Newshot

US startup unveils Frostbite, an AI-designed CPU architecture with 4.59x the energy efficiency of conventional processors

AI summary

Architect Labs, a startup based in Palo Alto, California, announced in a research blog post that its AI system autonomously discovered, designed, and verified a general-purpose CPU called Frostbite, which uses a new instruction set redesigned from the ground up to escape some of the architectural limits of traditional CPUs. In 25 workload tests, a Frostbite cluster of 4 compute tiles reached 4.59x the performance per watt of a similarly sized, silicon-validated dual-issue in-order CPU core, though the data comes from a simulated design based on a 7nm process and Frostbite has not yet been taped out, so no physical chip exists to verify it; for legal reasons, Architect Labs cannot currently disclose the model of the comparison CPU core.

Translation(model translation, not the source's own text)

US startup unveils Frostbite, a self-designed AI CPU architecture with 4.59x the energy efficiency of conventional processors

IT之家 November 22 news: Architect Labs, a startup based in Palo Alto, California, announced in a research blog that its AI system has autonomously discovered, designed, and verified a general-purpose CPU named Frostbite. The processor uses an entirely new instruction set and was redesigned from the ground up in an attempt to escape some of the architectural limitations of conventional CPUs.

Across 25 workload tests, a Frostbite cluster composed of 4 compute tiles achieved 4.59x the performance per watt of a similarly sized, silicon-validated dual-issue in-order CPU core. However, these figures come from a simulated design based on a 7nm process; Frostbite has not yet been taped out, so there is no actual chip to verify. For legal reasons, Architect Labs is currently unable to disclose the model of the CPU core used as the comparison baseline.

Conventional CPUs must determine dependencies between different instructions at runtime through hardware, re-identifying which instructions can execute in parallel. Multicore processors also require complex multi-level cache hierarchies and cache coherence protocols to ensure that every core sees a consistent view of shared memory data. However, these mechanisms not only consume chip area and power but also face declining scaling efficiency as processor scale grows.

IT之家 notes that Architect Labs' proposed alternative is called "Multithreading for Free." Its core idea is to have the program itself directly expose work that can be executed in parallel, rather than relying on conventional threads, lock mechanisms, or out-of-order execution hardware to find parallelism opportunities during runtime.

Frostbite consists of multiple compute tiles, each a small processing unit with an identical structure, interconnected through a network-on-chip and each possessing its own small local memory.

In this performance test, the company used a compute cluster composed of 4 tiles and 4 L3 nodes for comparison with the conventional CPU core. However, the complete Frostbite design actually contains 16 tiles.

During the early synthesis stage, the processor's design frequency reached 800MHz, but in the final layout design it dropped to 500MHz. In addition to achieving 4.59x the performance per watt of the comparison processor, the Frostbite cluster also achieved 8.77x the performance per unit area of the comparison core.

Specifically, the energy efficiency gains of Frostbite vary significantly across different workloads:

Test itemFrostbite performance per watt improvement
All 25 tests (geometric mean)4.59x
10 tests used for design tuning6.02x
15 tests introduced only in the final evaluation stage3.83x
Best performer: conv2d (not involved in tuning)23.12x
Worst performer: recursive Fibonacci (not involved in tuning)1.15x

Frostbite performance per watt improvement

All 25 tests (geometric mean)

10 tests used for design tuning

15 tests introduced only in the final evaluation stage

Best performer: conv2d (not involved in tuning)

Worst performer: recursive Fibonacci (not involved in tuning)

It is worth noting that the AI system involved in designing Frostbite initially did not adopt any preset processor architecture philosophy. Architect Labs calls this state "Zero Architectural Priors." However, the team still determined in advance the workloads that the processor needed to focus on optimizing.

Subsequently, multiple AI agents read academic papers and performance testing materials, built Python models, and selected an architectural direction from which to proceed with the design. The initially generated design document was 141 pages long, covering the programming model, runtime environment, and high-level hardware modules.

But after the first version of the register-transfer level (RTL) design was generated, this PDF document was deleted. The company explained that many of the numbers and estimates in the document were far from what was actually achievable in physical implementation.

This document was only ever a design starting point. After determining the architectural direction, the AI agents no longer relied on English to describe hardware specifications, because the team believed that natural language cannot accurately express all details and makes errors difficult to spot.

Instead, the AI used ASL, a hardware description language developed by Architect Labs itself, to write the specifications. ASL can be converted into formal expressions in the Lean proof system, which can then be machine-verified. The company believes this approach can eliminate a large number of potential errors while also reducing the possibility of AI agents exploiting loopholes in the rules to meet optimization goals.

On this basis, the AI agents also developed Arcmodel. This is a software model written in ASL that can simulate the runtime behavior of the entire chip, and it was completed before actual circuit design began.

Because directly simulating RTL is slow, the AI agents chose to test design proposals on Arcmodel first. The model runs at approximately 200x the speed of RTL simulation.

However, the initial power estimates from the model were inaccurate, with a maximum deviation of nearly 250%. The system then continuously adjusted the model through automated iteration, ultimately narrowing the deviation between its power estimates and the measured results of the actual circuit design to within 5%.

After completing the foundational model, a large number of AI agents began proposing design optimization proposals in parallel. Each modification had to pass specification verification, fast model testing, circuit design implementation, and power measurement; only proposals that passed all checks were retained.

Throughout the process, the AI agents tried a total of 58 optimization proposals and ultimately retained 37.

In addition, the system unexpectedly exhibited a behavior that human designers had never explicitly specified, namely so-called "Work Stealing": when a tile has no tasks to execute, it can take unfinished work from a neighboring tile; if no tasks are available, it enters an idle state to save power.

In terms of energy consumption distribution, Frostbite also shows a clear difference from conventional CPUs. According to the company's average measurements across 20 workloads (excluding clock distribution energy), the energy consumption breakdown of the two is as follows:

Energy consumption itemConventional CPUFrostbite
Memory57%38%
Instruction fetch, decode, and issue27%6%
Network-on-chip and schedulerNot separately counted26%
Clock distribution (share of total energy)49%41%

Clock distribution (share of total energy)

From the data, Frostbite reduced the share of energy consumed by instruction fetch, decode, and issue from 27% to 6%, but the network-on-chip and scheduler also introduced new energy overhead, accounting for 26%.

This is not the first time Architect Labs has demonstrated an AI-involved processor design. In August this year, the company also unveiled Redwood, an AI inference accelerator aimed at robots and drones.

For Redwood, Architect Labs said that two human architects only needed to provide design specifications, and the AI could generate RTL code, verification programs, firmware, drivers, and compute kernels in less than two weeks. However, Redwood currently also runs only on FPGA, and no actual chip has been manufactured yet. The company says it is advancing a tape-out plan in cooperation with TSMC.

Meanwhile, Tendrils Compute, a startup headquartered in Cambridge, UK, is also developing a processor in a similar direction. Earlier this month, the company completed a new round of financing and said it is developing a processor that can enable "thousands of cores to run in parallel." In addition, Tendrils Compute is also developing a programming language called Vine, and Frostbite's programs are likewise written in this language.

Architect Labs said it will publish more performance test results in the future, but this is only part of its plan. The company also hopes to launch an online FPGA-based software experimentation platform so that anyone can run Frostbite programs, along with a preliminary version of an operating system. Ultimately, the company plans to publicly release Frostbite's RTL design code and instruction set.

On the performance evaluation side, Architect Labs also plans to expand the range of test workloads and introduce more CPUs as comparisons, including comparing a larger-scale Frostbite mesh architecture with multicore versions equipped with large-capacity caches.

In addition, the company plans to compare with larger-scale out-of-order CPUs with wider instruction execution. The related tests are expected to be conducted in the coming months, using mainstream CPU benchmark suites such as SPEC and PARSEC.

It should be noted that the performance advantages currently announced for Frostbite still come from a simulated design and have not yet been verified on actual silicon chips. Whether it can maintain these energy efficiency advantages after actually being manufactured will require subsequent tape-out and measured results to provide the answer.

Related articles

  • Developer squeezes a tiny Transformer model onto a floppy disk, letting a 486-era PC run AI
  • New York's new concert hall Domino's official website becomes a joke, promoting "handmade" while being full of AI-generated traces
  • IMF Managing Director warns: excessive AI investment enthusiasm could lead to further adjustments in financial markets
  • Who is responsible when AI agents transfer money or make purchases without authorization? Multiple banks warn users may have to bear the losses themselves
  • JPR data: global PC client CPU shipments in 2026Q3 down 6.5% year-over-year and 3.4% quarter-over-quarter
  • Adobe unveils four AI experimental projects: photos instantly become 3D scenes, and AI characters can be designed through chat

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗