NPUs for Beginners: How On-Device AI Hardware Works
TL;DR
A complete, up-to-date breakdown of npus for developers and founders. It covers the core ideas, the trade-offs that matter, a practical workflow, real numbers, and the questions people ask most — written to be skimmed, applied, and shared.
Key takeaways
- Neuromorphic and photonic computing are promising but still mostly research-stage; treat them as long-horizon bets, not 2026 production defaults.
- For on-device and edge AI, look at NPUs in the SoC (Apple, Qualcomm, Intel, AMD) rather than discrete GPUs to hit power and latency budgets.
- 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.
- RISC-V is a credible base ISA for custom accelerators and control cores because it is open, royalty-free, and extensible with custom instructions.
- 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.
This is a practical, up-to-date guide to Npus — 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.
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.
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.
Why High-Bandwidth Memory Is the Real Bottleneck
For large models the scarce resource is usually not compute but the speed at which weights and activations can be moved to the compute units. High-bandwidth memory solves this by stacking DRAM dies vertically and connecting them to the processor through a silicon interposer with an extremely wide interface. The current mainstream generation, HBM3e, delivers multiple terabytes per second per stack, and next-generation accelerators pack several stacks around each compute die. Because HBM is hard to manufacture and yields are constrained, it has become a genuine supply bottleneck, with SK hynix, Samsung, and Micron as the only volume suppliers. Practitioners should read an accelerator's memory capacity and bandwidth as carefully as its FLOPS, since they often determine real-world LLM throughput.
Neuromorphic Computing
Neuromorphic computing takes design cues from the brain, using spiking neural networks where information is carried by discrete events (spikes) rather than continuous dense arithmetic. Chips like Intel's Loihi 2 and IBM's TrueNorth and NorthPole colocate memory and computation and process events only when they occur, which can make them extremely energy-efficient for sparse, event-driven workloads. This event-based model suits applications such as always-on sensing, gesture recognition, and certain robotics and optimization problems. The catch is that mainstream deep learning is built around dense tensor math and standard training pipelines, so neuromorphic hardware requires different algorithms and lacks a mature software ecosystem. It remains largely a research and specialized-deployment technology rather than a general-purpose replacement for GPUs.
TPUs and the Case for Custom Silicon
Google's Tensor Processing Unit is the best-known example of a company building its own accelerator rather than buying GPUs. TPUs are built around a large systolic array, a grid of multiply-accumulate units that streams data through in a tightly choreographed pattern to maximize compute per memory access. They are tightly co-designed with the JAX and TensorFlow software stacks and with Google's own optical interconnect, letting TPU pods scale to thousands of chips with high efficiency. Amazon (Trainium and Inferentia), Microsoft (Maia), and Meta (MTIA) have followed with their own in-house accelerators. The strategic logic is control: owning the silicon reduces dependence on a single vendor, tunes hardware to specific models, and can lower total cost at hyperscaler volumes.
Npus: Key Facts and Data
According to recent industry research and the official documentation linked below:
- As of 2025, high-bandwidth memory is a primary bottleneck for AI accelerators, and SK hynix, Samsung, and Micron are the three suppliers producing HBM3e stacks, with SK hynix widely reported as the leading HBM vendor.
- 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.
- The Hopper-based H100 SXM offers 80 GB of HBM3 memory delivering roughly 3.35 TB/s of bandwidth, while the Blackwell B200 pairs two reticle-limited dies into one package with 192 GB of HBM3e and around 8 TB/s of bandwidth.
Quick-Reference Summary
A map of what this guide covers:
| Topic | What you'll learn |
|---|---|
| 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 |
| 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 |
| Why High-Bandwidth Memory Is the Real Bottleneck | For large models the scarce resource is usually not compute but the speed at which weights and activations can be moved to the compute units. |
| Neuromorphic Computing | Neuromorphic computing takes design cues from the brain |
| TPUs and the Case for Custom Silicon | Google's Tensor Processing Unit is the best-known example of a company building its own accelerator rather than buying GPUs. |
How to Get Started with Npus
A simple path that works:
- Learn the fundamentals of Npus 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
Neuromorphic and photonic computing are promising but still mostly research-stage; treat them as long-horizon bets, not 2026 production defaults. 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 npus?
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. This guide covers npus end to end — core concepts, best practices, concrete data, and a step-by-step approach you can apply right away.
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.
Why is NVIDIA so dominant in AI chips?
NVIDIA's dominance comes as much from software as from hardware. CUDA, launched in 2007, plus libraries like cuDNN and deep integration with frameworks such as PyTorch mean nearly all AI code runs on NVIDIA GPUs with minimal effort. Combined with strong hardware, fast NVLink interconnects, and a large installed base, this creates an ecosystem lock-in that competitors find hard to overcome.
What are chiplets and why is the industry moving to them?
Chiplets are smaller dies made separately and assembled into a single package instead of building one large monolithic chip. They improve manufacturing yield, since a defect only ruins a small chiplet, and let designers mix process nodes to optimize cost. Modern high-end accelerators like NVIDIA's Blackwell and AMD's Instinct use this approach, and standards such as UCIe aim to let chiplets from different vendors work together.
What are FP8 and FP4, and why do they matter?
FP8 and FP4 are 8-bit and 4-bit floating-point formats that represent numbers with far fewer bits than the traditional FP16 or FP32. Using lower precision lets a chip do more operations per second and move more values per unit of memory bandwidth, boosting throughput and reducing cost, which is why NVIDIA's Hopper added FP8 and Blackwell added FP4. The tradeoff is potential accuracy loss, so teams should validate quantized models on their own evaluation sets before deploying.
Sandeep Kumar Chaudhary
Full Stack Software Developer· Nepal's SEO, AEO, GEO & AIO expert and share-market educator. More about me