Products / Data accelerators

Apache Iceberg GPU acceleration

Compact your lakehouse 7× faster, with the same SQL.

Talk to us about Apache Iceberg beestream-ai

At a glance● same answers

7.3×faster compaction, seven TPC-DS fact tables

21.9×faster reads with 1M equality deletes

6×faster TPC-H Q1 query

102/102TPC-DS queries identical to Spark

The problem

Lakehouse maintenance that never stops billing

Every update to an Iceberg table leaves delete files behind. Compaction has to merge them, on CPUs, every day, for every table. It is a cost line that never goes away, and today there is no full GPU path for it.

What Beestream does

A drop-in GPU path.

Beestream runs Iceberg compaction and merge-on-read scans on the GPU, inside the Spark job you already have. You call the same procedures. Any step that cannot run exactly on the GPU falls back to Spark.

What you get

Nothing to change: same SQL, same procedures, same tables
Per table, 6.6× to 10.0× faster, with the same rows and no delete files left
About 80% lower cost per compaction run, modelled on AWS list prices (L40S GPU vs a 32-vCPU instance)
Engine-neutral core that Flink, Trino or a standalone compactor can use too

Compaction time, seven TPC-DS fact tables · lower is better

Apache SparkCPU, 32 threads

38.7 s

Beestreamsame GPU host

5.3 s

RTX 4090 vs i9-13900K, 32 threads · Iceberg 1.12 · libcudf 26.08

Spark SQLunchanged

CALL system.rewrite_data_files(
  table => 'lake.sales'
);
-- same procedure, now on the GPU

The Beestream platform

Start with one product. Add the rest on the same GPUs.

Apache Iceberg is one part of the Beestream platform. Once it is in, the other products run on the same GPUs, under the same support, with no new stack to adopt.

UDFsSpark RAPIDS GPU accelerationcompiled to GPU kernels 19×JVector GPU accelerationfaster vector index builds 1.4×jitLLMllama.cpp prompt speed 2.4×jitNNfaster than PyTorch, ResNet-50 13×jitCastfaster forecasts

Contact

Tell us what is costing you CPU time. We'll show you the GPU path.

LinkedIn/company/beestream-ai X@beestream_ai Bluesky@beestream.ai GitHubgithub.com/beestream-ai