Forecasting 1 Million Time Series in Under 8 Minutes on a Single PC with Geneva Forecasting™ Engine

Using RoadMap Technologies’ Geneva Forecasting™ Engine API, we batch-forecasted 1 million time series at a sustained rate of 2,300 to 2,500 time series per second. That throughput far exceeds both the most widely used statistical forecasting packages and leading enterprise ERP systems, all within the confines of a single $2,000 desktop PC.

Benchmark Environment and Results

Category	Specification / Result
CPU	Intel Core i9-13900KF @ up to 5.8 GHz
RAM	64 GB DDR5 @ 4000 MT/s
Storage	2 TB Western Digital BLACK Edition Gen 4 NVMe
Operating System	Ubuntu Linux 24.04.4 LTS
Dataset	M4 Forecasting Competition Dataset (1,000,000 series)
Total Elapsed Time	7 minutes, 9 seconds (1,000,000 time series forecasts)
Sustained Throughput	~2,300 to 2,500 time series / second

The Jack of All Trades vs. The Master of Some

With the rise of large language models (LLMs), large reasoning models (LRMs), and agentic AI taking the industry by storm, there is no denying the appeal of a general-purpose model that can help a user with anything from straining pasta to summarizing tax rules. However, when it comes to applying probabilistic deep learning algorithms to structured numerical operations, unique challenges quickly surface around resource efficiency and reproducibility of results.

In practice, general-purpose LLMs and LRMs are often too broad in their training for high-volume quantitative pipelines and can produce inconsistent outputs due to the “overthinking” phenomenon, where reasoning models identify a valid solution early but burn compute cycles second-guessing themselves downstream.

Traditional, deterministic statistical algorithms, on the other hand, are mature, mathematically provable, and battle-tested across decades of industrial production, especially in one foundational discipline of business operations: forecasting.

A prime example of deterministic engineering at scale is RoadMap Technologies’ Geneva Forecasting™ Engine.

Written in C++, the Geneva Forecasting™ Engine and its API were architected from the ground up with execution speed and memory efficiency as primary design pillars. As a result, we achieved sustained speeds of 2,300 to 2,500 time series per second against the M4 Forecasting Competition dataset on a consumer-grade desktop PC. In a technology landscape where heavyweight machine learning clusters and sprawling cloud bills have become the norm, building sophisticated analytical software that runs natively at full speed on your own hardware has become something of a lost art.

Because the Geneva Forecasting™ Engine scales linearly with available CPU cores, it runs cleanly on any tier of hardware without requiring a dedicated GPU. Eliminating the GPU requirement saves organizations thousands of dollars in local workstation hardware or recurring cloud GPU instance fees while delivering reproducible, statistically rigorous forecasts. While most modern AI workloads depend on scarce and expensive GPU compute, Geneva turns standard multi-core CPUs into high-throughput forecasting engines.

The value of self-hosting on-premises software has been understated over the past decade with the shift toward cloud-only subscriptions. Running your forecasting engine on-premises or on dedicated private hardware lets your team control resource allocation directly without worrying about unpredictable cloud compute bills.

Respecting the Hardware

Have you ever noticed how a brand-new computer today can feel slower at basic tasks than an office machine from the 1990s?

Across modern software development, teams frequently trade away resource efficiency and low-level performance in favor of rapid feature shipping and heavy runtime abstractions. The result is familiar to anyone who has wondered why a simple spreadsheet viewer needs to spawn a dozen background processes just to open a file.

The Geneva Forecasting™ Engine is engineered to respect the hardware it runs on, scaling across every available CPU core with minimal memory overhead. Compared to LLMs and deep learning pipelines that demand high-wattage GPUs, CPU-native software aligns directly with real-world enterprise infrastructure, where standard servers and office workstations rarely include dedicated AI accelerators.

With ongoing volatility in memory and accelerator pricing, there is no guarantee that high-capacity RAM kits, enterprise GPUs, or storage pricing will stabilize soon. Even as new compression and quantization techniques emerge, Jevons Paradox reminds us that gains in AI model efficiency tend to increase total compute consumption rather than reduce it. The CPU market, by contrast, remains predictable, accessible, and ubiquitous.

At RoadMap Technologies, we take pride in making the most of every CPU cycle and byte of memory while delivering expert-level forecasting accuracy on your data, just as we have for nearly four decades. The release of the Geneva Forecasting™ Engine API is the next step in giving developers and planning teams that performance directly.

Ready to test the engine on your own time series? Explore the Geneva Developer Portal and get free API access at portal.roadmap-tech.com.