Stacking GPUs Like Memory If Heat Problems Are Removed

Vaire Computing has raised a total of $4.5 million in funding. They aim to produce useful reversible computing chips in 2027.

They ihad a seed round of $4 million, which was announced on July 1, 2024, bringing the total funding to $4.5 million when combined with previous pre-seed funding of $500,000.

IBM, Intel, MIT, Sandia and U of Florida have long histories researching reversible computing.

Reversible computing offers a path to continued performance improvements. It could help overcome the limit of physics (Landauer Limit) that will eventually limit further miniaturization and speed increases in conventional computing.

Reversible computing has the potential to achieve energy consumption significantly below the Landauer limit, which is the theoretical minimum energy required for irreversible computation. Here’s a breakdown of the potential energy savings:

Theoretical Potential
Reversible computing could theoretically reduce energy consumption to arbitrarily low levels, even approaching zero. This is because reversible operations do not erase information.

Practical Considerations
While the theoretical potential is immense, practical implementations face challenges:

Orders of Magnitude Improvement: Reversible computing could potentially achieve energy consumption many orders of magnitude below the Landauer limit

Approaching Zero: In principle, for bit operations performed using high-quality reversible transformations, the energy dissipated per operation can be made arbitrarily small – much smaller than kT ln 2 (0.69 kT), which is the Landauer limit

No Absolute Minimum: Unlike irreversible computing, reversible computing does not have an absolute minimum heat output, allowing it to potentially surpass the limitations imposed by the laws of thermodynamics on conventional computing.

Challenge with Reversible Computing?

It is not just that the logic gates we are currently using are irreversible. Most calculations today are also fundamentally irreversible. As an example, consider adding two integers, A and B. If you just want to know the result, A+B, it’s impossible to know which integers A and B were added together—there are infinitely many combinations of integers whose sum is A+B.

To be reversible, you need to calculate and store some extra information. The result of reversible addition would be A+B and A—storing the input A in addition. With this additional information, you can find out what the original A and B were, but the price is that you have to store A, which you’re not interested in, and that’s why it’s called ‘garbage.’

Reversible computing operation.

As the calculation continues unchecked, you accumulate more and more garbage until your memory overflows. If you override memory cells and thereby delete information, you’re expending energy, so you’re not saving energy overall. That’s why Rolf Landauer dismissed reversible computing immediately as impractical when he researched heat generation in computing. But there are ways to get around this problem.

How Do You Make Reversible Computing a Reality? Calculate Twice

The trick to avoiding the memory overflow problem in reversible computing is to do every calculation twice: First, you perform the calculation forward, storing all the bits and never wiping out information. Once you arrive at the final answer, you make a copy of it. Then, you run the entire calculation backward to get to where you started and eliminate all the garbage. This also means you only need to store one step backward.

You might think, wait a minute, that’s actually twice the computational effort. Is this really worth it just to save some energy?

First, doing the calculation forward and backward is the simplest approach to dealing with garbage. Over time, researchers discovered more sophisticated ways to reuse garbage bits in future computations, thereby reducing their number by orders of magnitude.

In the worst case, reversible computing requires twice as much circuitry, but it will allow us to build much larger and, therefore, much more powerful computers.

If a computer could be made entirely of reversible gates and operate completely adiabatically, it would not dissipate any heat from computation. In practice, some energy is always lost due to non-ideal characteristics of real materials and devices. But adiabatic switching of transistors combined with reversible computing will allow us to reduce energy consumption by several orders of magnitudes—more energy efficient than the human brain. That’s why Vaire Computing calls their chips Near-Zero Energy Chips (NZECs).

Without heat problems then many layer chips could be made. HBM memory can have 12-16 layers currently. The industry could see HBM stacks with capacities far beyond current levels, possibly up to 1,000 layers for NAND, although this is more speculative for DRAM. If Vaire succeeds then there could 10-100s of layers of reversible compute GPUs.

Vaire will trade single-core for multicore performance—operating transistors adiabatically impacts single-core performance, but being super energy-efficient will allow us to stack compute layers on top of each other and build truly 3D chips. Microprocessors today are 2D because if you stack them, they will trap even more heat inside that you can’t get rid of fast enough, and they will overheat.

Building 3D chips with many compute layers will be a game-changer. The starting point will be GPUs and machine learning as the first use case.

Challenges and Limitations

Engineering Hurdles: Achieving practical, high-quality, high-performance, and cost-efficient reversible computing below the Landauer limit is considered one of the most difficult engineering challenges of our time

Overhead Considerations: While individual operations might approach zero energy consumption, practical implementations still need to account for factors like error correction, input/output operations, and system-level overheads.

Temperature Dependence: The actual energy savings would depend on the operating temperature, as the Landauer limit itself is temperature-dependent.

Vaire Achievements in Reversible Computing:

Vaire Computing has focused on developing near-zero energy chips using reversible computing. They’ve made strides in creating silicon chips that could consume minimal energy and generate virtually no heat, aiming to address the energy inefficiency of traditional computing.

They’ve hired Mike Frank, a noted researcher in reversible computing, as a senior scientist, enhancing their expertise in the field.

This move also signifies their commitment to advancing the technology from theoretical to practical application.

Vaire has also been selected for the second UK cohort of Intel Ignite, Intel’s global startup accelerator program, which provides them with access to design tools, IP, and prototyping capabilities.

Vaire Computing has outlined an ambitious roadmap for the development and commercialization of its reversible computing technology:

First Prototype: The company aims to deliver its first chip prototype within the next 12 months (by mid-2025)

Test Chip Series: Vaire plans a series of test chips progressing from gates and simple functions to functional blocks, culminating in a software-programmable circuit

22nm Manufacturing Process: The first tape-out is being designed for a 22nm manufacturing process

Market Entry: Vaire expects to have a product in the market by 2027

Initial Focus: The company plans to enter the market with a generic Neural Processing Unit (NPU) chip for edge computing

Long-term Vision: Vaire aims to become the standard for computer processors, potentially replacing traditional chips across various applications

Performance Goals: The company is targeting a chip that is approaching 4,000 times more energy efficient than similar devices implemented classically

Roadmap:

Vaire Computing’s roadmap includes the development of a proof-of-concept for reversible computing, with plans to use their seed funding to hire additional engineering talent and fast-track their first prototype chip. They aim to have their first tape out in Q1 2025, with the goal of achieving full-scale production by 2027.
They envision a future where reversible computing becomes the standard, with their technology potentially leading to a 4000x improvement in energy efficiency over the next 10 to 15 years. This would significantly reduce the energy consumption associated with computing, particularly for AI and data center applications.
Their long-term goal is to extend Moore’s Law by decoupling energy and water resources from computational growth, allowing for continuous exponential growth in computing power without equivalent resource depletion.

Vaire Computing is positioning itself as a pioneer in reversible computing, aiming to revolutionize chip technology for more sustainable and efficient computing solutions.