Siam AI Service Level Agreement (SLA)
1. Definitions
Quantity
The number of GPUs the Service Provider guarantees will always be available for use. For example, if a user rents a 1K Cluster, the provider guarantees a minimum of 1,000 GPU cards.
SLA
Service Level Agreement refers to the agreed-upon service level in the contract. If the Service Provider fails to deliver the service as agreed upon, the availability will be calculated monthly using a formula. It will then be refunded to the user as a monthly service discount or compensated monthly service hours (Compensate Hours) according to the respective section’s formula.
SLA Credit
A discount or monthly service hour compensation. The monthly service fee will be calculated if the service does not meet the agreement (e.g., an SLA lower than stipulated in the contract). The SLA difference will be calculated and refunded as a monthly rental discount or service hour compensation, based on the contract terms.
Monthly Downtime
The accumulated duration within a month during which the user is unable to use the service. This counter resets to 1 on the 1st of every month. For instance, if the agreed SLA is 99.99% but the user experiences a 30-minute system outage proven to be the Service Provider’s fault, the provider will issue a 5% discount or service hour compensation to the user, as specified in the corresponding section tables.
Response Time
The duration from when the system receives an incident report or detects an alert to the time the team formally acknowledges, investigates, and responds to the issue.
Incident
An undesirable or unplanned event that causes service disruption or degrades service quality.
Security Incident
An undesirable cyber threat event caused by malicious external parties and/or internal individuals that impacts the service, potentially escalating to system disruptions and reputational damage to the organization.
Siam AI Corporation Co., Ltd. SLA Calculation Principles
GPU Service Level Agreement
For GPUs, the service level is guaranteed based on the total accumulated Monthly Downtime of all GPUs per month, in accordance with the table below. For conditions regarding the scope of GPU-related issues, please refer to the Demarcation Point section at the end of this topic.
| SLA | SLA Credit | Monthly Downtime |
|---|---|---|
| 99.0% | Guarantee | 7h 18m 17s |
| 98.0%–99.0% | 5% | 14h 36m 35s |
| 95.0%–98.0% | 10% | 1d 12h 31m 27s |
| Less than 95.0% | 15% | < 1d 12h 31m 27s |
Demarcation Point
The Service Provider’s SLA covers only the Hardware, BIOS/Firmware, Host Driver, and GPU Health levels.
Exclusions
- User OS & Software Stack Layer Exclusion: Excludes issues stemming from the user’s PyTorch/CUDA Code, OOM (Out of Memory) from the Model, Custom Container Crashes, and user configuration errors.
- Thermal Throttling Threshold Exclusion: Cases where the GPU reduces its operating speed (Thermal Throttling) due to:
- Processing workloads that exceed standard hardware specifications.
- Modifications to the software/Driver/GPU BIOS by the user.
- Temporary cooling system failures.
If the team can control the temperature to safe levels and restore the system to a normal state within the contractually specified time frame, this will not be calculated as Monthly Downtime or an SLA breach. This exception does not apply if the cooling system completely fails, causing a System Shutdown/Crash or an unrecoverable state according to standard protocols.
- Scheduled Maintenance Exclusion: System shutdowns for Firmware Updates (e.g., GPU BIOS, NVSwitch FW, InfiniBand Subnet Manager) require at least 7 days’ advance notice and must be performed during Off-peak hours. These will not be calculated as Monthly Downtime.
Data Center Facility Infrastructure Availability
The Service Provider guarantees the availability of utilities within the data center (Facility Infrastructure Availability), referencing and complying with the Service Level (SLA Standard / Tier Level) of the respective Data Center Operator, as detailed in the following table:
| SLA | SLA Credit | Monthly Downtime |
|---|---|---|
| 99.99% | Guaranteed by DC | 4m 23s |
| 99.9%–99.99% | 5% | 43m 50s |
| 99.5%–99.9% | 10% | 7h 18m 17s |
| Less than 99.5% | 15% | < 7h 18m 17s |
- Power Availability: Guarantees a continuous power supply from the main power distribution and backup systems (UPS & Generator) up to the rack power distribution units (PDU Feeds), in accordance with the Data Center Provider’s availability rate.
- Cooling Availability: Guarantees continuous cooling and air circulation in accordance with the Data Center Provider’s service standards.
CPU and Storage Server Service Level Agreement
System Availability and Uptime
The Service Provider guarantees the availability of the processing server hardware (CPU) and Storage System at a rate of 99.5% per month, as shown in the table below:
| SLA | SLA Credit | Monthly Downtime |
|---|---|---|
| 99.5% | Guarantee | 3h 39m 8.7s |
| 99.0%–99.5% | 5% | 7h 18m 17s |
| 98.0%–99.0% | 10% | 14h 36m 35s |
| Less than 98.0% | 15% | < 14h 36m 35s |
Demarcation Point
- CPU Infrastructure: Covers the integrity of the CPU hardware, System Memory (RAM), Motherboard, Power Supply Units (PSU), and server-level cooling systems.
- Storage Infrastructure: Covers the operation of data storage disk arrays (Local NVMe / SSD / HDD Array or Distributed Storage Volume) connected to the server to ensure standard Read/Write I/O Operations.
Downtime Definition
A Hardware Outage for processing and storage systems begins to count when one of the following conditions occurs:
- An Unplanned Server Shutdown / Crash or the inability to boot into the operating system due to a Hardware Failure.
- A fatal memory error that cannot be resolved (Uncorrectable ECC Memory Error) or a CPU Machine Check Exception (MCE) that halts the system.
- The storage system is completely inaccessible (Volume Unavailable) or suffers a Total I/O Blockage.
Exclusions
The following events will not be calculated as Monthly Downtime under the SLA agreement:
- Partial disk failures in a Redundant Array (e.g., RAID 1/5/6/10 or Distributed Storage) that can still read and write data normally without causing an I/O Interruption.
- Data loss, corruption, or an unresponsive operating system e.g. OS Crash / Kernel Panic caused by the user’s own software, commands, or applications (User-level Software Failure).
- Scheduled Maintenance that has been notified to the user at least 7 business days in advance.
Notice: The Service Provider reserves the right to modify the SLA conditions on the website. The user must be notified in writing at least 14 days in advance.
Note: The full SLA details are subject to the terms and conditions specified in each individual service agreement.