Physical GPU with 180GB total memory
AI Models & Assistants / Knowledge library
B200 HGX 180GB
NVIDIA accelerator; 180GB device memory.
Why learn B200 HGX 180GB?
Datacenter deployment, not a desktop upgrade. Do not interchange HGX and other Blackwell SKU specifications.
The problem it solves
Provides compute and memory for supported workloads using CUDA · TensorRT-LLM.
Where it came from
B200 HGX 180GB: NVIDIA accelerator; 180GB device memory. B denotes Blackwell; HGX is the integrated multi-GPU server platform. Datacenter deployment, not a desktop upgrade. Do not interchange HGX and other Blackwell SKU specifications.
Build your understanding
Work through these exercises in order. Try the task before opening a reference.
- Explain what B200 HGX 180GB does.
- Identify a use case for B200 HGX 180GB.
- Describe a limitation of B200 HGX 180GB.
- Name a compatible tool for B200 HGX 180GB.
- Compare B200 HGX 180GB against an alternative.
Documentation and sources
- Primary source / model card (docs.nvidia.com)
Catalog research notes
- See the AI Workbench for source dates, pricing evidence and configuration details.
Connect the ideas
Built on
Real World Example
Optimizing Genomic Sequencing Pipelines with B200 HGX 180GB
A clinical genomics research institute analyzing large-scale patient datasets for rare disease diagnostics.
The business problem
The institute faces significant latency in processing high-resolution genomic sequences because their current infrastructure lacks the device memory capacity to load entire reference genomes into a single GPU buffer, causing frequent, expensive data swaps between host memory and the accelerator.
How this topic solves it
By deploying the NVIDIA B200 HGX 180GB, the institute leverages the expansive 180GB device memory to cache massive genomic reference indices directly on the hardware, enabling parallelized processing of multiple patient samples without memory-bound bottlenecks.
- Provision the B200 HGX 180GB accelerator within the existing high-performance computing cluster.
- Configure Multi-Instance GPU (MIG) profiles to partition the 180GB memory into isolated compute instances for distinct genomic analysis tasks.
- Map the genomic reference indices into the allocated device memory buffers of the MIG instances.
- Execute parallel sequence alignment algorithms across the partitioned instances to maximize throughput.
- Monitor memory utilization per instance to ensure optimal distribution of the 180GB capacity.
This diagram illustrates how the B200 HGX 180GB memory is partitioned into distinct instances to handle parallel genomic workloads.
Dedicated memory for Patient A
Dedicated memory for Patient B
Reference data stored in device memory
- B200 HGX 180GB → partitions memory → MIG Instance 1
- B200 HGX 180GB → partitions memory → MIG Instance 2
- Genomic Index → loads into → MIG Instance 1
- Genomic Index → loads into → MIG Instance 2
Expected business outcome
The research institute achieves a streamlined workflow where genomic alignment tasks run concurrently in memory, significantly reducing the time-to-insight for patient diagnostics.
Limitations and checks
A primary limitation is that partitioning the 180GB memory via MIG reduces the maximum available memory for any single instance, which can be verified by checking the instance-specific memory allocation via the nvidia-smi command.