Guide, read 10 Nov 2026
GPU, FPGA and custom silicon for Spark, in plain words
What offloading means
A Spark executor is a JVM process on CPU cores. Hardware acceleration moves the heavy, repetitive parts of SQL and DataFrame operators (scan decode, filter, project, sort, join, aggregation) to a different chip that does that kind of arithmetic on many values at once. The plan, scheduling and fault tolerance stay in Spark. DualBird's own description says the same for its product: Spark still handles planning, scheduling and fault tolerance.
Three consequences follow for the person who owns the bill. The accelerator only helps the work it covers. The data has to travel to the chip and back, so the offloaded stage has to be large enough to cover that trip. And the hardware is something you rent, buy or move to, which is why these options appear in the hardware row of the Change Budget.
FPGAs: DualBird on AWS F2
An FPGA is a chip whose logic can be reconfigured after manufacture, so a vendor can build a pipeline for sort, join and aggregation in hardware. AWS describes F2 as its second-generation FPGA instance. The AWS accelerated-computing table lists three sizes: f2.6xlarge with 24 vCPU, 256 GiB of memory and one AMD Virtex UltraScale+ HBM VU47P FPGA; f2.12xlarge with 48 vCPU, 512 GiB and two FPGAs; and f2.48xlarge with 192 vCPU, 2,048 GiB and eight FPGAs.
DualBird says its Spark plugin runs on these instances: you change the EC2 instance type to F2 and add the plugin. A DualBird engineer says Python UDF-heavy work and unsupported operators fall back to the CPU and see less benefit for now. DualBird says 10x to 30x faster processing and 50% to 90% lower cost on its home page and 10x to 100x in its meta description; the two ranges differ on its own site. Its one worked benchmark is a 100 GB sort-based Iceberg compaction it ran itself on synthetic data (about 370 million rows).
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.
GPUs: the NVIDIA cuDF plugin for Apache Spark
A GPU has thousands of small cores built for the same operation across a column of values. NVIDIA's open-source cuDF plugin (also called the RAPIDS Accelerator for Apache Spark) swaps in GPU implementations of supported SQL and DataFrame operators. NVIDIA says the job still completes when it meets an operator or UDF the plugin cannot run, because that part runs on the CPU.
Hardware needed, from NVIDIA's download page: GPUs of the Volta, Turing, Ampere, Ada Lovelace, Hopper or Blackwell generations, Linux with glibc 2.28 or newer, and an NVIDIA driver R525 or later for CUDA 12 (R580 or later for CUDA 13). AWS lists its GPU families in the accelerated-computing instance docs, including G4dn, G5, G6, G6e, G7 and G7e, plus P4d, P5 and P6 for larger GPUs. The AWS and NVIDIA published test used g7.4xlarge, which has one NVIDIA RTX PRO 4500 Blackwell GPU, 16 vCPU and 64 GB of memory.
The plugin is free; per AWS's post you pay for the GPU instances. NVIDIA has an AutoTuner, a Profiling tool and a Qualification tool that reads CPU event logs and predicts which applications are candidates.
Custom silicon: Speedata's APU
A custom chip is designed from the start for analytics queries. Speedata calls its chip an Analytics Processing Unit (APU). The C200 is a PCIe Gen 5 card, built around the Callisto ASIC, so it needs a physical server. Speedata offers a pre-configured 2U server with two C200 cards, servers from Dell and HPE, and a "Speedata Acceleration for AWS" service in which you keep data in S3, sync changes to a Speedata-hosted environment, run there and push results back. Speedata says its Dash software plugs into the Catalyst optimizer and delegates UDFs and other non-accelerable SQL to the CPU.
Speedata's home page claims up to 100x acceleration and up to 90% lower TCO with zero code changes. Those are Speedata's claims. The Microsoft SIGMOD 2026 paper on Fabric hardware acceleration benchmarked the C200 on TPC-H queries at SF100 against a single-core SQL Server baseline, not against Spark, per Speedata's own summary of the paper. No price list is published.
Not every option fits a single type. DataPelago says its Nucleus engine targets CPU, GPU and FPGA, and publishes operator microbenchmarks against cuDF; its docs for the Spark product configure the Gluten plugin and do not state a GPU requirement.
Side by side, by what you have to change
The guide on GPU versus FPGA lists what each side has published, and the Change Budget lists what each option makes a team change.
| Hardware (alphabetical by example) | Example | Where it runs | What you add |
|---|---|---|---|
| FPGA | DualBird | AWS F2 instances, eight regions per AWS | F2 instance type and the plugin |
| GPU | NVIDIA cuDF plugin | Cloud GPU instances or your own GPU servers | GPU instances, plugin jar, memory and concurrency settings |
| Custom silicon (APU) | Speedata C200 | Your servers, OEM servers or Speedata-hosted AWS service | PCIe cards or a hosted service; Dash plugin |
Sources
- AWS documentation: Amazon EC2 accelerated computing instance types, read 10 Nov 2026
- Amazon EC2 F2 instances, read 10 Nov 2026
- AWS documentation: instance types by Region, read 10 Nov 2026
- NVIDIA cuDF plugin for Apache Spark: download and requirements, read 10 Nov 2026
- NVIDIA cuDF plugin for Apache Spark: overview, read 10 Nov 2026
- DualBird home page, read 10 Nov 2026
- DualBird: A simpler Spark, read 10 Nov 2026
- Speedata product page, read 10 Nov 2026
- Speedata home page, read 10 Nov 2026
- DataPelago technology page, read 10 Nov 2026
Last reviewed 10 Nov 2026.