IBM's Granite-TTM-R3 Runs 18,000 Forecasts per Second Without a GPU
IBM's third-generation TinyTimeMixer brings mixture-of-experts routing, trend-residual decomposition, and roughly 15 to 50 times faster forecasting than popular SOTA baselines.
- IBM released Granite-TTM-R3, the third generation of its TinyTimeMixer time-series foundation models.
- Mixture-of-experts routing across 1 to 35M parameter models with linear gating attention instead of self-attention.
- Roughly 7,500 samples/sec on GPU and 180 on CPU, with Lite hitting 18,000 and 800 respectively.
- GIFT-Eval scores of MASE 0.718 and CRPS 0.514 after fine-tuning, competitive with larger foundation models.
- Adds trend-residual decomposition, FFT embeddings, student-teacher pretraining and a multi-quantile probabilistic head.
- Apache-2.0 licensed, available through the granite-tsfm repo for zero-shot, few-shot and full fine-tune workflows.
IBM Research has released Granite-TTM-R3, the third generation of its TinyTimeMixer time-series models. Published under the Apache-2.0 license, R3 adds expert routing, trend-residual decomposition, probabilistic forecasts, and an architecture designed for high-throughput inference on CPUs and GPUs.
R3 matters for teams that forecast thousands of products, sensors, or energy meters. IBM reports hundreds of forecast samples per second on CPU, which could reduce infrastructure costs and simplify deployments that currently depend on GPUs. The paper and full ablation studies remain pending, so the published results should be treated as vendor-reported benchmarks.
Small experts, routed per series
TinyTimeMixer models are compact pretrained networks for multivariate time-series forecasting. R3 retains parameter counts between roughly 1 million and 35 million, while Lite configurations range from 1 million to 18 million parameters.
R3 organizes those models as experts of different sizes. A lightweight router selects or blends experts according to the input series, allowing the system to match model capacity to the data. The backbone uses linear gated attention to capture temporal relationships and contains no self-attention blocks, reducing memory use and computation.
IBM also introduced several changes to training and forecast generation:
- Trend-residual decomposition: The model separates slow-moving trends from higher-frequency residuals, learns each component, and then trains them jointly.
- Three-stage pretraining: Trend, residual, and joint training stages are intended to improve convergence stability.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
