-minute billing with no commitment. B200 SXM5 is available on Spheron on a per-GPU or 8-GPU node basis with on-demand and spot pricing. For most 70B to 200B parameter inference and fine-tuning jobs that don't need a sales conversation, B200 is the simpler call.

### Pick R100 if you can wait

R100 Rubin is available H2 2026 with 22 TB/s per-GPU bandwidth (2.75x B200) and 50 PFLOPS FP4 (5x B200 dense). For new training runs with flexible timelines, R100 is the higher-ceiling option. If you need to start now, GB200 is live today.

## Other GPUs on Spheron

For workloads that don't need GB200 specifically, Spheron also offers B200, B300, and H200 on per-minute billing with no commitments.

| | | | |
|---|---|---|---|
| B200 | 192 GB HBM3e | Same Blackwell silicon as GB200, billed by the minute. | From$2.14/hr live | Rent B200 → |
| B300 | 288 GB HBM3e | Blackwell Ultra. 50% more memory per GPU than B200. | From$2.63/hr live | Rent B300 → |
| H200 | 141 GB HBM3e | Hopper. Best value for inference and long-context serving. | From$1.40/hr live | Rent H200 → |

## GB200 FAQ

### How do I reserve GB200 NVL72 capacity on Spheron?

Submit the form on this page with your expected GPU count, timeline, and intended workload. Our team confirms availability, region, and pricing within one business day. Reservations are allocated per request across Tier 3 and Tier 4 liquid-cooled data center partner regions.

### What's the minimum and maximum quantity I can reserve?

Reserve any count. The smallest configuration is a single 8-GPU node. From there you can scale to 36 GPUs (half rack), 72 GPUs (full NVL72 rack), or multi-rack deployments in the hundreds. Smaller allocations still sit inside an NVL72 NVLink domain, so you keep the same interconnect at any scale.

### What's the difference between GB200 and B200?

B200 is the individual GPU: 192 GB HBM3e, 8 TB/s bandwidth, 10 PFLOPS dense FP4. It runs in standard 8-GPU SXM5 nodes connected to an x86 host. GB200 is the same B200 silicon deployed in NVL72 racks that pair 72 GPUs with 36 Grace CPUs over NVLink 5, exposing 13.5 TB of unified memory inside one NVLink domain. B200 is rented per GPU on per-minute billing. GB200 is reserved by request in any quantity and our team provisions.

### How is GB200 priced?

Pricing depends on quantity, commitment length, region, and networking requirements. Smaller reservations run closer to per-hour rates; multi-rack reservations come with committed pricing and priority allocation. Submit the form with your GPU count and our team shares a quote within one business day.

### How does GB200 compare to GB300?

GB300 swaps B200 GPUs (192 GB HBM3e) for B300 GPUs (288 GB HBM3e), raising total rack memory from 13.5 TB to 20.7 TB. Per-GPU dense FP4 increases from 10 to 15 PFLOPS and B300 adds 2x attention compute. Networking upgrades from ConnectX-7 (400 Gb/s) to ConnectX-8 (800 Gb/s) per GPU. NVLink 5 fabric, Grace CPU layout, and software stack remain consistent, so workloads port directly between the two systems.

### What workloads is GB200 built for?

Models in the 200B to 1T parameter range that benefit from rack-scale memory and NVLink interconnect. 1.44 EFLOPS FP4 per rack supports trillion-parameter dense and MoE training, frontier inference at high concurrency, long-context serving, and large-scale RL post-training. Smaller GB200 reservations also work for fine-tuning and inference on the latest Blackwell architecture at lower cost than GB300.

### How fast can I get provisioned after submitting the form?

Our team confirms availability within one business day. From there, smaller reservations (single 8-GPU nodes, half racks) typically come online within days. Multi-rack deployments take longer depending on region and networking requirements. We share the timeline along with the quote.

## Also consider

[GB300 288GB Blackwell Ultra · 288 GB HBM3e · 8 TB/s](https://www.spheron.network/gpu-rental/gb300/)
[B200 192GB Single-GPU Blackwell. Per-minute billing, no commit.](https://www.spheron.network/gpu-rental/b200/)
[R100 288GB HBM4 Rubin pre-order. Next-gen architecture, H2 2026.](https://www.spheron.network/gpu-rental/r100/)

More [Explore Browse all GPUs View catalog](https://www.spheron.network/gpu-rental/)
[Compare View pricing comparison See all rates](https://www.spheron.network/pricing/)