Iris Coleman
Mar 05, 2026 18:17
NVIDIA’s GB200 NVL72 sets new STAC-AI record for LLM inference in financial trading, delivering up to 3.2x performance over Hopper architecture.
NVIDIA’s Blackwell architecture just posted the fastest-ever results on the STAC-AI benchmark for financial LLM inference, with the GB200 NVL72 delivering up to 3.2x single-GPU performance improvements over the previous-generation Hopper. The March 5, 2026 results matter for trading firms racing to extract alpha from unstructured data analysis.
The Strategic Technology Analysis Center, which has benchmarked financial technology workloads for over 15 years, tested Blackwell against real-world scenarios using EDGAR 10-K filings—the dense annual reports that quant funds parse for investment signals. Running Meta’s Llama 3.1 models, the GB200 NVL72 hit 37,480 words per second on medium-length financial prompts, compared to 8,237 WPS for dual GH200 systems.
Raw Numbers Tell the Story
On the Llama 3.1 8B model with EDGAR4 data, Blackwell processed 224 requests per second versus 51.5 RPS for Hopper—a 4.3x improvement at the system level. The gap widened on computationally heavier tasks: the 70B parameter model on long-context EDGAR5 filings saw throughput jump from 41.4 WPS to 150 WPS.
What makes these gains possible? NVIDIA’s new NVFP4 quantization format, exclusive to Blackwell, squeezes models into smaller memory footprints without sacrificing accuracy. Hopper ran FP8 quantization; the architectural leap to four-bit precision on Blackwell unlocks the throughput delta.
Interactive Performance Matters for Trading
Batch processing is one thing. Real-time trading decisions require snappy responses. Here, Blackwell maintained lower reaction times (analogous to time-to-first-token) and better interword latency even when pushed toward maximum throughput. At matched utilization levels, the GB200 NVL72 consistently beat GH200 on responsiveness metrics across most test scenarios.
For trading desks running sentiment analysis on earnings calls or parsing breaking news, that latency advantage translates directly into faster decision-making. The benchmark explicitly tested the full inference pipeline including tokenization—work that real deployments can’t skip.
Market Context
NVIDIA shares traded at $181.41 on March 5, up 1.1% on the day, with the company’s market cap sitting at $4.42 trillion. The Blackwell architecture, announced at GTC 2024, was designed specifically for generative AI workloads. CEO Jensen Huang positioned it as powering “a new industrial revolution,” and these benchmark results provide concrete evidence for that claim in the financial sector.
The GB200 Grace Blackwell superchip combines two B200 GPUs with a Grace CPU, featuring redesigned AI Tensor Cores and fifth-generation NVLink for scaling up to 576 GPUs. Previous MLPerf results showed 2.2x training gains on Llama 3.1 405B; these STAC-AI numbers confirm similar advantages extend to inference.
Hopper Still Relevant
Worth noting: the three-year-old Hopper architecture posted respectable numbers. Trading firms with existing GH200 deployments aren’t obsolete overnight. But for new builds or firms where inference speed directly impacts returns, Blackwell’s economics look compelling—NVIDIA claims up to 25x reduction in LLM inference operating costs versus prior generations.
The full STAC reports, including detailed interactive mode metrics across various arrival rates, are available through STAC’s official channels. Financial institutions evaluating AI infrastructure upgrades now have audited third-party data to inform procurement decisions.
Image source: Shutterstock
Credit: Source link
