Nvidia B200 GPU: Specs, VRAM, Price, and AI Performance

The Nvidia B200 is the flagship GPU of Nvidia's Blackwell architecture, the most capable data-center GPU available in 2025. With 180 GB of HBM3e memory, 8 TB/s of memory bandwidth, and a dual-die design housing 208 billion transistors, it delivers roughly 4x the AI compute throughput of the H100 for training workloads and significantly higher gains on inference.

Runpod offers on-demand B200 GPU instances at $4.99/hr, making it one of the first cloud platforms to provide direct access to Blackwell-generation compute. This guide covers the full B200 specs, VRAM capacity and bandwidth, how it compares to the H100 and H200, and how to access B200 instances through Runpod.

How Much VRAM Does the B200 Have?

The Nvidia B200 has 180 GB of HBM3e memory with approximately 8 TB/s of memory bandwidth. This is the highest VRAM capacity available on any single GPU and more than double the H100's 80 GB.

For context across the current GPU landscape:

• B200: 180 GB HBM3e, ~8 TB/s bandwidth
• H200 SXM: 141 GB HBM3e, ~4.8 TB/s bandwidth
• H100 SXM5: 80 GB HBM3, ~3.35 TB/s bandwidth
• A100 80GB: 80 GB HBM2e, ~2 TB/s bandwidth
• RTX 5090: 32 GB GDDR7, ~1.79 TB/s bandwidth

The 180 GB capacity enables loading and serving the largest frontier models at full precision without tensor parallelism across multiple GPUs. Models in the 70B-180B parameter range that require multi-GPU setups on H100 or H200 can fit on a single B200, simplifying deployment significantly. The 8 TB/s bandwidth, more than 2x the H200 and nearly 2.4x the H100, is particularly impactful for inference workloads where token generation speed scales directly with memory bandwidth.

Nvidia B200 Specs

The B200 is built on Nvidia's Blackwell architecture using a dual-die design (two GB100 dies connected by a 10 TB/s inter-die interconnect), fabricated on TSMC's 4NP process. Key specifications:

• Architecture: Blackwell (dual GB100 die)
• Transistors: 208 billion
• VRAM: 180 GB HBM3e
• Memory Bandwidth: ~8 TB/s
• FP4 Throughput: 20 petaFLOPS per GPU
• FP8 Throughput: 9 petaFLOPS per GPU
• TF32 Throughput: 2.25 petaFLOPS per GPU (with sparsity)
• FP64 Throughput: 40 TFLOPS
• Tensor Cores: 5th generation with FP4 and FP8 support
• NVLink: NVLink 5.0, 1.8 TB/s bidirectional bandwidth per GPU
• MIG: Supported
• TDP: 700W (SXM form factor)
• Form Factor: SXM (B200 HGX baseboard in Runpod's infrastructure)

B200 vs H100: How Do They Compare?

The B200 represents a larger generational leap than the H100 to H200 transition, which kept the same die. The B200 is a fundamentally new architecture:

• VRAM: 180 GB (B200) vs 80 GB (H100). More than double, enabling models that required 2-3 H100s to fit on a single B200.
• Memory bandwidth: 8 TB/s (B200) vs 3.35 TB/s (H100 SXM5). Nearly 2.4x improvement, delivering significantly higher inference throughput for bandwidth-bound workloads.
• AI compute: The B200's 5th-gen Tensor Cores with FP4 support deliver approximately 4x the training throughput of the H100 on transformer models. FP4 inference is roughly 3x faster than H100 FP8.
• NVLink: B200 uses NVLink 5.0 at 1.8 TB/s bidirectional per GPU, vs H100's NVLink 4.0 at 900 GB/s. Double the inter-GPU bandwidth for distributed training.
• Cost: B200 instances on Runpod are $4.99/hr vs $2.69/hr for H100 SXM. The B200 costs approximately 1.9x more per hour but delivers 3-4x the throughput on most AI workloads, giving better performance per dollar on compute-intensive jobs.

B200 vs H200: How Do They Compare?

The H200 and B200 are both next-generation data-center GPUs, but they serve different tiers:

• VRAM: 180 GB (B200) vs 141 GB (H200). The B200 has 28% more VRAM.
• Memory bandwidth: 8 TB/s (B200) vs 4.8 TB/s (H200). The B200 has 67% more bandwidth.
• Compute: The B200's dual-die design and 5th-gen Tensor Cores with FP4 deliver substantially higher compute throughput than the H200, which retains the H100's GH100 die and Hopper-generation compute.
• Architecture: H200 is a memory upgrade on Hopper (same GH100 die). B200 is an entirely new architecture (Blackwell) with new Tensor Core generation, FP4 precision, and NVLink 5.0.
• Use case: H200 is the right choice when 141 GB VRAM is sufficient and Hopper-level compute meets the requirement. B200 is justified for workloads that need maximum throughput or models exceeding 141 GB at full precision.

B200 AI Performance

The B200's performance advantages come from three compounding improvements over the H100:

• FP4 Transformer Engine: The 5th-generation Tensor Cores introduce FP4 precision with automatic mixed-precision management. For transformer inference, this delivers roughly 3x the throughput of H100's FP8 and enables real-time serving of models that previously required multi-GPU setups.
• Memory bandwidth: At 8 TB/s, the B200 is the fastest GPU available for memory-bound workloads. Token generation throughput for large LLMs scales near-linearly with bandwidth, making the B200 particularly effective for serving large context windows.
• NVLink 5.0: 1.8 TB/s bidirectional bandwidth per GPU (double H100) reduces communication overhead in distributed training, enabling better scaling efficiency across multi-GPU configurations.

In practical terms: training jobs that run for days on H100 clusters can be completed in significantly less time on B200. Inference workloads that required 2-4 H100s for latency reasons can often run on a single B200 with better response times.

B200 Price on Runpod

Runpod offers B200 instances at the following rates:

• On-demand: $4.99/hr (180 GB VRAM, 283 GB RAM, 28 vCPUs)
• 6-month commit: $4.34/hr
• 1-year commit: $4.24/hr

B200 instances are currently available through Runpod's Secure Cloud. For large-scale or dedicated cluster deployments, Runpod also offers enterprise configurations through the sales team. See current availability and pricing on the Runpod pricing page.

For context on cost-effectiveness: at $4.99/hr on-demand, a B200 instance delivers approximately 3-4x the training throughput of an H100 SXM instance at $2.69/hr, making the B200 significantly more efficient per training FLOP for jobs that can saturate it.

When to Use a B200 vs H100 or H200

The B200 is the right choice when:

• Your model exceeds 141 GB at full precision (requires 180 GB VRAM)
• Memory bandwidth is your primary bottleneck and 8 TB/s would meaningfully accelerate inference
• You need maximum single-GPU training throughput for large transformer models
• You are serving large-batch inference at production scale where FP4 throughput matters

The H100 or H200 remains the better choice when:

• Your workload fits within 80 GB (H100) or 141 GB (H200) and cost efficiency is the priority
• You need H100-level performance at a lower per-hour rate
• Your software stack is not yet optimized for FP4 or Blackwell-specific features

All three are available on Runpod, allowing you to match GPU to workload without hardware lock-in.

Rent a B200 on Runpod

Runpod provides on-demand access to B200 GPU instances without procurement lead times, enterprise contracts, or infrastructure investment. You get full root access to a containerized GPU environment and can be running within minutes.

• On-demand instances: $4.99/hr with per-second billing, no minimum commitment
• Committed pricing: $4.34/hr (6-month) or $4.24/hr (1-year) for sustained workloads
• Serverless: Deploy models on B200 workers and pay per inference request with no idle GPU cost
• Instant Clusters: Deploy multi-node B200 HGX clusters for distributed training workloads
• Network Volumes: Attach persistent storage to keep model weights, datasets, and checkpoints available across sessions. Standard storage starts at $0.07/GB/mo (first TB, then $0.05/GB/mo), with high-performance storage at $0.14/GB/mo for maximum throughput on demanding AI and data pipelines. A single volume can be shared across multiple B200 instances simultaneously, making it easy to pre-load large models before launching compute.
• Templates: Pre-built environments for PyTorch, vLLM, and other frameworks with no manual CUDA setup required
• Flexibility: Switch between B200, H200, H100, and other GPU types based on workload requirements

See current B200 availability and pricing on the Runpod pricing page.

B200 FAQs

How much VRAM does the B200 have?

The Nvidia B200 has 180 GB of HBM3e memory with approximately 8 TB/s of memory bandwidth. This is the highest VRAM capacity available on any single GPU, more than double the H100 SXM5's 80 GB.

What is the Nvidia B200?

The Nvidia B200 is a data-center GPU built on the Blackwell architecture, featuring a dual-die design with 208 billion transistors. It is designed for large-scale AI training and inference, offering 180 GB HBM3e VRAM, 8 TB/s memory bandwidth, 5th-generation Tensor Cores with FP4 support, and NVLink 5.0 at 1.8 TB/s bidirectional bandwidth per GPU.

How much does the B200 cost on Runpod?

Runpod offers B200 instances at $4.99/hr on-demand, $4.34/hr on a 6-month commit, and $4.24/hr on a 1-year commit. Each instance includes 180 GB VRAM, 283 GB RAM, and 28 vCPUs. Per-second billing means you only pay for exactly what you use.

B200 vs H100: which is better for AI?

The B200 delivers approximately 3-4x better training throughput than the H100 on transformer workloads, with 2.2x higher memory bandwidth and more than double the VRAM. It costs roughly 1.9x more per hour on Runpod but is significantly more efficient per compute FLOP for large workloads. For smaller models that fit within 80 GB and don't require maximum throughput, the H100 remains the better value.

What is the difference between the B200 and DGX B200?

The DGX B200 is Nvidia's branded complete server system containing eight B200 GPUs, NVSwitch fabric, dual Xeon CPUs, 4 TB of RAM, and enterprise support in a single 10U chassis. It is sold as an on-premises enterprise appliance at roughly $300,000-$500,000. Runpod offers individual B200 GPU instances from B200 HGX infrastructure at $4.99/hr, giving developers and researchers access to the same Blackwell GPU compute without purchasing a complete DGX system.

Is the B200 available on Runpod?

Yes. Runpod was among the first cloud platforms to offer B200 GPU instances, available through Secure Cloud. Runpod also supports multi-node B200 HGX cluster deployments via Instant Clusters for distributed training workloads requiring multiple GPUs. See the pricing page for current availability.