What does the future hold for Bitcoin mining? As an industry that has made a 1000X power efficiency improvement over the past decade, can it make further breakthroughs? What are the theoretical physical limits of current ASIC designs?
Transcript:
Hey, it is great to be on a stage in Nashville. My name is Haris Basit, and I am the Chief Strategy Officer for Bitdeer.
Bitdeer is a company with a global presence. Our largest footprint is here in the United States. In fact, one of our facilities is right here in Tennessee, a couple of hours from here in Knoxville. We have our headquarters in Singapore, and also have large facilities in Europe and, more recently, South Asia.
The thing we are really known for, and the thing we are focused on, is technology.
We are one of the largest Bitcoin mining companies with over 22 Exahash currently under management. We are utilizing nearly 900MW of power today but have a total worldwide power capacity of 2.5GW and are planning for significant growth in mining over the next 18 months.
Among all the publicly traded bitcoin mining companies, we are the only one that focuses on technology. Indeed, we have 25% of our workforce dedicated to research and development.
And I‘m going to talk about the advanced technology that we’re working on at Bitdeer.
Bitdeer recently announced its ASIC Roadmap as shown here.
Our plan is to release new chips in a fast and furious manner with new designs coming out about every other quarter—much faster than foundries release new manufacturing processes.
Of course, in Bitcoin mining, energy efficiency is by far the most important metric. Therefore, each new chip must include significant design innovations to improve energy efficiency and not just rely on improvements from the foundry.
By the second half of 2025, we will release mining rigs with energy efficiencies at or below 5J/TH—3 times better efficiency than the best rigs available today.
Michael Saylor says that “Bitcoin is Digital Energy”.
Worldwide, we use tens of gigawatts of power, but where does that power actually go and how is it consumed?
And how much more efficient can we get? What are the physical limits?
And perhaps most importantly, can the techniques we develop for energy efficiency benefit the non-Bitcoin world such as in AI and other fields?
Let’s look at this much more closely.
If we start with 100MW at a substation located right next to a datacenter we can expect 80% of that power to be actually used by the computing, networking, and data storage equipment. The remaining 20 MW is “wasted” in cooling and power distribution within the facility.
I should note that Bitcoin mining datacenters are quite a bit more efficient in this regard than HPC and AI datacenters because we can get away with much less cooling.
Now, if we drill down one level deeper -- within the Compute servers (or ASIC Mining Rig in our case) about 90% of the power is utilized by the chips, and 10% is “Wasted” again for cooling and power distribution – the same type of waste as in the data center itself.
But our goal here is to dig even deeper and understand how the chip itself is consuming power. Here we see that the efficiency is only about 65%. About 35% of the power the chips receive is wasted due to Clocking, Leakage and power distribution.
Therefore, out of the full 100MW delivered to the facility, only 47MW is doing something we consider useful—switching transistors and thus calculating the SHA256 algorithm upon which Bitcoin is based.
Focus on that switching energy because that is the whole goal of Bitcoin mining and where it turns out we can make the most impact. We will come back to it again throughout this talk. But first, I want to take a detour into the history of Bitcoin mining because that history is pretty spectacular.
Going back in history here, the Genesis block was mined by Satoshi Nakamoto on January 3, 2009, and this started mining with what we are calling here the “CPU Era” Within two years, we saw the birth of the GPU Era, which was about 10 times more efficient.
Then, within one year after that, we transitioned to FPGAs, which gave us yet another factor of ten improvement in efficiency.
The biggest step, however, came in 2013 when we started using Application Specific Integrated Circuits.
Chips specifically designed for bitcoin mining and saw another efficiency improvement – this time of 50 X.
So, in these first 5 years, we saw an overall efficiency improvement of 5000 times.Let that sink in – 5000X in five years. Not bad.
So, let’s look more closely now at the ASIC Era – it has only been a little over 10 years – but the industry has managed to get another 1000X improvement in efficiency.
We have divided the time into phases that represent each silicon manufacturing process.
We started in 2013 with 110nm processes, and we are now down to 4nm and soon 3nm.
You can also see that within each manufacturing process there are improvements to efficiency.
For example, within the 28 nm process, a full-custom design was introduced which allowed significant efficiency improvement without depending on process improvements.
Now let’s focus on the right side of this chart where the orange points are Bitdeer’s current roadmap.
As you can see over the past few years the trend line appears to be flattening. Bitdeer’s intent is to get well below the trend line shown here through the use of novel design techniques.
Again, the key is in the design – not the manufacturing process alone.
So, let’s talk about chip design.
This is a bit of an oversimplification, but I think it helps in understanding.
SHA256 which is used for Bitcoin mining is a purely digital function, so one would think that an entirely digital design process would be ideal. Indeed, digital design allows you to use high-level building blocks provided by the foundry and use software to synthesize the entire solution. It requires the least effort and can be done by ordinary chip designers. Indeed, this technique is what virtually every digital chip outside of Bitcoin uses. However, if you use this technique today for bitcoin you will get very poor energy efficiency.
So virtually every Bitcoin chip designed today and over the past 9 years uses Analog Design techniques extensively to reduce energy consumption. Here we cannot use the high-level logical functions provided by the foundry and have to deal with transistors and wires and voltages and currents. This is called Full-Custom design and takes far more effort and skill than digital design but is necessary to get the energy efficiency we need.
However, there are even more aggressive chip design techniques. These are currently only used by RF or microwave design engineers but involve thinking of the basic electromagnetic waves and distributed elements like transmission lines. These circuits have to be entirely hand-crafted using designers with yet even more skill and labor than analog but allow for circuits that cannot be made any other way.
The question arises, “Why doesn’t everyone use analog techniques for their digital chips?” The reason is that it would take far too much time and effort – the chip would never see the light of day.
However, Bitcoin chips are different because they have a very large amount of hierarchy. While a full chip may be 1B transistors it is made up of hundreds or even a thousand SHA256 cores each with 1 million transistors. Each SHA256 core in turn is made up of about 128 pipeline stages.
So, we can focus all of our efforts on a relatively small 10 thousand transistor design and then cut and paste that to make up the full 1B transistor chip.
Up until recently, there weren’t other applications that were extremely energy sensitive and also had similar hierarchy. However, now we have GPUs for AI which are becoming increasingly energy constrained and have similar hierarchy.
This is why we feel that a lot of the approaches we are taking with Bitcoin today can be applied to AI tomorrow.
OK, now as promised I want to come back to Switching Power. This is the power consumed when transistors switch between on and off states.
What we are seeing here is an example of a chip where some nodes are one and shown in green and the rest are zero and shown in black. From one moment to the next different nodes become 1 or 0 but on average almost exactly 50% of the nodes are 1 at any given time.
Every time a node goes from 0 to 1 charge needs to be brought in from the power supply. And when that node goes from a 1 back to a 0 that excess charge is flushed to the ground. There is no recycling of the charge here–the charge is used once and then flushed.
The total charge in the chip hardly changes—it is always about 50% ones–however we do not just simply recycle charge by moving it about the chip. We always bring in new charge. This is why the energy efficiency of electronics in general (not just bitcoin) has been ruled by this one equation (shown here) since the time President Nixon started his second term in the early 1970s.
If we could move the charge from where it is not needed to where it is needed without bringing new charge from the supply. We could recycle the charge. This would be a huge breakthrough. But can it be done in a practical way?
So, yes indeed we can move charge around a chip without having to always get new charge from the power supply.
What you see here are two examples of charge moving around a chip with only minimal power coming from the power supply.
In this case the switching power can be 75% less than what the equation that has ruled over us for the past 50 years predicts.
I do want to acknowledge that we are standing on the shoulders of giants here. The equations that govern the circuit on the right originally were developed by Christian Huygens in the sixteen hundreds to explain the behavior of his invention, the Pendulum clock.
Besides switching power, Leakage Power is another type of power that can dominate in bitcoin mining (and other chips), especially at high temperatures.
Leakage is a quantum mechanical effect where an electron can tunnel through a very thin barrier that should normally be able to block it.
As transistor gates get narrower the ability for electrons to tunnel through increases.
So, tunneling and leakage power become increasingly problematic for newer processes.
Leakage power also increases exponentially with temperature and is the primary reason ASICs are less efficient at very high temperatures.
It is also why there is significant benefit from immersion or liquid cooling – since those techniques reduce maximum temperatures.
The foundries have been trying to hold the line on leakage power by using FinFET technology and in the near future RibbonFET or Gate All Around technology.
However, distributed design and charge recycling techniques can significantly reduce this type of power loss as well -- thereby making us much less dependent on newer manufacturing processes.
I would like to end this talk by saying we are just getting started.
The era of distributed circuit design and charge recycling has just begun.
Bitcoin and Bitdeer are leading the way.
I think that these techniques are going to be very applicable to the GPUs that are used for AI. The savings in power that we can get by applying it to much larger applications like AI might actually more than offset the amount of power we consume as Bitcoin miners.
So in the end, it could be that Bitcoin has a net negative energy usage if you attribute that savings to Bitcoin. Thank you.
本文提供的资讯仅供一般参考,并不构成也不应被理解为任何产品的广告、专业意见、要约、招揽或建议,也不涉及任何产品的交易。对于任何资讯的公正性、准确性、及时性、完整性或正确性,以及任何产品的未来回报、表现或结果,Bitdeer不作出任何明示或暗示的保证、陈述、担保或承诺。在适用法律允许的范围内,Bitdeer明确排除与本文提供的资讯有关的任何和所有责任,并且在任何情况下,Bitdeer都不对任何人因依赖本文的任何资讯而产生的任何损失或遭受的损害负责。