Guide, read 10 Nov 2026
GPU vs FPGA for Spark: the published facts
What is being compared
This page sets the public record for one GPU option and one FPGA option side by side. The GPU option is the NVIDIA cuDF plugin for Apache Spark (the RAPIDS Accelerator). The FPGA option is DualBird's Spark plugin on Amazon EC2 F2 instances. The tests were run on different workloads, scales and baselines by different parties, so the numbers below are not a head-to-head. They are what each side has published, in one place.
Neither technology is judged here, and nothing is derived from the figures. The plain-words explanation of both is in the guide on hardware acceleration.
What NVIDIA and its partners have published
The main test is by AWS with NVIDIA on Amazon EMR on EKS in us-east-1. TPC-DS at 3 TB, 103 queries, median of five iterations, Spark 3.5.6 and cuDF plugin 26.04.2. Eight g7.4xlarge nodes (one NVIDIA RTX PRO 4500 Blackwell GPU, 16 vCPU and 64 GB each) took 281 seconds. Comparable eight-node CPU clusters took 1,010 to 1,043 seconds, which AWS labels a 3.7x speedup. AWS reports $2.06 per run on the GPU cluster and $2.93 to $3.18 on the CPU clusters, with the GPU cluster at $26.35 per hour (eight instances at $3.042 plus $2.01 for EMR on EKS) and the CPU clusters at $10.43 to $10.99 per hour. AWS leaves the EKS control-plane fee, EBS, S3 and the driver pod out of the cost.
NVIDIA's overview page adds a comparison against separate NDS runs on EC2 with open-source Spark 3.5.6: 8.4x faster than m7gd.8xlarge and 11.0x faster than m9gd.4xlarge, with reported cost per run 77% and 72% lower. NVIDIA itself notes the stacks differ. NVIDIA publishes its NDS automation in a public repository and states NDS is derived from TPC-DS.
Coverage: the DataFrame API and a limited set of SQL and DataFrame operations; not RDDs, Dataset lambdas or GraphX. Anything the plugin does not cover, UDFs included, runs on the CPU. Releases support Spark 3.3.x to 4.2.0, and Databricks (14.3 ML LTS and 17.3 ML LTS) and Dataproc runtimes are listed. Amazon EMR on EKS supports the plugin from release 6.9.0, with a "-spark-rapids" variant that ships it. Independent benchmark: none found on 2026-11-10.
What DualBird has published, labelled as DualBird's
DualBird says its plugin gives 10x to 30x faster processing and 50% to 90% lower costs with no changes required. That is the range in its page text; its meta description says 10x to 100x. Its home page shows a 110 GB enrichment pipeline (join denormalization) with 50% to 85% lower EC2 cost and a 6x to 18x higher performance ceiling, with no hardware, baseline, date or method given.
Its detailed test is a sort-based Iceberg compaction on 100 GB of synthetic data (about 370 million rows, 64 columns, ZSTD level 3), compared with vanilla Spark and an unnamed C++ accelerated Spark. DualBird says roughly 12x to 20x faster single-task execution, 55% to 85% lower EC2 cost, a practical performance ceiling about 6x higher, and about $2 per terabyte compacted against about $6 per terabyte rising past $13 for CPU clusters. DualBird's page describes the method only in summary; the full methodology is in a download behind a form. Instance types and counts are not stated in the public post.
Hardware: AWS lists three F2 sizes, f2.6xlarge (one FPGA, 24 vCPU, 256 GiB), f2.12xlarge (two FPGAs, 48 vCPU, 512 GiB) and f2.48xlarge (eight FPGAs, 192 vCPU, 2,048 GiB). DualBird's engineer says a defined set of operators is accelerated (sort, join, aggregation and others) and Python UDF-heavy work falls back to CPU. No third-party benchmark: none found on 2026-11-10.
DualBird runs on AWS F2 instances only, so it is not an option on Azure, Google Cloud or a Databricks-managed cluster, and every benchmark published for it so far is its own. F2 is listed in eight AWS regions (not, for example, us-east-2 or eu-west-1), the method for its compaction benchmark is behind a sign-up form, and no price for the plugin is published.
Side by side
DataPelago, which lists CPU, GPU and FPGA as targets for its Nucleus engine, published operator microbenchmarks against cuDF 25.04.00 on one H100: up to 10.5x on project and 10.1x on filter, 4.3x on aggregate, as DataPelago's own figures. They are not end-to-end Spark jobs.
What to check next depends on your own cloud, regions and job mix. The trial guide covers how to measure it.
| Fact | DualBird (FPGA) | NVIDIA cuDF plugin (GPU) |
|---|---|---|
| Hardware | AWS F2: AMD Virtex UltraScale+ VU47P FPGAs | NVIDIA GPUs, Volta through Blackwell generations |
| Where it runs | AWS EC2 F2 only, in the customer's AWS account | EMR on EKS, EMR on EC2, Databricks, Dataproc, Kubernetes, on-premises GPU servers |
| Software cost | No price published | Free; pay for GPU instances |
| Largest published scale | 110 GB (enrichment tile) and 100 GB (compaction), both DualBird's | 3 TB, run by AWS with NVIDIA |
| Who ran published tests | DualBird | AWS with NVIDIA; NVIDIA |
| Method published | Compaction method behind a sign-up form | AWS cluster, config, release and prices in the post; NDS scripts public |
| Spark versions listed | None found on 2026-11-10 | 3.3.x to 3.5.9, 4.0.0 to 4.0.4, 4.1.1 to 4.1.3, 4.2.0 (release 26.10.0) |
| Fallback | CPU for unsupported operators and Python UDFs, per DualBird | CPU for unsupported operators and UDFs, per NVIDIA |
Sources
- AWS Big Data Blog: GPU-accelerated Apache Spark with Amazon EMR and NVIDIA on G7 instances, read 10 Nov 2026
- NVIDIA cuDF plugin for Apache Spark: overview, read 10 Nov 2026
- NVIDIA cuDF plugin for Apache Spark: FAQ, read 10 Nov 2026
- NVIDIA cuDF plugin for Apache Spark: download, read 10 Nov 2026
- DualBird home page, read 10 Nov 2026
- DualBird: Iceberg compaction benchmark, read 10 Nov 2026
- DualBird: A simpler Spark, read 10 Nov 2026
- AWS documentation: Amazon EC2 accelerated computing instance types, read 10 Nov 2026
- AWS documentation: instance types by Region, read 10 Nov 2026
- DataPelago: Nucleus versus NVIDIA cuDF, read 10 Nov 2026
Last reviewed 10 Nov 2026.