Summary

• Firebolt is 8x better in price-performance than Snowflake, 18x better than Redshift, and 90x better than BigQuery.
• To achieve Firebolt's performance characteristics, Snowflake is 37.5 times more expensive, whereas Redshift and BigQuery failed to achieve anywhere close to similar performance numbers.
• For equivalently priced engines, Firebolt delivers 5.5x QPS compared to Snowflake and 10x QPS compared to Redshift. Firebolt also exhibits near-linear scaling when adding clusters, delivering low latency of 120 milliseconds with a throughput of 2500 QPS.

Why We Created FireScale

Industry-standard benchmarks often serve as an initial reference point for comparing data warehouse performance. However, they typically don’t capture the concurrency, query complexity, and volume demands of many modern data environments.

With this in mind, Firebolt developed FireScale¹ to better capture the complexity of modern, analytical workloads. FireScale is centered around a benchmark framework that includes three key components:

• A comprehensive set of queries representing customer usage patterns that simulate real-world conditions.
• An extended version of the AMPLab dataset, with additional dimension tables to accurately reflect the complexity of production query patterns.
• A range of workload execution scenarios that determine how these queries are run to simulate diverse operational conditions, such as varying concurrency levels and data sizes.

This approach ensures that FireScale can rigorously test a cloud data warehouse’s (CDW) performance under practical, high-demand conditions. The FireScale benchmark is publicly available on GitHub, allowing anyone to explore and reproduce the results as well as test Firebolt's capabilities firsthand.

Benchmark Methodology at a Glance

For the purposes of this benchmark, we performed two types of runs on FireScale: a power run to test query latency and cost efficiency and a multi-cluster concurrency run to measure throughput under high loads. We compared Firebolt, Snowflake, Redshift, and BigQuery configurations that were as closely aligned on size and/or pricing as possible, to ensure an accurate and fair comparison. Since each platform uses a different naming terminology for its compute resources, here is a quick explanation of the labeling used in our benchmark results:

• Firebolt: Compute resources are called engines and have different attributes such as node size (Small, Medium, Large, X-Large), # of nodes (1-128), # of clusters (1-10) and compute family type (storage-optimized or compute-optimized). We used only compute-optimized engines for this benchmark, so a label of Firebolt (1 x M) means a compute-optimized engine with a single medium-sized node and single cluster, whereas a label of Firebolt (1 x L) 8C means a compute-optimized engine with a single large-sized node and eight clusters.
• Snowflake: Compute resources are called warehouses, and have only two dimensions of scaling (unlike Firebolt’s multi-dimensional scaling options); size (X-Small, Small, Medium, Large, X-Large) and # of clusters. So a label of Snowflake (L) means a large size warehouse, whereas a label of Snowflake (S) 4C means a small size warehouse with four clusters.
• Redshift: Compute resources are called instances and are defined by node type and # of nodes. So a label of Redshift (3 x ra3.large) means 3 nodes of ra3.large, whereas a label of Redshift (7 x ra3.4xlarge) means 7 nodes of ra3.4xlarge
• BigQuery: We ran the benchmark on BigQuery serverless (single region, on-demand) and are just using BigQuery as the label.

Below is a quick summary of how we conducted these runs and the corresponding results. For a detailed look into the benchmark methodology and the full set of results, please read our technical blog. 

Power Run

The power run tested 25 queries based on the AMPLab dataset at a 1 TB scale. Each of the 25 queries was submitted sequentially five times. The first run was excluded from the query latency calculation, and the time to run the workload was measured as the sum of the median of the next four runs for each query. The cost to run the workload was calculated as a factor of the time taken to run the workload and the corresponding pricing on each platform.

Time vs. Cost: Firebolt Achieves the Fastest Performance at the Lowest Price Point

This chart maps each engine’s total workload runtime (horizontal axis) against overall cost (vertical axis). All of Firebolt’s engines (with single node sizing of S, M, L, XL) converge in the upper-right corner, indicating near-minimal execution time at a fraction of the cost compared to Snowflake and Redshift. Notably, even Snowflake’s largest warehouse (XL) or Redshift’s largest instance (7 x ra3.4xlarge) cannot match Firebolt’s sub-second query latency while incurring substantially higher expenses.

Workload Completion Times: Firebolt Consistently Outruns All Other Alternatives

This chart tracks how many seconds each engine takes to finish the benchmark workload. All four of Firebolt’s engines outperform even the fastest workload execution times for Snowflake, Redshift, and BigQuery. When comparing for equivalently sized/priced configurations, Firebolt is 3.7x faster than Snowflake, 6x-16x faster than Redshift, and 6.5x faster than BigQuery.

Lowest Cost per Workload: Firebolt Delivers High Performance at a Fraction of the Price

This chart compares the expense of running the benchmark workload across all four platforms. All four of Firebolt’s engines are significantly less expensive than even the least costly option for Snowflake, Redshift, and BigQuery. When comparing for the best price-performing engine across platforms, Firebolt (1 x S) is 8x better than Snowflake (XS), 18x better than Redshift (3 x ra3.large), and over 90x better than BigQuery.

To Achieve Similar Performance, Snowflake is 37.5 times more expensive than Firebolt

Firebolt’s smallest engine with a single small size node (1 x S) completes the benchmark workload in 16.77 seconds at a cost of $0.0065, while Snowflake’s largest warehouse (XL) takes 18.33 seconds and costs $0.2444. That’s a 37.5x higher expense for Snowflake, while still taking longer execution time than Firebolt. Additionally, both Redshift and BigQuery failed to achieve anywhere close to the 16.77 seconds time to complete the workload.

Concurrency Run

For the concurrency run, we selected five of the fastest (vendor-agnostic) queries from the original 25 and generated 50,000 variations, executing them in random order at scale. To measure throughput, we tested each system under simulated multi-user conditions to ensure full hardware utilization across all vendors, while avoiding over-saturation, to maximize the resulting QPS (queries per second).

Higher Concurrency Throughput: Firebolt Leads by a Wide Margin

When comparing equivalently priced configurations, Firebolt (1 x M) 8C reaches ~1700 QPS, which is 5.5x compared to Snowflake (S) 4C and 10x compared to Redshift (8 x ra3.4xlarge). Firebolt’s high throughput ensures more concurrent users can query data simultaneously without significant slowdowns.

Firebolt’s Concurrency Scaling Outpaces Both Snowflake and Redshift

While both Firebolt and Snowflake exhibit near-linear scaling upon adding clusters, Snowflake tops out at 640 QPS which is almost a quarter of the 2500 QPS delivered by Firebolt. Redshift on the other hand tops out at just 165 QPS, and fails to demonstrate any substantial gains in concurrency scaling upon moving to larger instances.

Near-Linear Concurrency Scaling: Firebolt Maximizes QPS with Added Clusters

Upon adding clusters, Firebolt’s (1 x M) and (1 x L) engines exhibit near-linear scaling for concurrency. Firebolt (1 x M) goes from 223 QPS on a single cluster to almost 1700 QPS on an eight cluster configuration. Firebolt (1 x L) goes from 355 QPS on a single cluster to almost 2500 QPS on an eight cluster configuration, while delivering low latency of 120 milliseconds.

Conclusion

We built FireScale to hold modern cloud data warehouses to a higher standard—one that accurately reflects real-world  production workloads. From the power run’s single-query speed tests to the concurrency run’s intense multi-user scenarios, the results show Firebolt consistently outperforming Snowflake, Redshift and BigQuery across query latency, concurrency scaling and overall price-performance. In an era defined by large-scale AI and hyper-personalized data apps, these performance and cost advantages aren’t just “nice to have”—they’re essential for creating responsive, engaging experiences at scale while keeping costs under control.

Ready to try Firebolt yourself? Sign up today for a trial and receive $200 in free credits to experience sub-second queries and high concurrency scaling in your own data environment. If you need help getting started or have questions, please contact us.

¹ For a detailed look at how we developed the FireScale benchmark, please read the Benchmark Methodology section of our 'Firebolt Unleashed: High Efficiency and Low-Cost Concurrency in Action' blog.