Optimizing SHA256 in Bitcoin Mining

Nicolas T. Courtois1 , Marek Grajek2 , and Rahul Naik1,3

1
University College London, UK
[email protected]
2
Independent researcher and writer, Poland
3
Royal Bank of Scotland, UK

Abstract. Bitcoin is a “crypto currency”, a decentralized electronic pay-

ment scheme based on cryptography. It implements a particular type of
peer-to-peer payment system. Bitcoin depends on well-known crypto-
graphic standards such as SHA-256. In this paper we revisit the cryp-
tographic process which allows one to make money by producing new
bitcoins. We reformulate this problem as a specific sort of Constrained
Input Small Output (CISO) hashing problem and reduce the problem to
a pure block cipher problem, cf. Fig. 1. We estimate the speed of this
process and we show that the amortized cost of this process is less than
it seems and it depends on a certain cryptographic constant which is
estimated to be at most 1.89. These optimizations enable bitcoin miners
to save countless millions of dollars per year in electricity bills.

Keywords: electronic payment, crypto currencies, bitcoin, hash func-

tions, SHA-256, bitcoin mining, CICO problem (Constrained Input Con-
strained Output), cryptanalysis of block ciphers.

1 What Are Bitcoins and Bitcoin Mining

Bitcoin is a collaborative virtual currency and a decentralized peer-to-peer pay-
ment system without trusted central authorities. It has been invented in 2008
[27] and launched in 2009. Ever since Bitcoin was launched in 2009, it has been
presented by its technical architects as an experimental rather than mature elec-
tronic currency ecosystem. Initially bitcoin was a sort of social experiment, how-
ever bitcoins have been traded for real money for several years now. One year
ago, in April 2013 the leading newspaper The Economist explains that bitcoin
is here to stay, that it is a future of payments and calls it digital gold [14]. This
has coincided with the total value of bitcoins in circulation exceeding 1 billion
dollars. Bitcoin became a mainstream financial instrument.
In this paper we study bitcoin with particular attention paid to the process
of bitcoin mining, which is the specialist term given to the process by which
new bitcoins are created. Then later in Section 3 we are going to study in a lot
of detail one particular technical question in symmetric cryptography to see if
there exists an improved method which allows one to mine bitcoins faster.
Bitcoins are a type of digital currency which can be stored on a computer.
Bitcoins use the concept of so called “Proofs of Work” which are solutions to

Z. Kotulski et al. (Eds.): CSS 2014, CCIS 448, pp. 131–144, 2014.

c Springer-Verlag Berlin Heidelberg 2014
132 N.T. Courtois, M. Grajek, and R. Naik

certain very hard cryptographic puzzles based on hash functions. However these
solutions are NOT bitcoins. The puzzles are rather part of the bitcoin trust
infrastructure. In fact the puzzles are connected together to form a chain and
as the length of this chain grows, so does the security level. Bitcoins are simply
awarded to people who produce these “Proofs of Work”. Ownership of bitcoins
is achieved through digital signatures: the owner of a certain private key is the
owner of a certain quantity of bitcoins. This private key is the unique way to
transfer the bitcoin to another computer or person.
The operation of so called bitcoin mining or creating bitcoins out of the thin
air is not only possible. It is essential, it is encouraged, and it is a crucial and
necessary part of the Bitcoin ecosystem.
Cryptographic computations executed in a peer-to-peer network with tens of
thousands of independent hashing nodes are the heart of the security assurance
provided by this virtual currency system. It would be very difficult and extremely
costly for one entity to corrupt all these independent miners. The sum of all
this collective computational work provides some sort of solid cryptographic
proof and prevents attacks on this system. This is also how the network polices
itself: miners are expected to approve only correctly formed transactions. Bitcoin
implements a specific sort of distributed and decentralized electronic notary
system without a central authority. Well almost. Certain decisions about how
the system works, what exactly the bitcoin software does and how [3], are still
pretty centralized. In particular mining activity concentrates in extremely few
very large mining pools one of which controls more than 40 % of hash power,
see Table 2 in [9].
In a nutshell, bitcoin miners make money when they find a 32-bit value which,
when hashed together with the data from other transactions with a standard
hash function gives a hash with a certain number of 64 or more leading zeros.
This is an extremely rare event. It is in general believed that there is no way to
produce these data otherwise than by very long and costly computations. This
question of how to improve this process is the central question in this paper.

2 Detailed Study of Mining Internals

The goal of the miner is to solve a certain cryptographic puzzle which we will
later call a CISO Hash Problem. The solution will be called a CISO block.
Great majority of miners ignore what exactly they are doing, either running
open source software, or having purchased specialized hardware to do mining
very efficiently. However miners must know that the operation is very timely
and that they need to be permanently connected to the network. The solutions
to these puzzles are linked to each other and form a unique chain of blocks.
This is usually called the block chain. The whole block chain is public and the
whole of it can for example be consulted at http://blockexplorer.com/. All
new blocks which are found need to be broadcast to all network participants as
soon as possible. The miners need to be very reactive and they do it because
it is in their interest. They need to listen to broadcasts in order to receive the
 Optimizing SHA256 in Bitcoin Mining 133

data about recent transactions which they are expected to approve. Then they
need to broadcast any solution (CISO block) which they have found as soon as
they found it, because their solution is likely to be part of the "main chain of
blocks" only if it is widely known. Once the solution is known it "discourages"
other miners from searching for the same block. Instead they can concentrate on
searching for the next block which will confirm the present block and will make
the miner be able to claim his a reward for producing this CISO block.
Our goal is to clarify how this system works. In this paper we mostly consider
this to be a static problem which needs to be solved. We refer to [8] for a
discussion on how the difficulty of this problem changes with time.
The problem of bitcoin mining is very closely related to well known problems
in cryptography. One crucial question is as follows: how does the bitcoin min-
ing differ from traditional questions in cryptanalysis of block ciphers and hash
functions and is there a more efficient way to mine bitcoins.
First we are going to briefly describe the problem as a static computation
problem about a certain block cipher. Then we are going to look at how the
problem evolves in time and how solutions to the CISO problem are converted
to shares in the bitcoin currency. Finally we are going to study what the possible
solutions and optimizations are.

2.1 Constrained Input Small Output (CISO) Hashing Problem

New bitcoins can be created when the miner hashes some data from the bitcoin
network together with a 32-bit random nonce, and obtains a number on 256
bits which starts with a certain number of zeros. We call this problem CISO:
Constrained Input Small Output.
This can be seen as a special case of a problem which is sometimes called
CICO which translates into Constrained Inputs Constrained Outputs problem.
This terminology has been introduced in the study of the most recently stan-
dardized U.S. government standard hash function SHA-3 a.k.a. Keccak. SHA-3
is the latest hash function in the SHA family and a successor to SHA-256 used
in bitcoin [1]. It is possible to claim that this means that SHA-256 of Bitcoin
is considered by the United States NIST and a broader cryptographic engineer-
ing community as NOT sufficiently secure for long term security. This sort of
CISO/CICO problems are not new, they are very frequently studied in crypt-
analysis of hash functions since ever, and endless variants of these problems exist
for specific hash functions, some examples can be found in [1,11,26].
The exact details of the specific Constrained Input Small Output (CISO)
problem which we have in bitcoin are described below. It can be obtained by the
inspection of the bitcoin source code, see [3].

2.2 (CISO) Hashing Problem Internals

On Fig. 1 we show the cryptographic computation which is executed many many
times by bitcoin miners. This picture emphasizes the internal structure inside
SHA-256 hash function.
134 N.T. Courtois, M. Grajek, and R. Naik

SHA-256 is a hash function built from a block cipher following the so called
Davies-Meyer construction. The principle of the Davies-Meyer construction is
that the input value is at the end added to the output and that it transforms an
encryption algorithm into a “hashing” algorithm, a building piece of a standard
hash function. The underlying block cipher has 64 rounds and thus a 2048-bit
expanded internal key (64x32 bits). This key is obtained from the message block
to be compressed, which has 512 bits at the input and is expanded four times
to form this 2048-bit internal key for our block cipher. In one sense on Fig.
1 we convert the problem of bitcoin mining or of solving CISO hash puzzles,
to a specific problem with three distinct applications of the block cipher which
underlies SHA-256 connected together to form certain circuit.

Fig. 1. The Block Hashing Algorithm of bitcoin revisited and seen as a Constrained
Input Small Output (CISO) problem. We see two applications of SHA-256 together with
internal details of the Davies-Meyer construction. We can view it as a triple application
of a specific block cipher. An interesting question is whether there is a more efficient
cryptographic shortcut or inversion attack or some non-trivial optimizations which
allow to save a constant factor. Such optimizations, if they exist, could be worth some
serious money as they would allow to produce bitcoins cheaper
 Optimizing SHA256 in Bitcoin Mining 135

2.3 The Main Objective of CISO Hashing

The goal of CISO hashing is to produce solutions which are correctly formed
in the sense that they satisfy all the required conditions and constraints, which
we are going to explain in details in Section 2.4. The miner is trying to find a
solution to the CISO problem such that:

H2 < target . (1)

Here target is a large integer which is a global variable for the whole bitcoin
system worldwide, and on which all the participants worldwide are expected to
agree. The value of target is adjusted approximately every 2 weeks, see [8] for
more details. The job of bitcoin miners is to find these solutions and publish
them. They are rewarded with some bitcoins for their work. In 2014 the reward
is 25 BTC (25 bitcoins) per valid solution. The reward also changes very slowly
with time, see [8].
It is generally believed that there is no other method to achieve success than
trial and error; hashing at random as depicted on Fig. 1 until a result with a
sufficient number of leading zeros is found. However this is unlikely to be true,
there is always a better way, at least slightly, see Section 4.

2.4 The Inputs and Input Constraints In CISO Hashing

A number of data fields are present as inputs to the CISO problem, cf. Fig. 1.
Some fields are fixed or change very slowly and these are indicated in green on
Fig. 1. Other data need to be adjusted by the miner in order to obtain a solution
however they still change very slowly. Such data are indicated in yellow on the
main picture. Data which change the most frequently are indicated in red: these
are “hot” data which need to be re-computed each time the nonce changes.
Bitcoin is a live distributed system and the exact conditions required for the
data to be consider valid are essentially fixed and known, however they are likely
to evolve with time in subtle ways. Rules have already been and are likely to
be altered during the operation of the system. They depend on the consensus of
participants in the system. We have:

1. Version number on 32 bits. Has been equal to 1 then it became 2.

2. The previous block on 256 bits. Or more precisely a hash on 256 bits of
all the data of the last CISO block which was broadcast in the peer-to-peer
network: the miner works to extend the chain of hashes by one more block.
The process of generating new blocks is a sort of lottery and sometimes there
will be ambiguity on which block to use (less than 1 % of the time cf. [9]).
Miners in principle follow the longest chain or choose at random. However
miners could also disagree for profit or in order to attack the network and
there are endless subversive scenarios, see [8,9].
3. The Merkle root on 256 bits. This is a sort of an aggregated hash of many
recent events in the bitcoin network which certifies that the system recognizes
all of them simultaneously as being valid. It also contains the public key of
136 N.T. Courtois, M. Grajek, and R. Naik

the future owner of this freshly created portion of bitcoin currency which
this block is intended to embody. The current CISO block which the miner
is trying to create by solving the current CISO problem and all subsequent
CISO blocks will provide an accumulation of evidence about all these events.
This security guarantee increases with time. In principle miners can decide
which transactions are going to be recognized by the system. However the
transaction fees which will be obtained by the miners for the transactions
included in this block are an incentive to include every single transaction.
4. Timestamp on 32 bits. This is the current time in seconds.
5. Target on 32 bits. More precisely the global variable target is on 256 bits
and what is stored here is a compressed version of target which is frequently
called difficulty. We have difficulty = 6, 119, 726, 089 ≈ 232.5 as of 14
April 2014. We have difficulty · 232 = 1/probability = 2256 /target.
6. Nonce on 32 bits. This nonce is freely chosen by the miner. Interestingly
the nonce has only 32 bits while the current value of target makes that the
probability of obtaining a suitable H2 by accident is as low as 2−64.5 .
This means that the miner needs to be able to generate different versions of
the puzzle with a different Merkle root (or with other differences)
7. Padding+ Len has 384 bits for H1 and 256 bits for H2. Two constants
due to the specification of SHA-256 hash function which is used here twice
with data of different sizes: the input hashed has 640 and 256 bits respectively
in each application of SHA-256. These two values never change.

With respect to the input data requirements and constraints above and the
output constraint H2 < target we have:

Definition 1 (CISO Problem). We call the CISO Problem the problem of

finding a valid Merkle root, 32-bit nonce and other data as illustrated on Fig. 1
which is correctly formed and such that H2 < target.

3 How to Speed Up the Bitcoin Mining Process

In this part we look at a pure specialist question which pertains to symmetric
cryptography, of whether there is a cryptographic “shortcut” attack: simply a
method of mining bitcoins faster than brute force, or faster than the trivial
method in which the SHA-256 hash function is a black box. The question is what
is the fastest possible method for bitcoin mining, given the specific structure of
Fig. 1. We were the first to develop such techniques openly and publish them.
Related Search: One could also try to solve this problem by formal coding
and “a software algebraic attack”, see [10,23,28].
Our analysis follows the NIST specification of SHA-256 [15] and the inspection
of the Bitcoin source code [3]. SHA-256 is a hash function built from a block
cipher following the well-known Davies-Meyer construction in which the input is
at the end added to the output. This construction is one of the known methods
to transform a block cipher into a compression function. A compression function
is a building block of a hash function with a fixed input size. It is typically equal
 Optimizing SHA256 in Bitcoin Mining 137

Fig. 2. One compression function in SHA-256. It comprises a 256-bit block cipher with
64 rounds, a key expansion mechanism from 512 to 2048 bits, and a final set of eight
32-bit additions

to twice the output size. In our case we have a compression function from 512
to 256 bits, cf. Fig. 2.
The block size in this block cipher is 256 bits, the key size is 512 bits which
is expanded to 64 subkeys on 32 bits each for each of 64 rounds of the cipher.
The first 16 subkeys for the first 16 rounds are identical to the message and are
copied in the same order cf. [15] and later (Algorithm [1])
In addition in order to hash a full message, SHA-256 applies a Merkle-Damgard
padding and length extension which makes it a secure hash function for messages
of variable length. In the pre-processing stage, we must append one binary 1 and
many zeros to the message in such a way that the resulting length is equal to
448 modulo 512, cf. [15]. Then we append the length of the message in bits as a
64-bit big-endian integer. Full SHA-256 is applied twice. In the first application
of SHA-256 in bitcoin mining the message has a fixed length of 640 bits which
requires two applications of the compression function. In the second application
SHA-256 is applied to 256 bits.
It may therefore seem that a bitcoin miner needs to compute the compression
function 3.0 times for each nonce and for each Merkle hash. In the following
sections we are going to work on reducing this figure down to about 1.89.
We recall from Section 2.1 that new bitcoins can be created when the miner
succeeds to hash some data from the bitcoin network together with a 32-bit
random nonce and is able to obtain a number on 256 bits which starts with
a certain number of 64 or more zeros. We called it Constrained Input Small
Output problem or shortly the CISO problem, cf. Fig. 1. The process needs to
be iterated with different values of MerkleRoot and different 32-bit nonces until a
suitable “CISO configuration” is found in which the output satisfies H2 < target
as explained in Section 2.3.
138 N.T. Courtois, M. Grajek, and R. Naik

3.1 Improvement 1: Remove First Compression Function

We can reduce the cost factor from 3.0 to 2.0 almost instantly by making the
following observation. In the process of bitcoin mining the first compression
function does not depend on the random nonce on 32 bits. Therefore we can
compute it once every 232 nonces. On average we need:
1
2.0 + (2)
232
compression functions. The added factor is the amortized cost of the first hash
and can be neglected.
Discussion. In more advanced bitcoin mining algorithms the miner does not
have to compute the output for every nonce. He can do it only for some well
chosen nonces. They may be chosen in such a way as in order to obtain specific
values which make the computation easier. Moreover, some well chosen nonces
could be generated in some specific order in order to enable incremental compu-
tations. In an incremental computation some computations could made easier by
reusing all the (known) internal values in one or several previous computations.
There is a lot of highly non-trivial optimizations which can be developed. One
simple example of incremental computation will be given in Section 3.4, another
in Section 3.9.

3.2 Improvement 2: Save 3 Rounds at the End

We look at the computation of H2 on Fig. 1. The second computation of the
hash function and the third compression function. A close examination reveals
that in the last rounds of the underlying block cipher the two words on 32 bits
in which we want to have at least 64 zeros, after addition of a suitable constant,
are created at rounds 60 and 61 if we number from 0. We basically want to force
values created at rounds t = 60 and 61 to two fixed constants which come from
the SHA-256 IV constants, and which would produce zeros at the output. For
this most of the time we just need to compute the first 61 rounds out of 64
and we can early reject most cases. Only in 1/232 of cases we need to compute
62 rounds in the third compression function. Then only in some 1/264 of cases
where we have actually obtained at least 64 zeros, we would need compute the
full 64 rounds. Thus overall one only needs to compute the whole compression
function an equivalent of very roughly 1 + 61/64 ≈ 1.95 times on average. Most
of the time one only needs to compute H1 and 61 rounds of H2 to early reject
the 32-bit value obtained which must be equal to the IV constant.

3.3 Improvement 3: Gain Three Rounds at the Beginning

Now we look at the second computation of the hash function in the second com-
pression function, the computation of H1 on Fig. 1. Here we use the observation
that in SHA-256 the key for the first 16 rounds are exactly the 16 message blocks
in the same order, cf. [15] and Algorithm [1]. It is possible that in the second
 Optimizing SHA256 in Bitcoin Mining 139

compression function on Fig. 1 the nonce enters at round 3 (numbered from 0)

and therefore in most cases we just need to compute the last 61 out of 64 rounds
of the block cipher. The first three rounds are the same for every nonce and their
(amortized) cost is nearly zero.

3.4 Improvement 4: Incremental Calculations in Round 3

This improvement requires us to delve more deeply into the structure of the
block cipher inside SHA-256. We recall that the state of the cipher after round 2
is constant and does not yet depend on the value of the nonce. The 32-bit nonce
will be precisely copied to become the session key for the round 3 of encryption.
In the lower-left corner of Fig. 1 we show the circuit for one round of encryption
where at round 3 the nonce enters as W3 = nonce as shown on later in Table 1.
Here  denotes one addition on 32 bits.
Wt is the key derived from the message and Kt is a certain constant [15]. For
t = 3 we have W3 = nonce. Now it is obvious that the whole round 3 can be
computed essentially for free in the incremental way. We just need two 32-bit
increments instead of one whole round which is about 7 additions and 4 other 32-
bit operations. Each time we increment the nonce we simply need to increment
two values Thus we have saved one more round in each of our computations.

3.5 Improvement 5: Exploiting Zeros and Constants in the Key

The next improvement comes from the fact that the key in the first 16 rounds
of the block cipher is an exact copy of the message. Many parts of this key are
constants. Many are actually always equal to zero. This allows one to save a lot
of additions in the computation of SHA-256.

3.6 Improvement 6: Saving Two More Additions with Hard Coding

It is easy to see that 2 more additions can be saved. Looking at Table 1 we should
not count the three first constants on the left in yellow which are identical for
all the 232 different nonces. This is because this saving was already done in
Section 3.3. However we have two additional constants in the last line in green.
Then in these two last rounds, one in each computation, and because the Kt are
constants and do not depend on the message being hashed, we can pre-compute
the constants KWt = Kt  Wt on 32 bits which saves us 2 additions.

Remark 1. In Section 3.5 above and here in Section 3.6 we saved 18+2 additions.
However these savings are illusory, because we can save many more additions
by another method. On Fig. 3 we show that one only needs essentially 2 additions
in order to implement the whole round function of SHA256, this instead of 6+1
full adders in each compression function cf. Fig. 1.

We are going to use Carry Save Adders (CSA) in order to delay the propaga-
tion of carries and save a lot of circuit area. The main idea which is attributed to
140 N.T. Courtois, M. Grajek, and R. Naik

Table 1. Key in the first 16 rounds out of 64 in each computation and their provenance

Computation of H1 Computation of H2
Round t 32 bit Wt Description Round t 32 bit Wt Description
last 32 bits of 0 XXXXXXXX H10
0 XXXXXXXX hash Markle Root 1 XXXXXXXX H11
1 XXXXXXXX timestamp 2 XXXXXXXX H12
2 XXXXXXXX target 3 XXXXXXXX H13
nonce (00000000
4 XXXXXXXX H14
3 XXXXXXXX to FFFFFFFF) 5 XXXXXXXX H15
4 0x80000000 Padding starts 6 XXXXXXXX H16
5 0x00000000 | 7 XXXXXXXX H17
6 0x00000000 | Padding
7 0x00000000 | 8 0x80000000 starts
8 0x00000000 | 9 0x00000000 |
9 0x00000000 | 10 0x00000000 |
10 0x00000000 | 11 0x00000000 |
11 0x00000000 | 12 0x00000000 |
12 0x00000000 | Padding
13 0x00000000 Padding ends 13 0x00000000 ends
14 0x00000000 length H 14 0x00000000 length H
15 0x00000280 length L 15 0x00000100 length L

John von Neumann, is to propagate the carries only locally delaying a complete
propagation to the very end. This allows a dramatic reduction in the cost of
implementing multiple additions: three or more additions do NOT cost much
more than one single addition.
More precisely Carry Save Adders (CSA) allow to add n numbers for any n ≥ 3
and to form two numbers which need to be added to obtain the final result. This
is obtained by a successive transformation of 3 numbers into 2 numbers with a
Carry Save Adder (CSA) which has a very low cost and a final addition of 2
numbers. A Carry Save Adder takes 3 integers a, b, c on k bits written in binary
and outputs two numbers ps (partial sum) and sc (shift-carry) as follows:

psi =ai ⊕ bi ⊕ ci ,
(3)
sci+1 =ai bi ∨ ai ci ∨ bi ci .

In the case of addition modulo 2k there is a slight simplification as the most

significant digits can be discarded but the result remains essentially the same.
Overall the application of Carry Save Adders (CSA) allows us to implement
each round of SHA256 with only two full additions cf. Fig. 3.
Now we are also going to look beyond the 64 rounds of SHA256 seen as a
block cipher. What remains is the key expansion which expands the message to
be hashed into the 64x32 bit keys Wt .
 Optimizing SHA256 in Bitcoin Mining 141

Fig. 3. How to compute one round of SHA-256 with just two full adders

3.7 Short Description of Message Expansion in SHA-256

Before we can propose additional optimizations, we need to explain how the
message expansion works in the NIST specification of SHA-256 [15]. We refer to
[15] for definitions of σ0 and σ1 .

Algorithm 1. The message scheduler expanding a 512 message block into a

2048-bit key for the SHA-256 block cipher.
for 0 ≤ t ≤ 15 do
Wt = Mt
end for
for 16 ≤ t ≤ 63 do
Wt = σ1 (Wt−2 )  Wt−7  σ0 (Wt−15 )  Wt−16
end for

3.8 Improvement 7: Saving Two More Additions

We consider the computation of H1. It is possible to see that the first two non-
trivial keys W16 and W17 are also constants and do not yet depend on the
142 N.T. Courtois, M. Grajek, and R. Naik

nonce. Therefore they can be pre-computed and as in Improvement 6 we can

replace the traditional hard-coded constants Kt by new hard-coded constants
KWt = Wt  Kt for the whole range of 232 nonces. Thus we save two more
additions.

3.9 Improvement 8: One More Additions Saved by Incremental

Computation
Now still in the case of H1 we have:

W19 = σ1 (W17 )  W11  σ0 (W4 )  W3 . (4)

Here W3 is the nonce which is incremented by 1. Other elements do not
depend on the nonce and change at a much slower rate. We can thus save one
more addition and simply increment W19 directly for each consecutive nonce.
Summary: There remains 48-3 values Wt to be computed. If we use again
Carry Save Adders (CSA) we see that in each of the 45 cases we need just one
full adder instead of 3 full adders.

4 Our Overall Result on the Speed of Hashing

We have obtained the following result:

Theorem 1. [Hash Speed] The amortized average cost of trying one output H2
to see if it has 64 or more leading zeros is at most about 1.89 computations
of the compression function of SHA-256 instead of 3.0, which represents an
improvement by 37%. 


Justification: We have saved about 7 rounds and many additions. However known
ASIC implementations also save many additions and actually the designs which
achieve the lowest possible area are not necessarily the fastest. Therefore we
are just going to estimate the RELATIVE savings w.r.t. best standard ASIC
implementations of full SHA256 such as in [5,12,13,16,17,18,19,20,21,25,29,30].
Thus overall cost minus savings are equivalent to a total of
64 + 64 7 7
− =2− ≈ 1.89 (5)
64 64 64
compression functions.
This 1.89 compression functions is equivalent to saving of 37% compared to
the initial cost of 3.0 compression functions as per Fig. 1. It also shows how
much can be gained in bitcoin mining compared to using an optimized SHA256
ASIC implementation three times.

Remark 2. Our problem is essentially the same as a brute force attack on a block
cipher. The same computation is done a very large number of times, yet cheaper,
maybe just a small factor cheaper. It is not correct to believe that block ciphers
are well understood in cryptography. On the contrary, it appears that for more
 Optimizing SHA256 in Bitcoin Mining 143

or less any block cipher there may exist an attack which will be just slightly
faster than brute force, see [24]. An efficient low-data software algebraic attack
could also be a solution to this problem, cf. [10,23,28].

5 Conclusion

In this paper we explain how bitcoin electronic currency works and show that
the profitability of bitcoin mining depends on a certain cryptographic constant
which we showed to be at most 1.89. Normally very few people care about this
sort of fine cryptographic engineering details. However here it is different. This
observation allows bitcoin miners to save many millions of dollars each year.

References
1. Aumasson, J.-P., Khovratovich, D.: First Analysis of Keccak (2009),
http://131002.net/data/papers/AK09.pdf
2. Barber, S., Boyen, X., Shi, E., Uzun, E.: Bitter to Better — How to Make Bitcoin a
Better Currency. In: Keromytis, A.D. (ed.) FC 2012. LNCS, vol. 7397, pp. 399–414.
Springer, Heidelberg (2012)
3. Nakamoto, S., et al.: Bitcoin QT: http://bitcoin.org/en/download
4. Boyar, J., Matthews, P., Peralta, R.: Logic Minimization Techniques with Appli-
cations to Cryptology. Journal of Cryptology 26, 280–312 (2013)
5. Chaves, R., Kuzmanov, G., Sousa, L., Vassiliadis, S.: Improving SHA-2 hardware
implementations. In: Goubin, L., Matsui, M. (eds.) CHES 2006. LNCS, vol. 4249,
pp. 298–310. Springer, Heidelberg (2006)
6. Courtois, N.T., Hulme, D., Mourouzis, T.: Solving Circuit Optimisation Problems
in Cryptography and Cryptanalysis. In: Proceedings of SHARCS 2012 Workshop,
UK, pp. 179–191 (2011)
7. Courtois, N.T., Hulme, D., Mourouzis, T.: Multiplicative Complexity and Solv-
ing Generalized Brent Equations With SAT Solvers. In: COMPUTATION TOOLS
2012, The Third International Conference on Computational Logics, Algebras, Pro-
gramming, Tools, and Benchmarking. ARIA, Nice (2012)
8. Courtois, N.T., Grajek, M., Naik, R.: The Unreasonable Fundamental Incertitudes
Behind Bitcoin Mining (2013), http://arxiv.org/abs/1310.7935
9. Courtois, N.T., Bahack, L.: On Subversive Miner Strategies and Block Withholding
Attack in Bitcoin Digital Currency (2014), http://arxiv.org/abs/1402.1718
10. Courtois, N.T., Bard, G.V.: Algebraic Cryptanalysis of the Data Encryption Stan-
dard. In: Galbraith, S.D. (ed.) Cryptography and Coding 2007. LNCS, vol. 4887,
pp. 152–169. Springer, Heidelberg (2007)
11. Courtois, N.T., Mourouzis, T.: Black-Box Collision Attacks on the Compression
Function of the GOST Hash Function. In: Proceedings of 6th International Con-
ference on Security and Cryptography SECRYPT, Spain (2011)
12. Dadda, L., Macchetti, M., Jeff Owen, J.: An ASIC design for a high speed imple-
mentation of the hash function SHA-256 384, 512. In: ACM Great Lakes Sympo-
sium on VLSI, pp. 421–425. ACM (2004)
13. Dadda, L., Macchetti, M., Owen, J.: The Design of a High Speed ASIC Unit for
the Hash Function SHA-256 (384, 512). In: DATE 2004, pp. 70–75. IEEE (2004)
144 N.T. Courtois, M. Grajek, and R. Naik

14. Virtual currencies: Mining digital gold, From the print edition: Finance and eco-
nomics, The Economist (2013)
15. National Institute of Standards and Technology (NIST). FIPS PUB 180-2, SHA256
Standard (2002),
http://csrc.nist.gov/publications/fips/
fips180-2/fips180-2withchangenotice.pdf
16. Feldhofer, M., Rechberger, C.: A Case Against Currently Used Hash Functions in
RFID Protocols. In: Meersman, R., Tari, Z., Herrero, P. (eds.) OTM 2006 Work-
shops. LNCS, vol. 4277, pp. 372–381. Springer, Heidelberg (2006)
17. Knezevic, M.: Efficient Hardware Implementations of Cryptographic Primitives.
PhD thesis, Katholieke Universiteit Leuven (2011)
18. Lee, Y.K., Chan, H., Verbauwhede, I.: Iteration bound analysis and throughput
optimum architecture of SHA-256 (384,512) for hardware implementations. In:
Kim, S., Yung, M., Lee, H.-W. (eds.) WISA 2007. LNCS, vol. 4867, pp. 102–114.
Springer, Heidelberg (2008)
19. Macchetti, M., Dadda, L.: Quasi-Pipelined Hash Circuits. In: IEEE Symposium on
Computer Arithmetic, pp. 222–229 (2005)
20. Michail, H.E., Athanasiou, G., Kritikakou, A., Goutis, C.E., Gregoriades, A., Pa-
padopoulou, V.G.: Ultra High Speed SHA-256 Hashing Cryptographic Module for
IPSec Hardware/Software Codesign. In: SECRYPT, pp. 309–313 (2010)
21. Michail, H.E., Athanasiou, G., Gregoriades, A., Panagiotou, C.L., Goutis, C.E.:
High Throughput Hardware/Software Co-design Approach SHA-256 Hashing
Cryptographic Module. Global Journal of Computer Science and Technology 10,
15 (2010)
22. Guo, J., Matusiewicz, K.: Preimages for Step-Reduced SHA-2 (2008),
http://eprint.iacr.org/2009/477.pdf
23. Heusser, J.: SAT solving - An alternative to brute force bitcoin mining (2013),
http://jheusser.github.io/2013/02/03/satcoin.html
24. Huang, J., Lai, X.: What is the Effective Key Length for a Block Cipher: an Attack
on Every Block Cipher (2012), http://eprint.iacr.org/2012/677
25. Kim, M., Ryou, J., Jun, S.: Efficient Hardware Architecture of SHA-256 Algorithm
for Trusted Mobile Computing. In: Yung, M., Liu, P., Lin, D. (eds.) Inscrypt 2008.
LNCS, vol. 5487, pp. 240–252. Springer, Heidelberg (2009)
26. Matusiewicz, K., Pieprzyk, J., Pramstaller, N.: Rechberger, Ch., Rijmen, V.:
Analysis of simplified variants of SHA-256:
http://www2.mat.dtu.dk/people/K.Matusiewicz/papers/SimplifiedSHA256.pdf
27. Nakamoto, S.: Bitcoin: A Peer-to-Peer Electronic Cash System:
http://bitcoin.org/bitcoin.pdf
28. Raddum, H., Semaev, I.: New Technique for Solving Sparse Equation Systems. In:
ECRYPT STVL (2006), http://eprint.iacr.org/2006/475/
29. Sklavos, N., Koufopavlou, O.G.: On the hardware implementations of the SHA-2
(256, 384, 512) hash functions. ISCAS 5, 153–156 (2003)
30. Tillich, S., Feldhofer, M., Kirschbaum, M., Plos, T., Schmidt, J.-M., Alexander
Szekely, A.: Uniform Evaluation of Hardware Implementations of the Round-Two
SHA-3 Candidates. In: Second SHA-3 Conference (2010),
http://csrc.nist.gov/groups/ST/hash/sha-3/Round2/
Aug2010/documents/papers/TILLICH_sha3hw.pdf