Support Matrix | NVIDIA NIM Harmonizer

Harmonizer NIM uses one GPU per container.

Supported GPUs

All supported GPUs serve the model in bfloat16 (BF16) precision. Each GPU builds a set of fixed-shape optimized executor states at startup, sized to the GPU’s memory.

GPU	Precision	Notes
H100 80 GB HBM3	`bf16`	Primary validation target.
H100 PCIe	`bf16`	Full startup specialization set.
H100 NVL	`bf16`	Full startup specialization set.
A100 80 GB	`bf16`	Full startup specialization set.
H200	`bf16`	Full startup specialization set.
H200 NVL	`bf16`	Full startup specialization set.
B200	`bf16`	Blackwell coverage.
L40S 48 GB	`bf16`	Reduced startup specialization set to fit 48 GB.
RTX 6000 Ada	`bf16`	Reduced startup specialization set to fit 48 GB.
RTX 4090	`bf16`	Minimal startup specialization set to fit 24 GB.

Profiles

The model manifest ships latency and throughput profiles for every supported GPU. These are selector aliases and benchmark-scenario labels. Both aliases use the same BF16 checkpoint, model artifact, and runtime settings; they are not separately tuned configurations.

Profile	Precision	Scenario	Supported GPUs
`latency`	BF16	Alias used for online/concurrency-1 benchmark selection.	All supported GPUs
`throughput`	BF16	Alias used for high-concurrency benchmark selection.	All supported GPUs

Latency Notes

End-to-end latency includes model execution, image encoding and decoding, and transport. The non-model stages significantly affect total latency, but work is overlapped in the serving pipeline and they do not consistently dominate model execution. The shipped fast PNG path and raw uint8 tensor path have similar end-to-end performance. Choose between them based on the client integration, and prefer gRPC on port 50051 for latency-sensitive traffic to avoid the HTTP base64 and JSON adapter path.

Multi-GPU deployment is not used for Harmonizer. The recommended deployment shape is one NIM container per GPU.