Speedata.io

Big news. Microsoft's paper "CoddSpeed: Hardware Accelerated Query Processing in Microsoft Fabric" just won Best Industry Paper at SIGMOD 2026, a top academic conference in the field of database research, and Speedata's C-200 Analytics Processing Unit (APU) is in it, integrated and benchmarked by Microsoft inside their next-gen Fabric architecture.

The paper documents Microsoft's multi-year effort to move analytics onto hardware accelerators. Microsoft ran the APU on their own hardware-agnostic abstraction layer, the same one their GPU and FPGA paths use.

The results, measured by Microsoft: up to 259× on TPC-H Q13 vs. single-core SQL Server, the compute-heaviest query in the set, where the APU shines.

Microsoft put our thesis in print - GPUs are "powerful but costly and power-hungry." The APU offers a far better price performance path.

This is what validation looks like. A peer-reviewed paper from one of the world's largest data platform providers. Every author on this paper is a Microsoft engineer. The measurements, the integration, the conclusions, all theirs. What put the APU there was two and a half years of our engineers building side by side with Microsoft's.

Congratulations to Microsoft and our team on this award.

We'll be at Databricks Data Engineering Data + AI Summit in San Francisco next week, meeting with teams rethinking the economics of their analytics stack. Set a session here: https://calendly.com/adi-speedata/30min

Paper in the first comment.

Most of the AI spend conversation is about models and GPUs. The $$$ hefty line item nobody's modeling yet is what it costs to run analytics once agents, not people, are driving the query volume. That bill scales fast, and it doesn't scale on the hardware most enterprises are running today.

Test our Workload Analyzer to see what your SQL analytics actually cost on the Analytics Processing Unit (APU) versus CPU and GPU. Link in comments.

Most organizations are bolting AI onto infrastructure built for humans. It works as a starting point, but it's a ceiling, not a strategy.

Our AI expert, Adi Fuchs lays out the practical rules for running AI coding agents inside real VLSI environments: skills, explicit context, traceability, minimal diffs, evidence-based gating. The stuff that separates AI "help" from actual engineering capability.

But the bigger shift is "Build for AI": redesigning the VLSI process from first principles so human engineers and AI agents collaborate natively. Machine-readable artifacts by default. Context and state as the primary, not legacy knowledge scattered across scripts, wikis, and Slack.

9 practical rules, including the one most teams skip - define your human-only decisions up front. Let agents draft fixes and scaffolding. Reserve architecture, CDC sign-off, and tapeout readiness for people.

Read more: https://www.speedata.io/post/know-your-unknowns-building-knowledge-bases-for-your-vlsi-organization-practical-rules

Speedata is growing by leaps and bounds. This week's role: Software Engineering Manager, Performance Modeling.

Lead the team building the simulators that serve as the source of truth for our ASIC, functional and performance models that drive architectural exploration.

What you'll do:
- Lead and grow a team of experienced engineers building hardware simulators from the ground up
- Drive co-design initiatives across Architecture, VLSI, and Software teams
- Provide insights that shape chip architecture, compiler optimization, and system-level performance

Purpose-built silicon. Purpose-built careers. Join us. Email [email protected] or

SW Engineering Manager – Performance Modeling | Speedata Speedata is modernizing analytics infrastructure with the first purpose-built ASIC processor, the Analytics Processing Unit (APU), for analytics and AI data workloads. Delivering up to 100x faster Apache Spark performance while cutting infrastructure TCO by 90%, the APU executes analytics operations...

Model training had scaling laws. A clear improvement trajectory. Data pipelines don't. No roadmaps, no best practices, and it's a big reason most enterprise AI pilots quietly fail.

A deployed model only knows what it was trained on, public data or whatever fixed dataset it shipped with. Making it valuable inside an enterprise means feeding it the company's own structured data, cleanly. That's the gap behind most failed AI pilots, and the workload our APU is built for. Our CEO Adi Gelvan goes deeper on it with Daniel Nenni at SemiWiki.com The Open Forum for Semiconductor Professionals.

CEO Interview with Adi Gelvan of Speedata Speedata created the world's first Analytics Processing Unit (APU), a purpose-built processor designed specifically to accelerate big data...

The Speedata Workload Analyzer projects how much faster your real analytics or AI data prep pipelines would run on the Analytics Processing Unit (APU).

Three ways to test:

🟧 Upload a Spark event log in the browser - results in minutes
🟧 Run our CLI locally - your logs never leave your environment
🟧 Compare against TPC-DS benchmarks

How much faster would your Spark workloads run on dedicated analytics silicon?

Try our Workload Analyzer: https://tinyurl.com/SpeedataWorkloadAnalyzer

Why does a GPU running SQL feel like it's barely trying? Why are TPUs cloud-only? What does an LPU actually do that a GPU can't?

The architectures are different because the workloads are different, and at production scale, those differences compound into real money.

A breakdown of the five processors defining modern compute, and where each one fits https://www.speedata.io/post/apu-vs-lpu-vs-tpu-vs-gpu-vs-cpu-when-to-use-each-one

Speedata is growing fast. This week's role: Lead SoC Architect.

Own the architecture of an ASIC built from the ground up for analytics and AI data prep, not a repurposed processor.

What you'll do:
- Define hardware specifications for subsystems across the SoC
- Break down sub-system architecture into microarchitectural blocks with the related firmware and SW components
- Design the SoC to fit precisely into the datacenter systems it integrates with

Purpose-built silicon. Purpose-built careers. Join us. Email [email protected]

https://www.speedata.io/careers/lead-soc-architect

The Speedata analytics processing unit (APU) is purpose-built silicon for Apache Spark analytics acceleration, and every Wednesday we host a live session for anyone who wants to see it in action.

We spend 20 minutes running Spark SQL or AI data prep workloads on the APU and walking through the architecture, then open up the final 10 minutes for live Q&A.

Sessions alternate between 10am PDT and 10am CEST so it works whichever side of the Atlantic you're on.

Pick a Wednesday: https://www.speedata.io/live-apu-demo

We hosted our webinar on the modern AI compute stack.
The GPU-first model made sense when AI was experimental. In production, efficiency is the priority. Running the wrong workload on the wrong chip means overpaying in power, memory, and infrastructure costs.

We broke down where each processor fits:

🟧 CPUs for orchestration and control logic
🟧 GPUs for model training and inference
🟧 TPUs and cloud ASICs for hyperscale AI
🟧 APUs for analytics-heavy data pipelines and AI data preparation
🟧 LPUs for low-latency inference decoding

And why the agentic era is creating an entirely new pressure point on data infrastructure that most architectures weren't designed to handle.

Full recap and recording now live on the Speedata blog. Link in comments.