System Requirements for TimesFM 2.5: Hardware Tiers and Memory Configuration

TimesFM 2.5 supports four hardware tiers ranging from 4 GB CPU‑only environments to production‑grade GPUs with 16 GB+ VRAM, with total memory requirements calculated as approximately 800 MB for model weights plus 0.5 GB overhead and 0.2 MB per 1,000 series per context point.

TimesFM 2.5 is a 200 million‑parameter time‑series forecasting model developed by Google Research. Understanding the system requirements for TimesFM 2.5 is essential before deployment, as the model’s memory footprint scales linearly with context length and batch size. The google-research/timesfm repository provides a pre‑flight validation tool and detailed hardware tier specifications in timesfm-forecasting/references/system_requirements.md to help you match your infrastructure to your forecasting workload.

Hardware Tiers for TimesFM 2.5

The repository defines four distinct hardware tiers in the official requirements file. Each tier specifies RAM or VRAM minimums, recommended batch sizes (per_core_batch_size), and maximum context lengths (max_context).

Tier 1 – Minimal (CPU‑only)

Target Hardware: CPU with 4 – 8 GB RAM.
Configuration: Set per_core_batch_size=4 and max_context=512.
Performance: Processes 100‑point series in 2 – 5 seconds.
Memory Profile: Requires approximately 100 MB per 1,000 series for 512‑point contexts according to the memory formula in system_requirements.md.

Tier 2 – Standard (CPU or GPU)

Target Hardware: CPU with ≥ 16 GB RAM or GPU with 4 – 8 GB VRAM.
Configuration: Use per_core_batch_size=32 on CPU or per_core_batch_size=64 on GPU, with max_context=1024.
Performance: GPU inference typically completes in 0.5 – 1 second for 100‑point series.
Memory Profile: 1,024‑point contexts require ~200 MB per 1,000 series.

Tier 3 – Production (High‑End GPU)

Target Hardware: GPU with ≥ 16 GB VRAM or Apple Silicon with ≥ 32 GB unified memory.
Configuration: Set per_core_batch_size between 128 and 256, with max_context=4096 or higher.
Performance: Typical inference speed of 0.1 – 0.3 seconds per 100‑point series.
Memory Profile: Contexts of 4,096 points require substantial VRAM; the absolute maximum of 16,384 points demands significant memory resources.

Tier 4 – Legacy Models (TimesFM 2.0)

Target Hardware: CPU with ≥ 16 GB RAM or GPU with ≥ 8 GB VRAM.
Configuration: Applies to the legacy 500 M‑parameter TimesFM 2.0 model.
Notes: This tier supports the earlier model architecture with higher baseline memory requirements than the 200 M‑parameter TimesFM 2.5.

Memory Calculation Formula

The canonical memory estimation formula is documented in timesfm-forecasting/references/system_requirements.md (lines 12‑20):

RAM ≈ model_weights + 0.5 GB + (0.2 MB × num_series × context_length / 1000)

• model_weights: Approximately 800 MB for TimesFM 2.5.
• context_length: Defined by the max_context parameter in your ForecastConfig.
• num_series: Number of time‑series processed in a single batch.

Larger contexts are feasible but require proportionally more memory. For example, 2,048‑point contexts typically require ~16 GB RAM, while 4,096‑point contexts need a GPU or ≥ 32 GB RAM.

Pre‑Flight System Validation

Before executing training or inference, run the pre‑flight system checker located at timesfm-forecasting/scripts/check_system.py. This script evaluates your machine’s RAM and VRAM against the selected tier and prints warnings if the configuration exceeds available resources, preventing out‑of‑memory crashes.

python -m timesfm-forecasting.scripts.check_system \
    --max_context 1024 \
    --per_core_batch_size 64

The checker reads hardware specifications and confirms whether your requested max_context and batch size fit within the limits defined for your hardware tier. For a complete example of tier‑specific configuration in practice, see timesfm-forecasting/scripts/forecast_csv.py.

Configuration Examples by Hardware Tier

The following ForecastConfig implementations correspond to the hardware tiers defined in the system requirements file and are implemented in the core Torch backend at src/timesfm/timesfm_2p5/timesfm_2p5_torch.py.

Minimal Tier Configuration

For Tier 1 (CPU‑only, 4 – 8 GB RAM) as specified in lines 35‑41 of the requirements file:

import timesfm
model = timesfm.ForecastConfig(
    max_context=512,
    max_horizon=128,
    per_core_batch_size=4,
    normalize_inputs=True,
    use_continuous_quantile_head=True,
    fix_quantile_crossing=True,
)
model.compile()

Standard Tier GPU Configuration

For Tier 2 (GPU with 4 – 8 GB VRAM) as specified in lines 54‑61:

import timesfm
model = timesfm.ForecastConfig(
    max_context=1024,
    max_horizon=256,
    per_core_batch_size=64,   # GPU batch size

    normalize_inputs=True,
    use_continuous_quantile_head=True,
    fix_quantile_crossing=True,
)
model.compile()

Production Tier Configuration

For Tier 3 (GPU with ≥ 16 GB VRAM) as specified in lines 73‑80:

import timesfm
model = timesfm.ForecastConfig(
    max_context=4096,
    max_horizon=256,
    per_core_batch_size=128,  # or 256 for very large batches

    normalize_inputs=True,
    use_continuous_quantile_head=True,
    fix_quantile_crossing=True,
)
model.compile()

Summary

• TimesFM 2.5 defines four hardware tiers from minimal 4 GB CPU setups to production 16 GB+ GPU environments.
• Memory requirements scale according to the formula RAM ≈ 800 MB + 0.5 GB + (0.2 MB × num_series × context_length / 1000).
• Context limits vary by tier: 512 points (Minimal), 1,024 points (Standard), and up to 4,096+ points (Production).
• Validation tool: Use timesfm-forecasting/scripts/check_system.py to verify your hardware against configuration parameters before running inference.
• Reference documentation: Canonical specifications live in timesfm-forecasting/references/system_requirements.md.

Frequently Asked Questions

Can TimesFM 2.5 run on a CPU without a GPU?

Yes. The Minimal tier supports CPU‑only operation with 4 – 8 GB RAM using max_context=512 and per_core_batch_size=4. The Standard tier also supports CPU operation with ≥ 16 GB RAM, though GPU acceleration significantly improves inference speed.

How much RAM is required for a context length of 4096 points?

A context length of 4,096 points requires the Production tier hardware: either a GPU with ≥ 16 GB VRAM or Apple Silicon with ≥ 32 GB unified memory. According to the memory formula, this context length demands substantial memory allocation proportional to the batch size.

What is the per_core_batch_size parameter and how does it affect memory?

The per_core_batch_size parameter controls how many time‑series are processed simultaneously per core. Higher values increase throughput but linearly increase RAM consumption based on the formula 0.2 MB × num_series × context_length / 1000. The Minimal tier uses 4, Standard uses 32–64, and Production uses 128–256.

Where can I find the official system requirements documentation?

The canonical hardware specifications and memory formulas are documented in timesfm-forecasting/references/system_requirements.md within the google-research/timesfm repository. The companion validation script is located at timesfm-forecasting/scripts/check_system.py.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}