For years, improving data warehouse performance meant spending more: more servers, more memory, more money. Microsoft just turned that logic upside down at Microsoft Build 2026, announcing GPU acceleration for the Fabric Data Warehouse, with query speed gains of up to 7x, and without rewriting a single line of SQL.
The early access preview is scheduled for July 2026. Here's what's actually happening under the hood.
Meet CoddSpeed
The technology powering the acceleration has a name: CoddSpeed. This isn't a minor tweak to the system. It's a query execution engine developed over several years by Microsoft Research, derived from a prototype called TQP (Tensor Query Processor).
The technical work behind it was solid enough to earn the Best Industry Paper award at ACM SIGMOD 2026, the most prestigious database conference in academia. The paper is titled "CoddSpeed: Hardware Accelerated Query Processing in Microsoft Fabric."
CoddSpeed is built around something called the Coprocessor Abstraction Layer (CAL), a hardware-agnostic API that allows the system to automatically route query fragments to different types of hardware accelerators: GPUs, FPGAs, or ASICs. If a query isn't eligible for GPU processing, the system falls back to CPU seamlessly, with no errors and no interruptions.
The benchmark numbers
Microsoft published internal benchmarks conducted in May 2026, testing up to 64 simultaneous users querying the Data Warehouse. The results:
- 1 concurrent user: 3x faster than CPU-based competitors
- 16 concurrent users: 6x faster
- 64 concurrent users: up to 7x faster
The pattern here is worth noting. The acceleration isn't linear: it compounds as concurrency increases. GPUs are designed for massive parallelism, so the more users competing for resources simultaneously, the more the hardware works in your favor.
On TPC-H SF=100 benchmarks with a single NVIDIA A100 GPU, CoddSpeed recorded a 7.9x warm speedup and 4.7x cold speedup. With 8x H100 GPUs on a single node at TPC-H SF=1000 (1 TB dataset), the speedup reached 27.1x compared to CPU execution.
For context: CoddSpeed was measured as approximately 2x faster than HeavyDB, already considered one of the fastest GPU query processors available.
How to enable it
The most surprising part of this announcement isn't the numbers. It's how simple adoption is. No data migration. No query rewrites. No special infrastructure provisioning.
GPU acceleration is enabled through a single toggle in workspace settings. From there, Fabric's query optimizer automatically decides which parts of each query go to the GPU and which stay on the CPU. End users don't need to know or even notice it's happening.
All SQL Analytics Endpoints and Data Warehouses within the affected workspace receive the acceleration automatically.
Why this matters right now
The timing isn't coincidental. Corporate analytics is changing. Historically, the data warehouse served primarily batch reporting: heavy queries running overnight, dashboards updated once a day. Latency was tolerable.
The new demand comes from AI systems, autonomous agents, and applications that query data in real time to make decisions. These workloads require response speeds that CPUs, no matter how good, struggle to deliver as concurrency scales up.
Microsoft is repositioning the Fabric Data Warehouse from a reporting repository into an active execution layer for AI, agents, and systems that continuously reason over data. GPU is the piece that makes this viable at scale.
What changes for current Fabric users
If you already work with Microsoft Fabric, the main takeaway is simple: you don't need to do anything differently to benefit from this when the preview becomes available. Your existing SQL queries will run exactly as before. Just faster.
For teams dealing with slow dashboards during peak hours, reports that lock up when too many users connect simultaneously, or semantic models that take too long to process complex queries, this announcement is directly relevant.
The preview starts in July. Worth keeping an eye on.
Sources: Microsoft Fabric Blog (Build 2026), Microsoft Research (CoddSpeed paper), ACM SIGMOD 2026, Microsoft internal benchmarks (May 2026).