AI Accelerators Beyond GPUs: TPU, Trainium, Gaudi and Wafer-Scale
For senior hardware and infrastructure engineers who must evaluate non-GPU AI accelerators on architecture, system design, software maturity and total cost, and decide when one beats a GPU for a given workload.
The curriculum is being prepared and reviewed. Enrollment is not open yet.
About this course
GPUs dominate AI compute, but a growing share of training and inference runs on accelerators built around different bets: large systolic arrays with compiler-scheduled dataflow, on-chip SRAM instead of HBM, whole-wafer silicon, or chips co-designed with a single company's models. This course teaches the architecture behind those bets well enough to predict where each wins and where it struggles. It opens with the design space (spatial versus temporal architectures, systolic arrays, dataflow, SRAM-versus-HBM memory strategies and deterministic scheduling) and the roofline and memory-capacity reasoning that separates them.
It then examines the platforms current in October 2026: Google TPUs through Trillium and Ironwood (TPU v7) with optical circuit-switched pods and the JAX and XLA stack; AWS Trainium2 and Trainium3 UltraServers, Inferentia2 and the Neuron SDK; Intel Gaudi 3 and Intel's announced Crescent Island inference GPU; Cerebras WSE-3 and CS-3 systems; Groq LPUs, including NVIDIA's Groq 3 LPX inference accelerator for Vera Rubin following NVIDIA's December 2025 licensing agreement with Groq; SambaNova reconfigurable dataflow units; Tenstorrent; Qualcomm AI200 and AI250; and in-house chips such as Microsoft Maia 200 and Meta MTIA. Each is treated through the same lens: compute and number formats, memory hierarchy, scale-up and scale-out interconnect, compiler and framework support, availability model (cloud-only or purchasable), and operational maturity. The course closes with a rigorous evaluation method: workload characterization, porting cost, benchmark design that avoids vendor-favorable traps, and total-cost modeling. Vendor performance figures are presented as vendor claims; announced-but-not-shipping products are labeled. The course ends with an automated exam and a Wirewalk Institute certificate of completion and assessment; it is not a vendor certification.
What you will be able to do
Classify accelerator architectures by compute organization, memory strategy and scheduling model, and predict the workloads each favors
Explain systolic-array and dataflow execution and the role of the compiler in mapping models onto them
Compare Google TPU, AWS Trainium and Inferentia, Intel Gaudi, Cerebras, Groq, SambaNova and hyperscaler in-house accelerators on compute, memory, interconnect and software maturity
Analyze SRAM-centric and wafer-scale designs for capacity limits, model-sharding requirements and latency advantages
Assess the porting cost and operational risk of moving a training or inference workload from GPUs to a non-GPU accelerator
Design a fair benchmark comparing an accelerator with a GPU baseline, controlling for precision, batch size, sequence length and system scale
Build a total-cost and risk model that supports a defensible accelerator decision for a stated workload
Syllabus
6 modules · 21 lessons · about 18 hours
1
Module 1: The Accelerator Design Space
The architectural choices that distinguish non-GPU accelerators and how to reason about them.4 lessons
Spatial Versus Temporal Architectures: Systolic Arrays, Dataflow and SIMT
Memory Strategies: HBM-Centric, SRAM-Centric and Hybrid Designs
Compiler-Scheduled Execution and Determinism
Roofline and Capacity Reasoning Across Architectures
2
Module 2: Google TPU
TPU architecture, pods and the software stack, from Trillium to Ironwood.3 lessons
TPU Architecture: MXUs, SparseCores and HBM
Ironwood and TPU Pods: ICI, Optical Circuit Switching and Slices
XLA, JAX and PyTorch on TPU: What Ports Easily and What Does Not
3
Module 3: AWS Trainium and Inferentia
AWS's training and inference silicon, UltraServers and the Neuron stack.3 lessons
Trainium Architecture: NeuronCores, HBM and Number Formats
Trainium3 UltraServers and NeuronLink Scale-Up
Inferentia2 and the Neuron SDK: Compilation, Kernels and Framework Support
4
Module 4: Intel Gaudi and Merchant Alternatives
Gaudi's Ethernet-native design, Intel's roadmap, and other purchasable accelerators.3 lessons
Gaudi 3: Matrix Engines, TPCs and Integrated Ethernet Scale-Out
Intel's Roadmap: Crescent Island and Beyond
Tenstorrent, Qualcomm AI200 and AI250 and Other Merchant Options
5
Module 5: Wafer-Scale, LPU and Dataflow Designs
Cerebras, Groq and SambaNova, and the in-house chips of hyperscalers.4 lessons
Cerebras WSE-3 and CS-3: Wafer-Scale Compute and Weight Streaming
Groq LPUs and NVIDIA Groq 3 LPX: Deterministic SRAM Inference
SambaNova Reconfigurable Dataflow and Three-Tier Memory
Hyperscaler In-House Chips: Microsoft Maia and Meta MTIA
6
Module 6: Evaluating and Choosing an Accelerator
A disciplined method for comparing accelerators with GPUs and making a defensible decision.4 lessons
Workload Characterization: Prefill, Decode, Training and Fine-Tuning
Fair Benchmarking: Precision, Batch, Sequence Length and Scale
Porting Cost, Software Maturity and Vendor Risk
Total Cost Modeling and the Decision Memo
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 25 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 55 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by Google Cloud, AWS, Intel, Cerebras, Groq, NVIDIA, SambaNova, Microsoft, Meta, Qualcomm, Tenstorrent, or AMD. Product and company names are trademarks of their respective owners.
A plain-English, one-hour walk through an AI data center: what a model asks of hardware, what is inside a GPU server, how GPUs talk over NVLink, InfiniBand and Ethernet, where the data lives, and why AI racks need so much power and cooling.
A plain-language tour of the hardware that runs AI: why AI is limited by compute, memory bandwidth and data movement; what GPUs, HBM and precision formats are; how servers, networks, storage, power and cooling fit together; and how to size a simple inference or training system and weigh cloud against buying.
An engineer's tour of the modern data center GPU: how SMs and CUs, warps and wavefronts, tensor and matrix cores and the on-chip memory hierarchy work; what FP64 through FP4 and sparsity really buy you; how HBM, PCIe, NVLink, Infinity Fabric, UALink and CXL move data; and how power, cooling and reliability features shape real deployments. Reviewed against vendor sources as of October 2026.