Neural Nova

Blog

Benchmark

Neural Nova

Unlock More Performance from Your Existing GPUs

Neural Nova accelerates training and inference while reducing cost per token across models, frameworks, and GPUs.

Explore Nova AI Platform

Talk to Our Team

Lower AI infrastructure costs without changing your GPUs.

Built for AI infrastructure companies who need more performance per GPU

gemma-4-26B-A4B-it

gpt-oss-120b

753B parameters

Qwen3-235B-A22B

Mistral-7B-Instruct-v0.3

1.93X tok/s | SGLang | H100
48% Cost ↓

1.24X tok/s | vLLM | H100

19% Cost ↓

1.24X tok/s | vLLM | H100

19% Cost ↓

1.27X tok/s | vLLM | MI355X

21% Cost ↓

1.27X tok/s | vLLM | MI355X

21% Cost ↓

2.38X tok/s | vLLM | H100

58% Cost ↓

2.38X tok/s | vLLM | H100

58% Cost ↓

72% Faster Training | H100

42% Cost ↓

72% Faster Training | H100

42% Cost ↓

Nova AI Platform

Nova AI Engine delivers higher throughput at lower GPU cost, proven on the models you already run.

Start Optimizing

How the Nova AI Platform Optimizes

An illustration from Carlos Gomes Cabral

Run Configuration

Model Configuration

Model ID

Qwen/Qwen3-4B-Instruct-2507

Hardware

h100-1x

CUDA Version

01.

Define Your Workload

Configure your AI deployment from the model, hardware, and performance targets, to establish your baseline before optimizing.

02.

Optimize the Full Stack

03.

Validate Correctness & Benchmark Performance

03.

Continuous Optimization

01.

Define Your Workload

Configure your AI deployment from the model, hardware, and performance targets, to establish your baseline before optimizing.

02.

Optimize the Full Stack

03.

Validate Correctness & Benchmark Performance

03.

Continuous Optimization

Works With Your Existing AI Stack

Connect Nova AI Platform to Hugging Face, vLLM, and SGLang to benchmark, optimize, and validate workloads from testing to production.

Optimize and validate across Hugging Face, vLLM, and SGLang

Inference

Training

Hugging Face

Portable Improvement Across Different Architectures

Portable Gains Across Architectures

Neural Nova delivers consistent gains across NVIDIA, AMD, and more incoming hardware.

Coming Soon

Proven Performance

Measurable gains across real workloads, with validated improvements in fine-tuning speed and inference throughput against your baseline.

Measurable gains on real workloads, validated against your baseline.

Fine-tuning

Mistral-7B-Instruct-v0.3

Qwen3-4B-Instruct-2507

72.60% faster · 42.1% GPU cost savings

61.09% faster · 37.9% GPU cost savings

Inference

OpenAI/GPT-oss-120b

40.6% lower GPU cost

Explore Optimized Models on HF

Explore models

Optimized Training and Inference

Powered by the Nova AI Platform, reduce GPU cost and improve performance across fine-tuning and inference without rebuilding your stack or adding specialized performance engineers.

Nova Training

Run supervised fine-tuning and LoRA jobs using optimized training configurations on customer-provided or provisioned GPU infrastructure.

Supervised fine-tuning and LoRA

Lower training time and GPU cost

Training metrics, checkpoints, and performance reporting

Dataset validation and training configuration

Plan a Training Run

Technology infrastructure and network cables

Nova Inference

Deploy optimized inference for public, private, or fine-tuned models without rebuilding your serving stack.

Optimized vLLM and SGLang deployments

Higher throughput and lower cost per request

OpenAI-compatible API configuration

Re-optimization as workloads and infrastructure evolve

Plan an Inference Deployment

Built with trusted ecosystem partners

READY FOR PRODUCTION

Improve the Performance of Your AI Workload

Benchmark, optimize, and validate faster models with measurable throughput gains before they reach users.

Talk to Our Team

Blog

Benchmark

Blog

Benchmark

Blog

Benchmark