CommNet on NVIDIA SM120: 47% Lower VRAM and 41% Higher Decode Throughput with Fused Packed INT8

Hello NVIDIA community,

I am developing CommNet, a hardware-aware compiler and runtime designed to optimize pretrained LLMs for the NVIDIA hardware available to developers and enterprises.

CommNet automatically evaluates deployment alternatives under memory, quality, context-length, and performance constraints, then produces a verified, OpenAI-compatible model deployment.

The objective is:

Enable companies to run larger and faster private AI models on existing hardware, with less manual optimization and measurable quality protection.

Current results

On an NVIDIA SM120 GPU with Qwen3-0.6B:

Configuration Peak VRAM Offline decode Serving decode Quality
FP16 baseline 1.22 GB 222 tok/s 183–197 tok/s Reference
CommNet optimized INT8 0.65 GB 312 tok/s ~246 tok/s KL ≈ 0.003

Measured improvements:

  • 47% lower peak VRAM
  • 41% higher offline decode throughput
  • approximately 25–34% higher serving throughput
  • TTFT around 17–19 ms
  • preserved quality under automated verification

7B deployment result

On Qwen2.5-7B-Instruct:

Configuration Hardware Peak VRAM Decode
FP16 Split across 2×16 GB GPUs 25.3 tok/s
CommNet optimized INT8 Single 16 GB GPU 9.79 GB 28.0 tok/s

CommNet enabled the 7B model to run entirely on one 16 GB GPU while outperforming the available two-GPU FP16 deployment.

Broader quality validation on larger models is continuing, but initial outputs remain coherent and closely aligned with the reference model.

What CommNet provides

CommNet is designed to automatically:

  • adapt model execution to the target NVIDIA GPU;
  • reduce real peak VRAM;
  • select quality-safe precision strategies;
  • optimize inference and context configuration;
  • verify quality before deployment;
  • generate a deployment manifest and performance report;
  • serve the resulting model through an OpenAI-compatible API.

Developers provide constraints such as:

Available GPU memory
Required model quality
Minimum throughput
Required context length
Target hardware

CommNet searches for and validates the strongest feasible deployment.

Benefits for developers

  • Run larger models on available GPUs.
  • Reduce manual experimentation with model formats and runtime settings.
  • Avoid configurations that fit nominally but fail in real peak memory.
  • Receive measured VRAM, throughput, latency, and quality results.
  • Deploy through existing OpenAI-compatible applications.

Benefits for companies

  • Lower AI infrastructure costs.
  • Deploy stronger private models on-premises.
  • extend the useful life of existing GPU investments;
  • reduce deployment engineering time;
  • increase inference capacity per GPU;
  • standardize model optimization across different hardware environments;
  • verify quality before production rollout.

Current direction

CommNet is progressing toward:

  • automated hardware-aware optimization;
  • execution fusion and scheduling;
  • structure-aware pruning;
  • communication-aware model planning;
  • large-model and MoE deployment;
  • heterogeneous and multi-device execution.

Evaluation opportunities

We are engaging with companies, AI infrastructure providers, system integrators, and development teams facing practical LLM deployment constraints.

Relevant situations include:

  • a model does not fit the available GPU;
  • CPU offloading makes inference too slow;
  • quantized deployments still use excessive memory;
  • different runtimes produce inconsistent performance;
  • long-context workloads reduce throughput;
  • private AI requires stronger models on existing on-premises hardware.

Teams interested in evaluating CommNet can share:


Model name
GPU configuration
Current peak VRAM
Current throughput
Required context length
Quality requirements
Deployment objective

We can provide a before-and-after evaluation covering memory usage, TTFT, decode throughput, quality, and deployment feasibility.

Thank you.