Google Tensor Processor: A Brief History & The Road To 3nm

A Brief History of Google’s Tensor Processor

Pin

Google’s journey into crafting its own system-on-chip (SoC) kicked off in April 2016, post-launching their inaugural Pixel smartphone. 

Sundar Pichai and Rick Osterloh, the big guns at Google, knew it was a long game.

Fast forward to 2017, Google’s hardware division is buzzing. A team of 76 semiconductor aficionados, AI and ML specialists, are all hands on deck, working on the chip. 

The Pixel series starts flaunting custom-designed co-processors. The Pixel Visual Core and the Pixel Neural Core make their debuts.

By April 2020, the buzzword is “Whitechapel”. Google’s custom ARM-based processor is in the works, aimed at Pixel and Chromebook devices. 

Pichai’s excitement over Google’s “deeper investments” in hardware in an October investor call fuels speculations about Whitechapel.

Come 2020, the Pixel 5 drops the Neural Core. Why? The Snapdragon 765G SoC had it covered, hitting the camera performance targets Google was aiming for. 

And in April 2021, the word is out. 

Whitechapel is set to power Google’s upcoming Pixel smartphones.

August rolls in, and Google pulls the curtain on Tensor, as a sneak peek into the Pixel 6 and Pixel 6 Pro smartphones. 

It’s a shift from the Qualcomm Snapdragon chips, marking 2021’s Pixel 5a as the last of its kind. 

Pichai hints at more off-the-shelf solutions for Pixel phones, thanks to the development of Tensor and the Pixel 6. 

But, September 2022 brings some blues. A Tensor-powered successor to the Pixelbook laptop, slated for a 2023 release, gets the axe due to cost-cutting.

Design

Pin

“Tensor” is a nod to Google’s TensorFlow and Tensor Processing Unit techs. 

Developed by the Google Silicon team and the Google Research division, it’s a brainchild of Phil Carmack and Monika Gupta.

Tensor’s design is a break from the norm. It packs two large cores, two medium, and four small ones, a rare combo for octa-core SoCs. 

Carmack spills that it’s all about keeping Tensor efficient during intense workloads, managing co-processors by running both large cores simultaneously at a low frequency. 

Performance metrics were low compared to Apple and Qualcomm’s chips, but this wasn’t the point. 

Tensor’s performance isn’t about synthetic benchmarks, according to Google

Tensor is about the ML capabilities it unlocks—advanced speech recognition, real-time language translation, unblurring photographs, and HDR-like processing for videos, frame by frame.

Tensor Capabilities & Features – It’s Main USPs Unpacked

Pin

• Enhanced User Experience:
  • The chip enables more personalized and intuitive user experiences, with phones adapting and becoming more helpful over time.
  • The AI and ML capabilities allow Pixel phones to understand user context, preferences, and unique needs.
• Photography Advancements:
  • Google Tensor aids in capturing better photos by understanding and adjusting focus and brightness according to the subject.
  • Features like Face Unblur and Magic Eraser allow users to rectify blurry faces and remove unwanted items from pictures without losing background details.
• Communication Efficiency:
  • It facilitates faster communication through features like Live Caption and Live Translate, translating languages in real-time, either speech-to-text or directly to audio.
• Power Efficiency and Data Protection:
  • Despite running advanced ML models, Google Tensor consumes less power compared to previous Pixel phones.
  • It includes a separate core for sensitive tasks and controls to run in an isolated environment, making it resilient to attacks.
  • The Tensor Security core, combined with the next-gen Titan M2™ co-processor, adds an additional layer of protection for personal information.
• Adaptive Learning and Suggestions:
  • Google Tensor helps Pixel phones learn user behaviors and preferences, providing relevant information and suggestions when needed.
  • The Pixel’s always-on display, powered by Google Tensor, keeps you informed about appointments and detected songs through Now Playing song detection.

The First Google Tensor Processors

Pin

Out of the gates, Google’s first Tensor processor did not blow the doors off the competition. But it did pack in plenty of smarts that Apple and Qualcomm’s lacked. 

So, while the performance was a bit iffy, there were plenty of “extras” to get excited about. 

Here’s how Google’s first ever Tensor processor, dubbed “Whitechapel”, compared to its second-generation Tensor G2 processor. 

Google Tensor G1 vs Google Tensor G2

Launch Date:

• Original: October 19, 2021
• G2: October 6, 2022

Model Number & Codename:

• Original: GS101 (Whitechapel)
• G2: GS201 (Cloudripper)

Both chips are made by Samsung, using its 5 nm fabrication process.

CPU:

• Original: It’s an octa-core beast, featuring 2.8 GHz Cortex-X1, 2.25 GHz Cortex-A76, and 1.8 GHz Cortex-A55.
• G2: The successor ups the ante with a 2.85 GHz Cortex-X1, 2.35 GHz Cortex-A78, and retains the 1.8 GHz Cortex-A55.

Security:

TrustZone with Trusty OS is the guardian angel for both, ensuring top-notch security.

GPU:

• Original: Mali-G78 MP20 operates at 848 MHz.
• G2: Enter Mali-G710 MP7. The frequency? It’s still under wraps!

RAM:

LPDDR5 is the common thread, ensuring swift operations and multitasking.

Storage:

UFS 3.1 is the storage type, offering high-speed read and write capabilities for both.

Connectivity:

• Original: It’s equipped with an Exynos 5123 modem.
• G2: The modem gets a boost with Exynos 5300.

Wireless & Navigation:

Wi-Fi 6 and Wi-Fi 6E, Bluetooth 5.2, and Dual-band GNSS are constants, ensuring seamless connectivity and precise navigation.

ISP NPU:

The edgeTPU is the brain behind both, handling image signal processing with finesse.

Tensor G2 Versus The Competition

Tensor G2’s Reception

Pin

Performance has never been Tensor’s strong suit, not on the first iteration nor the G2. 

But as Google constantly likes to point out: Tensor is NOT about benchmarks. 

And reviewers and critics seemed to understand this when they got their hands on the Pixel 7 and Pixel 7 Pro. 

[snippet]

Here’s some choice comments made about the first iteration of Google’s Tensor processor inside the Pixel 6 and Pixel 6 Pro: 

• Philip Michaels from Tom’s Guide was all in, giving a nod to the Pixel 6 and Pixel 6 Pro’s Tensor-driven features and video enhancements. And guess what? Marques Brownlee and Wired’s Julian Chokkattu were on the same page!
• But it doesn’t stop there. Lily Hay Newman, also from Wired, spotlighted the chip’s security capabilities. She didn’t hold back, calling them Tensor’s top selling point!
• CNN Underscored’s Jacob Krol didn’t shy away either, stating Tensor brought “some of the most fluid and fastest performance” to the smartphone game. However, not everyone was sold. Android Authority’s Jimmy Westenberg seemed to be on the fence.
• Here’s the kicker. Ryne Hager from Android Police felt the chip did the job for the everyday user but threw a curveball. He wished for more years of Android updates from Google, especially since they’ve broken free from Qualcomm’s chains.
• TechRadar’s James Peckham didn’t mince words, labeling Tensor a “standout feature.” However, his colleague, David Lumb, threw in a twist, describing the chip’s performance as “strong but not class-leading.”

I tested the Pixel 6 Pro and the Pixel 6. I also used the Pro model as my daily driver until, well… things took a turn for the worse.

Tensor G3

Pin

Tensor G3

Google’s Tensor G3 chipset debuted inside the Pixel 8 series, the Pixel 8 and the Pixel 8 Pro.

Picking up where its Tensor G2 chipset left off, the G3 brings a host of performance improvements and doubles-down on machine learning and AI smarts in general – something Google is heavily investing in right now. 

The Tensor G3 marks a departure from traditional configurations, adopting a 1+4+4 architecture. This setup includes:

• 1x Cortex-X3: A high-performance core optimized for the most demanding tasks, ensuring swift and efficient processing.
• 4x Cortex-A715: Mid-range cores designed for a balance between power efficiency and performance, perfect for everyday applications.
• 4x Cortex-A510: Low-power cores that manage background operations with minimal battery consumption, enhancing overall device longevity.

Improvements & Upgrades Over Tensor G2

• GPU Enhancements: The Tensor G3 chip’s Immortalis-G715 GPU delivers a performance increase of 15% compared to the prior year’s Mali-G710 model. But it is not the latest and greatest, getting launched in 2022. ARM’s Immortalis-G720 offers yet another 15% performance boost and superior power management, but this GPU was not used by Google.
• Manufacturing Mastery: Samsung fabricates Google’s Tensor chipsets and with the Tensor G3, Samsung Foundry’s new 4nm process was used. This is still a ways behind TSMC’s 3nm process used for chips found in Apple’s iPhone, but it is a step up for Google’s Pixel 8 phones; the Pixel 7 series used 5nm chips, for reference. 
• Storage Speeds: Integration of UFS 4.0 storage technology ensures rapid data access and transfer, making loading times and app performance snappier than ever.

Google’s Tensor G4 CPU

Pin

When the Google Pixel 9 series launches later this year, it will run on Google’s new silicon – the Tensor G4. This next-generation system-on-chip (SoC) represents a major step up in performance and efficiency over the Tensor G3.

The headline feature of the Tensor G4 is its 4nm manufacturing process, which will once again be handled by Samsung Electronics’ foundry business. And this new process, which employs something known as FOWLP, is expect to yield impressive results with respect to performance and power management.

Fan-out wafer-level packaging (FOWLP) allows for significantly improved heat dissipation compared to traditional methods. With FOWLP, the Tensor G4 can run at higher clock speeds and voltages while still maintaining optimal thermal limits.

Here’s a breakdown of how FOWLP works:

• Chip dies are placed face-down on an adhesive film without a carrier substrate.
• The dies are then encapsulated in a mold compound for protection.
• A re-distributed layer (RDL) of wiring routes “fans out” from the original die pads, enabling more external connections.
• A ball grid array (BGA) of solder balls is added to the bottom to interface with a device’s circuit board.
• The package receives protective coatings and final testing.

By combining Samsung’s 4nm process with FOWLP’s thermal advantages, the Tensor G4 should provide a considerable uplift in CPU and GPU performance for the Pixel 9 compared to the previous Tensor G3.

Tensor G5 Moving To 3nm Process

But wait, there’s more! Google wasn’t about to sit still. They shifted gears and jumped straight into the development of Tensor G5, giving it the codename “Laguna.”

And here’s the twist: it’s being designed fully in-house, leaving Samsung in the rearview mirror and partnering with TSMC for manufacturing.

And the cherry on top? It’s being built on TSMC’s groundbreaking 3 nm process which is the same process it used to make Apple’s new A17 CPU.