Electronics World : TPU VS. GPU

In today's booming era of artificial intelligence and high-performance computing, GPUs (graphics processing units) and TPUs (tensor processing units) are the two core engines driving this revolution. While both demonstrate impressive prowess in executing complex mathematical calculations, their underlying design philosophies differ significantly: the GPU is a versatile "jack of all trades," while the TPU is a specialized "special forces soldier." This fundamental difference, like its very DNA, is deeply embedded in every aspect of their design, from chip architecture to circuit board design. This article will delve into the differences between these two processors in their core mission, internal architecture, and even the final printed circuit board design and assembly, revealing why GPUs are the cornerstone of electronic design automation, while TPU circuit boards are sophisticated modules tailor-made for cloud-based AI training.

Graphics Rendering vs. AI Training

GPU:

Design Purpose: To handle large-scale, highly parallel graphics rendering computations. Graphics rendering involves numerous matrix operations, floating-point calculations, and geometric transformations.

Relevance to PCB Design: The graphical interface display, real-time rendering, 3D model viewing, and layer management of PCB design software (such as Altium Designer, Cadence Allegro, and Mentor PADS) are essentially graphics computations. When you zoom, pan, and rotate a complex PCB with dozens of layers and millions of components, it's the GPU that's performing the real-time computations and rendering of the images you see. Furthermore, some simulation tasks (such as post-processing and visualization of electromagnetic and thermal simulations) also rely heavily on the GPU's parallel computing capabilities.

TPU:

Design Purpose: A tensor processor specifically designed for neural network inference and training. It is highly optimized for specific large-scale, low-precision (e.g., INT8) matrix operations such as "multiply-accumulate."

Relevance to PCB Design: Traditionally, the PCB design process has rarely required large-scale neural network computations. While the TPU excels in processing AI models, PCB design primarily relies on deterministic computations based on geometry, electrical rules, and physical laws.

Architectural Differences

GPUs, like NVIDIA's, are general-purpose parallel computing beasts. They not only feature thousands of computing cores, but also a complex memory hierarchy, caches, and numerous high-speed interconnects.

PCIe interface: Communicates with the CPU.

NVLink: Direct high-speed communication between multiple GPUs requires a large number of additional ultra-high-speed differential pairs.

Memory interface: Modern high-end GPUs use HBM, connecting to the GPU core via Silicon Interposer, which manifests as a massive substrate package on the PCB.

PCB design impact: The PCB for a GPU is more like a "transportation hub," requiring the routing of a wide variety of high-speed signal lines. This requires extremely high layout and routing density, and a high number of layers (often over 20).

TPU: Google's TPU is a type of ASIC, tailored for matrix multiplication. Its architecture is relatively more focused.

It also communicates with the host CPU (via PCIe).

But its core is the ultra-high-speed connection between the TPU core and HBM memory. The control logic and interconnect within the TPU are likely to be simpler than those in a GPU, dedicating more transistor resources and power budget to the compute units and memory bandwidth.

PCB Design Impact: While similarly high-density, the interconnect types are likely to be fewer, with a clear focus on ensuring the data paths between the compute units and memory are extremely wide and fast.

Designing PCBs for both TPUs and GPUs challenges the physical limits of high-speed PCB design.

A GPU PCB is more like a fully functional supercar, required to handle a variety of complex "driving conditions" (general computing tasks). Its interconnect network is extremely complex and requires the utmost flexibility.

A TPU PCB, on the other hand, is like a drag racer designed to break a single record. Its single, clear goal is to complete matrix calculations with the highest energy efficiency. While its design allows for more extreme and focused refinement in certain areas, it still faces the same challenges of power consumption, heat dissipation, and signal integrity.

GPU PCBA vs. TPU PCBA

1. Form and Integration: Standard Parts vs. Custom Bricks

GPU PCBA: Despite being a top-tier product, it still requires certain industry compatibility considerations. For example, the PCIe version of the GPU can be plugged into any standard server motherboard. Even the SXM form factor adheres to NVIDIA's own specifications, allowing integration with various server vendors (such as Dell and HPE). This makes its PCBA a relatively complete and independent product.

TPU PCBA: This is tailor-made for Google's data centers. It is not sold separately, but rather as an integral part of the entire server rack. Its form factor, connectors, and power supply interface are likely highly customized to achieve the highest density, energy efficiency, and maintainability at the data center level. It's more like a Lego brick, designed specifically for Google's "baseboard."

2. Components and Complexity: Full-featured vs. Feature-focused

GPU PCBA: To achieve versatility, GPU chips integrate a variety of specialized units. This translates to the need for peripheral circuits and components supporting these functions within the PCBA. For example, the interfaces and signal conditioning components used for NVLink interconnects increase the complexity and component count of the PCBA.

TPU PCBA: As an ASIC, its functionality is focused. Google can "subtract" by removing all hardware logic and peripheral circuitry that is irrelevant or marginally relevant to the core matrix computations. This potentially simplifies the TPU PCBA's bill of materials (BOM) in certain aspects, allowing for a more focused investment of cost and area on compute and memory bandwidth.

3. Interconnect: Hardware Interconnect vs. System-Level Interconnect

GPU PCBA: Multi-card interconnection relies on NVLink, an electrical interconnect technology implemented on the PCBA. This technology requires PCB traces and dedicated bridges, offering short distances and high bandwidth.

TPU PCBA: Google emphasizes node-level interconnection. For example, the TPU v4 utilizes optical circuit switching technology. This means the TPU PCBA may require an integrated optical engine or a dedicated high-speed electrical interface to connect to it. This shifts the interconnection task from the board level to the rack level, enabling greater flexibility and scalability.

In short, from the macroscopic architectural mission to the microscopic PCBA design, the differences between GPUs and TPUs reflect the distinct distinction between "general-purpose flexibility" and "specialized efficiency." The GPU's circuit board is like a fully functional "superhub," dedicated to handling a wide range of complex computing tasks, with its design and ecosystem centered around versatility. The TPU's circuit board, on the other hand, is more like a deeply customized "efficient engine," where every component, trace, and thermal design serves a single goal: achieving extreme performance and energy efficiency in AI training. Understanding this difference not only helps us understand the causes of the current computing landscape but also foresee the inevitable trend of continued hardware customization in specific areas.

One-Stop HDI PCB Manufacturer and Its PCB Via Filing Capabilities

If you're looking for turnkey HDI electronics manufacturing services (EMS) from hardware development to PCBA fabrication and box-build assembly, you can work with the one-stop HDI PCBA manufacturer PCBONLINE.

Founded in 1999, PCBONLINE has R&D capabilities for HDI projects and EMS manufacturing capabilities, including via filling for stacked vias. It provides 4-to-64-layer HDI PCB fabrication, assembly, and PCBA box-build assembly. You can order various HDI PCBs from PCBONLINE, such as FR4, polyimide (flexible PCB), polyimide + FR4 (rigid-flex PCB), and PTFE/Rogers (high-frequency PCB).

Here'e the PCB via filing capabilities at PCBONLINEL:

• Micriavia filling with copper: laser via size 0.1-0.125mm, priority 0.1mm
• Finished hole size for via-in-pad filling with resin: 0.1-0.9mm (drill size 0.15-1.0mm), 0.3-0.55mm normal (drill size 0.4-0.65mm)
• Max aspect ratio for via-in-pad filling with resin PCB - 12: 1
• Min resin plugged PCB thickness: 0.2mm
• Max via-filling ith resin PCB thickness: 3.2mm
• Making different hole sizes with via filling in one board: Yes
• Via filling with copper/silver: Yes

If you need HDI PCBAs or any other PCBAs requiring via filling, please send your email to PCBONLINE at info@pcbonline.com. We will provide one-on-one engineering support to you.

Conclusion

Via filling is used for creating stacked vias in HDI PCB fabrication, BGA/CSP/QFN IC packaging, and filling PCB via-in-pad with resin during multilayer PCB fabrication. If you need one-stop electronics manufacturing for your HDI PCBA project, contact the one-stop advanced PCB manufacturer PCBONLINE for high-quality PCBA and box-build solutions tailored to your project's needs.