Estimating Snowflake Resource Usage
Description of your new file.
Overview
Kumo’s Native App uses Snowflake resources depending on the size of the data and the complexity of the models being trained. These resources are primarily consumed in two ways:
-
Running containers in
Snowpark Container Services (SPCS)
.
-
Executing
for data processing using Snowflake warehouses
The total resource usage in Snowflake is determined by:
-
The number of hours the app is running.
-
The selected hardware and warehouse size configuration.
-
The number of training or prediction jobs executed.
-
The size of the input data for each job.
SPCS Resource Usage
Kumo runs two key components on Snowpark Container Services:
-
Kumo Control Plane
(UI and API management).
-
Kumo AI Engine
(model training and inference).
How Resource Consumption is Measured: As of 11-Feb-2025 Snowflake charges based on:
-
Hardware configuration
(CPU, GPU, Memory allocated to the compute instance used).
-
Duration
for which the compute instance runs
Compute Resources:
-
Both the
Control Plane
and
AI Engine
run on a single
GPU compute instance
in SPCS.
-
As of 11-Feb-2025, tThe available instance types are
GPU_NV_M
and
GPU_NV_L
, selected at app launch.
Managing Resources Efficiently:
-
The compute resource is used continuously while the app is running.
-
The app
can be suspended
without losing data to save resources.
-
Instructions on suspending the app are available
.
Estimated Credit Usage: The table below provides an estimate of Snowflake credits consumed per hour based on the selected container size. The most up to date numbers can be found in the official Snowflake Credit Consumption Table.
| SPCS Container | Number of Rows | Snowflake Credits Per Hour |
|---|---|---|
| GPU_NV_M | < 250M | 2.68 |
| GPU_NV_L | < 1B | 14.12 |
Warehouse Resource Usage
To process data for training and predictions, Kumo executes Snowpark queries using Snowflake virtual warehouses. This includes:
-
Reading input tables from Snowflake.
-
Preparing data for training and prediction.
-
Generating artifacts that are transferred to the GPU.
Factors Affecting Credit Usage:
-
The
shape and structure
of the input data.
-
The
complexity
of queries being executed.
-
The
configuration of the job
(e.g., if a job processes an expensive Snowflake view, extra compute resources will be used).
-
Using the
Graph Snapshot
feature can reduce resource usage across multiple jobs.
The following table provides an estimated credit consumption per job based on data size.
| Data Size | Number of Rows | Warehouse Size | Snowflake Credits Per Job |
|---|---|---|---|
| 1-9GB | 1-24M | S | 1-3 |
| 10-90GB | 25-240M | M | 3-9 |
| 100-250GB | 250M-1B | L | 10-20 |
Other Resource Usage
Storage:
-
Kumo stores various types of data within Snowflake, including:
-
Temporary artifacts from training.
-
Metadata for tables, graphs, and jobs.
-
Model binaries.
-
-
A typical deployment with around
100 jobs
is expected to consume less than 1TB of data.
-
This would be billed at your account’s standard rate for storage.
Example Calculation
Scenario: Weekly Email Recommendations
-
Data Size:
4 tables (users, products, purchases, browsing history) with a total of
100M rows, 50GB
.
-
Training Frequency:
Once per month.
-
Prediction Frequency:
Once per week.
Estimated Monthly Resource Usage:
-
SPCS Container:
GPU_NV_M
-
App Run Time:
~12 hours per week
-
SPCS Credits per Month:
12 * 4 * 2.68 =
128.64 credits
-
Snowpark Credits Per Job:
6
-
Jobs per Month:
5
-
Snowpark Credits per Month:
5 * 6 =
30 credits
-
Total Monthly Credit Usage:158.64 credits
-
Storage Cost: 500GB
Upcoming Feature: Kumo Autoscaling on SPCS
In Q1 2025, Kumo will introduce dynamic autoscaling for GPUs in SPCS. This will:
-
Automatically scale
GPU resources up or down based on demand.
-
Eliminate the need for manual app launch and suspension.
-
Optimize infrastructure usage, particularly for large workloads.
We expect that this change will simplify management and reduce infrastructure resources requirements.