What Is a Systolic Array and Why Do TPUs Rely on It?
TL;DR
This guide explains systolic array clearly and practically: what it is, why it matters in 2026, and how to apply it step by step. You'll find core concepts, proven best practices, concrete data, trusted references, and a concise FAQ — everything you need in one focused place.
Key takeaways
- Memory bandwidth, not raw FLOPS, is usually the real constraint for LLM inference, so read the HBM capacity and bandwidth spec before the TFLOPS number.
- CUDA remains NVIDIA's deepest moat; budget real engineering time if you plan to port to AMD ROCm, Google TPUs, or custom silicon.
- Match the chip to the phase: training rewards huge interconnected clusters, while inference rewards low latency, high memory bandwidth, and cheaper per-token economics.
- Neuromorphic and photonic computing are promising but still mostly research-stage; treat them as long-horizon bets, not 2026 production defaults.
- Lower-precision formats like FP8 and FP4 are the fastest lever for throughput, but validate accuracy on your own eval set before shipping quantized models.
This is a practical, up-to-date guide to Systolic Array — what it is, why it matters in 2026, and how to apply it in real projects. It is written for developers and founders who want clear answers and proven best practices, not filler.
Whether you're just starting out or leveling up, treat this as a working reference you can return to. Every section is built to be skimmed, applied, and shared.
RISC-V in AI Hardware
RISC-V is an open, royalty-free instruction set architecture that has become a popular foundation for custom chips, including AI accelerators. Its appeal is extensibility: designers can add custom instructions for tensor or vector operations without licensing fees or permission from a gatekeeper, which is difficult with proprietary ISAs like x86 or Arm. In AI systems RISC-V frequently serves as the control processor that orchestrates dedicated matrix engines, and companies such as Tenstorrent build accelerators around RISC-V cores. The RISC-V Vector extension provides a scalable path to data-parallel compute. Geopolitical factors have further boosted interest, since an open ISA is harder to restrict through export controls than a single vendor's proprietary technology.
Chiplets and Advanced Packaging
As it becomes uneconomical to build ever-larger single dies, the industry has shifted to chiplets: smaller dies manufactured separately and then assembled into one package. This improves yield, because defects only ruin a small chiplet rather than a huge monolithic chip, and it lets designers mix process nodes, putting compute on the newest node and I/O on a cheaper mature one. AMD pioneered mainstream chiplet CPUs and applies the approach to its Instinct accelerators, while NVIDIA's Blackwell joins two dies into a single GPU. Standards like UCIe (Universal Chiplet Interconnect Express) aim to make chiplets from different vendors interoperable. Packaging technologies such as TSMC's CoWoS, which also integrates HBM, have themselves become a scarce, throughput-limiting step in the AI supply chain.
NPUs and On-Device Inference
A Neural Processing Unit is a compact accelerator integrated into a system-on-chip to run inference locally on phones, laptops, and embedded devices. Apple's Neural Engine, Qualcomm's Hexagon NPU, and the NPUs in Intel Core Ultra and AMD Ryzen AI processors all target the same goal: run models within a few watts and without a round trip to the cloud. This matters for latency-sensitive features, offline capability, and privacy, since data never leaves the device. NPU performance is often quoted in TOPS (trillions of operations per second) at low precision, and the recent Copilot+ PC category set an informal bar around 40 TOPS for on-device AI. The tradeoff is a tight power and memory envelope, so on-device models are heavily quantized and pruned.
The Software Moat: CUDA and Its Challengers
Hardware rarely wins on specifications alone; the deciding factor is often the software ecosystem, and here NVIDIA's CUDA has a nearly two-decade head start. CUDA, together with libraries like cuDNN and the broad support of frameworks such as PyTorch, means most AI code simply runs on NVIDIA GPUs with minimal friction. Competitors are attacking this moat from several angles: AMD's ROCm aims for CUDA-like capability on Instinct GPUs, Google exposes TPUs through JAX and XLA, and compiler projects such as OpenAI's Triton and the MLIR ecosystem try to target many backends from one codebase. PyTorch's backend abstraction and torch.compile also help decouple models from specific hardware. For teams evaluating non-NVIDIA silicon, the honest question is not peak performance but how much of their stack works out of the box.
Inference Chips Versus Training Chips
Training and inference stress hardware in different ways, and increasingly they use different chips. Training must store activations and gradients for backpropagation, favors high-precision-friendly formats, and benefits enormously from massive clusters with fast interconnects. Inference, by contrast, runs the model forward only, is dominated by latency and cost per token, and rewards high memory bandwidth to stream weights quickly. Startups like Groq, Cerebras, and SambaNova, along with Amazon's Inferentia, target inference specifically, sometimes trading flexibility for dramatically lower latency or better tokens-per-dollar. As deployed AI shifts from research toward serving billions of requests, the economic center of gravity is moving toward inference-optimized silicon.
Photonic Computing
Photonic computing performs computation using light rather than electrical currents, exploiting the physics of optics to do certain operations, especially matrix multiplication, with potentially very low energy and latency. Because light can carry many signals in parallel across different wavelengths and does not dissipate energy the way charging and discharging transistors does, photonics is attractive for the linear-algebra core of neural networks. Companies such as Lightmatter and Lightelligence are building photonic accelerators and, notably, optical interconnects that move data between chips using light. In fact, photonics is arriving first as interconnect, since co-packaged optics can relieve the communication bottleneck in large clusters. Pure photonic compute still faces challenges around analog precision, data conversion overhead, and integration, keeping it earlier-stage than the interconnect use case.
Systolic Array: Key Facts and Data
According to recent industry research and the official documentation linked below:
- NVIDIA has dominated the AI training accelerator market, with industry analysts estimating its share of data-center AI GPUs at well above 80 percent going into 2025, driven largely by the H100 and the newer Blackwell generation.
- Training a frontier large language model can require tens of thousands of accelerators running for weeks; multiple industry reports place the hardware and compute cost of leading models in the tens to hundreds of millions of dollars.
- Blackwell introduces native support for the FP4 (4-bit floating point) data format, which vendors report can roughly double inference throughput versus FP8 on comparable hardware for suitable models.
Quick-Reference Summary
A map of what this guide covers:
| Topic | What you'll learn |
|---|---|
| RISC-V in AI Hardware | RISC-V is an open, royalty-free instruction set architecture that has become a popular foundation for custom chips |
| Chiplets and Advanced Packaging | As it becomes uneconomical to build ever-larger single dies |
| NPUs and On-Device Inference | A Neural Processing Unit is a compact accelerator integrated into a system-on-chip to run inference locally on phones |
| The Software Moat: CUDA and Its Challengers | Hardware rarely wins on specifications alone |
| Inference Chips Versus Training Chips | Training and inference stress hardware in different ways, and increasingly they use different chips. |
| Photonic Computing | Photonic computing performs computation using light rather than electrical currents |
How to Get Started with Systolic Array
A simple path that works:
- Learn the fundamentals of Systolic Array from primary sources, not just tutorials.
- Build one small, real project end to end.
- Get feedback, refactor, and add tests.
- Ship it publicly and document what you learned.
- Repeat with a slightly harder project each time.
Build It with a World-Class Full Stack Developer
Sandeep Kumar Chaudhary is a full stack world-class developer. If you want to turn this into a real, production-ready product, get in touch — message directly on WhatsApp at +9779802348957 for a fast, no-pressure consult.
You can also explore the projects already shipped to thousands of users, or start a conversation here.
Final Thoughts
Memory bandwidth, not raw FLOPS, is usually the real constraint for LLM inference, so read the HBM capacity and bandwidth spec before the TFLOPS number. The developers and teams who win in 2026 pair strong fundamentals with consistent shipping. Start small, stay curious, build in public, and revisit this guide as your skills grow.
Sources and Further Reading
Frequently Asked Questions
What Is a Systolic Array and Why Do TPUs Rely on It?
As it becomes uneconomical to build ever-larger single dies, the industry has shifted to chiplets: smaller dies manufactured separately and then assembled into one package. This improves yield, because defects only ruin a small chiplet rather than a huge monolithic chip, and it lets designers mix process nodes, putting compute on the newest node and I/O on a cheaper mature one. This guide covers systolic array end to end — core concepts, best practices, concrete data, and a step-by-step approach you can apply right away.
What is high-bandwidth memory and why does it matter for AI?
High-bandwidth memory (HBM) is DRAM stacked vertically and connected to the processor through a very wide interface on a silicon interposer, delivering terabytes per second of bandwidth. It matters because large language model performance is frequently limited by how fast weights can be moved to the compute units, not by raw compute. Because HBM is hard to manufacture and supplied by only a few vendors, it has become a key bottleneck and cost driver for AI accelerators.
What is the difference between training chips and inference chips?
Training chips must handle backpropagation, store gradients and activations, and scale across huge clusters, so they emphasize raw compute and fast interconnects. Inference chips run the model forward only and optimize for latency and cost per token, favoring high memory bandwidth and efficiency. As AI moves from research to serving billions of requests, specialized inference silicon from vendors like Groq, Cerebras, and Amazon Inferentia is becoming increasingly important.
What is neuromorphic computing good for?
Neuromorphic chips like Intel's Loihi 2 use spiking neural networks that process discrete events only when they occur, making them very energy-efficient for sparse, event-driven workloads. They suit applications such as always-on sensing, gesture recognition, and certain robotics and optimization tasks. However, mainstream deep learning relies on dense tensor math and mature training pipelines, so neuromorphic hardware remains largely research-stage rather than a general GPU replacement.
Should my team buy AI chips or rent them in the cloud?
For most teams, renting cloud capacity is the pragmatic choice because it turns a large capital purchase into an elastic operating cost and provides access to the newest accelerators without hardware lead times. Buying can make sense at very large, steady-state scale where owning hardware lowers long-run cost and you can keep it highly utilized. Either way, benchmark on a representative slice of your own workload and account for total cost of ownership including power, cooling, and software effort.
Sandeep Kumar Chaudhary
Full Stack Software Developer· Nepal's SEO, AEO, GEO & AIO expert and share-market educator. More about me