EIP-8297: Partitioned Binary Tree

Switch Ethereum state tree to a partitioned binary tree

Abstract

Introduce a new binary state tree to replace the hexary Patricia tries. Account and storage tries are merged into a single tree with 32-byte keys that also holds contract code. Account data is broken into independent leaves grouped by 256 to provide locality.

The tree is partitioned into zones. The high 4 bits of every key are a zone identifier that labels the category of state the key holds: account headers, contract code, or storage. Storage takes the entire upper half of the zone space (any key with the high bit set), so it is rooted at depth 1 and occupies about half the tree. Account headers and code take fixed low zones, and the remaining low zones are reserved for future categories.

Note: the hash function used in this draft is not final. The reference implementation uses BLAKE3 to reduce friction for clients experimenting with this EIP, but the choice remains open.

Motivation

Ethereum’s long-term goal is to let blocks be proved with validity proofs so chain verification is as simple and fast as possible. Part of this work consists of proving the state read during EVM execution.

The Merkle Patricia Trie (MPT) is unfriendly to validity proofs: it uses RLP for node encoding, Keccak for hashing, is a “tree of trees”, and does not allow for the efficient proving of segments of bytecode. It also produces large Merkle proofs. As an example, the account trie today reaches a maximum depth of about 12, so the expected size of a single branch is 15 * 32 * 12 = 5760 bytes: 15 sibling hashes of 32 bytes at each of the 12 levels. From a worst-case block perspective, spending all 60M gas to touch a single byte of many distinct account codes, none of which is chunked, needs 60M/2400 * (12*480 + 64k) ≈ 1.8GB. Here 2400 is the cheapest gas to touch a fresh account (the EIP-2930 access-list address cost. A cold account access under EIP-2929 is 2600), 12*480 is the branch to that account, and 64k is the EIP-7954 64 KiB code size limit that must be revealed in full to prove any single byte of unchunked code.

A binary tree shrinks regular Merkle proofs, because proof size scales with siblings * log_arity(N) and arity 2 minimizes it. Switching from Keccak to a more proving-friendly hash improves circuit performance.

Partitioning the tree into zones adds two properties on top of a flat unified tree:

Structural boundaries. A node at a known depth is always the root of a known category: all storage at depth 1, the account zone and code zone at depth 4, one account’s storage bucket at depth 61. Protocols can reference these subtree roots as commitments without a side structure. This is what later proposals for state expiry and partial statelessness build on.

Code deduplication. Code beyond the first chunks is content-addressed by code hash rather than by account, so thousands of contracts deployed from the same factory share their code leaves instead of each storing a copy.

Specification

The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “NOT RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in RFC 2119 and RFC 8174.

Notable changes from the hexary structure

• The account and storage tries are merged into a single trie.
• RLP is no longer used.
• The account’s code is chunked and included in the tree.
• Account data (balance, nonce, first storage slots, first code chunks) is co-located to reduce branch openings.

Notable changes from EIP-7864

EIP-7864 specifies a flat unified binary tree where every key is key_hash(address || tree_index)[:31] || sub_index. This proposal keeps that tree, its node types, and its merkelization, and changes how keys are assigned:

• A 4-bit zone prefix is added to every stem. In EIP-7864 all of an account’s data (header, storage, code) is derived from its address and scattered through the tree by tree_index. Here each category lives in its own zone.
• Overflow code (chunks at index 128 and above) moves to zone 0x1 and is keyed by code_hash instead of by account, so contracts with identical bytecode share leaves. EIP-7864 keys all code per account, storing a copy for every contract.
• Overflow storage (slots 64 and above) moves to the storage zone, where a stem is the storage high bit, a 60-bit address prefix, and a 187-bit suffix. EIP-7864 uses key_hash(address || tree_index)[:31] with no per-account bucket boundary.
• MAIN_STORAGE_OFFSET is removed. EIP-7864 offsets main storage by 256**31 to keep it clear of the header and code region. Here the storage zone bit separates storage structurally, so no numeric offset is needed.
• Non-storage stems are a 4-bit zone followed by a 244-bit hash. The hash is truncated to 244 bits rather than EIP-7864’s 248, since the zone takes the high 4 bits of the stem.

The account header layout (BASIC_DATA packing, CODE_HASH, the first 64 storage slots, the first 128 code chunks), the code chunkification, and the four node types are unchanged from EIP-7864.

Tree structure

The tree stores key-value entries where both key and value are 32 bytes. The first 31 bytes of the key are the entry stem, and the last byte is the sub-index within that stem. Two keys with the same stem live in the same group of 256 leaves.

There are four node types:

• InternalNode has left_hash and right_hash.
• StemNode has stem, left_hash, and right_hash.
• LeafNode has a value that is a 32-byte blob or empty.
• EmptyNode represents an empty node or subtree.

Keys are traversed in big-endian bit order. A key is expanded to a bit list by _bytes_to_bits (defined below), most significant bit (MSB) first: bit 0 of the path is the MSB of the first byte, and the last bit is the least significant bit (LSB) of the last byte. Depth d in the tree selects bit d of this list, so the zone identifier in the high 4 bits of the stem governs the top four levels.

The path to a StemNode is the stem’s 248 bits in this MSB-to-LSB order. From that node a subtree of 256 values is indexed by the last byte. The path does not use the full 248 bits: it contains the minimal number of InternalNodes needed to separate stems that share a prefix. Walking the path to insert a new stem ends on an EmptyNode (replace it with the stem) or a StemNode (create as many InternalNodes as the two stems share prefix bits). There are no extension nodes.

This minimal-internal-node rule keeps the zone prefix cheap. Only a few low zones are populated, so the zone bits between the root and a populated zone collapse to the minimal set of internal nodes rather than a full four-level chain.

def _bytes_to_bits(data: bytes) -> list[int]:
    return [(byte >> (7 - i)) & 1 for byte in data for i in range(8)]

def _bits_to_bytes(bits: list[int]) -> bytes:
    return bytes(
        sum(bits[i + j] << (7 - j) for j in range(8))
        for i in range(0, len(bits), 8)
    )

class StemNode:
    def __init__(self, stem: bytes):
        assert len(stem) == 31, "stem must be 31 bytes"
        self.stem = stem
        self.values: list[Optional[bytes]] = [None] * 256

    def set_value(self, index: int, value: bytes):
        self.values[index] = value

class InternalNode:
    def __init__(self):
        self.left = None
        self.right = None

class BinaryTree:
    def __init__(self):
        self.root = None

    def insert(self, key: bytes, value: bytes):
        assert len(key) == 32, "key must be 32 bytes"
        assert len(value) == 32, "value must be 32 bytes"
        stem = key[:31]
        subindex = key[31]

        if self.root is None:
            self.root = StemNode(stem)
            self.root.set_value(subindex, value)
            return

        self.root = self._insert(self.root, stem, subindex, value, 0)

    def _insert(self, node, stem, subindex, value, depth):
        assert depth < 248, "depth must be less than 248"

        if node is None:
            node = StemNode(stem)
            node.set_value(subindex, value)
            return node

        stem_bits = _bytes_to_bits(stem)
        if isinstance(node, StemNode):
            if node.stem == stem:
                node.set_value(subindex, value)
                return node
            existing_stem_bits = _bytes_to_bits(node.stem)
            return self._split_leaf(
                node, stem_bits, existing_stem_bits, subindex, value, depth
            )

        bit = stem_bits[depth]
        if bit == 0:
            node.left = self._insert(node.left, stem, subindex, value, depth + 1)
        else:
            node.right = self._insert(node.right, stem, subindex, value, depth + 1)
        return node

    def _split_leaf(self, leaf, stem_bits, existing_stem_bits, subindex, value, depth):
        if stem_bits[depth] == existing_stem_bits[depth]:
            new_internal = InternalNode()
            bit = stem_bits[depth]
            if bit == 0:
                new_internal.left = self._split_leaf(
                    leaf, stem_bits, existing_stem_bits, subindex, value, depth + 1
                )
            else:
                new_internal.right = self._split_leaf(
                    leaf, stem_bits, existing_stem_bits, subindex, value, depth + 1
                )
            return new_internal
        else:
            new_internal = InternalNode()
            bit = stem_bits[depth]
            stem = _bits_to_bytes(stem_bits)
            if bit == 0:
                new_internal.left = StemNode(stem)
                new_internal.left.set_value(subindex, value)
                new_internal.right = leaf
            else:
                new_internal.right = StemNode(stem)
                new_internal.right.set_value(subindex, value)
                new_internal.left = leaf
            return new_internal

Node merkelization

Define hash(value) as:

• hash([0x00] * 64) = [0x00] * 32
• hash(value) = H(value), where H is the selected hash function.

The value length is 32 or 64. Merkelize each node type as:

• internal_node_hash = hash(left_hash || right_hash)
• stem_node_hash = hash(stem || 0x00 || hash(left_hash || right_hash))
• leaf_node_hash = hash(value)
• empty_node_hash = [0x00] * 32

def _hash(self, data):
    if data in (None, b"\x00" * 64):
        return b"\x00" * 32
    assert len(data) == 64 or len(data) == 32, "data must be 32 or 64 bytes"
    return blake3(data).digest()

def merkelize(self):
    def _merkelize(node):
        if node is None:
            return b"\x00" * 32
        if isinstance(node, InternalNode):
            left_hash = _merkelize(node.left)
            right_hash = _merkelize(node.right)
            return self._hash(left_hash + right_hash)

        level = [self._hash(x) for x in node.values]
        while len(level) > 1:
            new_level = []
            for i in range(0, len(level), 2):
                new_level.append(self._hash(level[i] + level[i + 1]))
            level = new_level
        return self._hash(node.stem + b"\0" + level[0])

    return _merkelize(self.root)

Zones

The high 4 bits of every stem are the zone identifier Z. The most significant bit separates storage from everything else.

New categories MUST be allocated from 0x2–0x7. The storage range MUST NOT be subdivided for other categories, since that would remove address entropy from storage keys.

The zones place deterministic boundaries at fixed depths:

Tree embedding

All state is embedded into the single key/value space. Data accessed together is co-located in one stem to reduce branch openings. The account header stem holds an account’s basic data, code hash, first 64 storage slots, and first 128 code chunks. Storage slots and code chunks beyond those live in their own zones, grouped 256 to a stem.

It is a required invariant that STEM_SUBTREE_WIDTH > CODE_OFFSET > HEADER_STORAGE_OFFSET and that HEADER_STORAGE_OFFSET is greater than the leaf keys.

Addresses are passed as Address32. Convert a legacy address by prepending 12 zero bytes:

def address20_to_address32(address: Address) -> Address32:
    return b'\x00' * 12 + address

Stems in a non-storage zone are a 4-bit zone prefix followed by 244 bits of hash:

def key_hash(inp: bytes) -> bytes32:
    return blake3(inp).digest() 

def zone_stem(zone: int, digest: bytes) -> bytes:
    # 4-bit zone || 244-bit hash = 248-bit (31-byte) stem
    zone_bits = [(zone >> (ZONE_BITS - 1 - i)) & 1 for i in range(ZONE_BITS)]
    return _bits_to_bytes(zone_bits + _bytes_to_bits(digest)[:248 - ZONE_BITS])

Header values

The account header stem is in ACCOUNT_ZONE and is keyed by the address alone, so each account has exactly one header stem.

def get_tree_key_for_header(address: Address32, sub_index: int) -> bytes:
    stem = zone_stem(ACCOUNT_ZONE, key_hash(address))
    return stem + bytes([sub_index])

def get_tree_key_for_basic_data(address: Address32):
    return get_tree_key_for_header(address, BASIC_DATA_LEAF_KEY)

def get_tree_key_for_code_hash(address: Address32):
    return get_tree_key_for_header(address, CODE_HASH_LEAF_KEY)

version, balance, nonce, and code_size are packed big-endian in the value at BASIC_DATA_LEAF_KEY:

Bytes 1 through 3 are reserved. The 4-byte code_size holds values up to 2^32 - 1 bytes, far beyond any foreseeable contract size limit. Packing these fields into one leaf needs one branch opening instead of three or four, which lowers gas and simplifies witness generation. Setting any header field also sets version to zero. code_hash and code_size are set on contract or EOA creation.

Code

Code chunks 0 through 127 live in the account header stem at sub-indices CODE_OFFSET..255. Chunks at index 128 and above live in CODE_ZONE, content-addressed by code_hash so contracts with identical bytecode share leaves.

def get_tree_key_for_code_chunk(address: Address32, code_hash: bytes32, chunk_id: int):
    if chunk_id < STEM_SUBTREE_WIDTH - CODE_OFFSET:        # chunk_id < 128
        return get_tree_key_for_header(address, CODE_OFFSET + chunk_id)
    overflow = chunk_id - (STEM_SUBTREE_WIDTH - CODE_OFFSET)
    tree_index = overflow // STEM_SUBTREE_WIDTH
    sub_index  = overflow %  STEM_SUBTREE_WIDTH
    stem = zone_stem(CODE_ZONE, key_hash(code_hash + tree_index.to_bytes(32, "big")))
    return stem + bytes([sub_index])

Chunk i stores a 32-byte value where bytes 1..31 are the i’th 31-byte slice of the code and byte 0 is the number of leading bytes that are PUSHDATA. For example, if code is ...PUSH4 99 98 | 97 96 PUSH1 128 MSTORE... where | begins a new chunk, the latter chunk begins 2 97 96 PUSH1 128 MSTORE, recording that its first 2 bytes are PUSHDATA.

PUSH_OFFSET = 95
PUSH1 = PUSH_OFFSET + 1
PUSH32 = PUSH_OFFSET + 32

def chunkify_code(code: bytes) -> Sequence[bytes32]:
    if len(code) % 31 != 0:
        code += b'\x00' * (31 - (len(code) % 31))
    bytes_to_exec_data = [0] * (len(code) + 32)
    pos = 0
    while pos < len(code):
        if PUSH1 <= code[pos] <= PUSH32:
            pushdata_bytes = code[pos] - PUSH_OFFSET
        else:
            pushdata_bytes = 0
        pos += 1
        for x in range(pushdata_bytes):
            bytes_to_exec_data[pos + x] = pushdata_bytes - x
        pos += pushdata_bytes
    return [
        bytes([min(bytes_to_exec_data[pos], 31)]) + code[pos: pos+31]
        for pos in range(0, len(code), 31)
    ]

Storage

Storage slots 0 through 63 live in the account header stem at sub-indices HEADER_STORAGE_OFFSET..127. Slots 64 and above live in the storage zone.

A storage stem is 248 bits: the high bit is 1 (the storage zone), the next 60 bits are an address prefix, and the last 187 bits are a suffix bound to both the address and the tree_index. The address prefix groups one account’s storage under a common bucket at depth 61. The suffix hashes the address with tree_index, so two accounts that share a prefix still produce independent stems.

STORAGE_ADDR_PREFIX_BITS = 60
STORAGE_SUFFIX_BITS      = 187

def storage_stem(address: Address32, tree_index: int) -> bytes:
    prefix = key_hash(address)
    suffix = key_hash(address + tree_index.to_bytes(32, "big"))
    bits = [1]                                              # storage zone high bit
    bits += _bytes_to_bits(prefix)[:STORAGE_ADDR_PREFIX_BITS]
    bits += _bytes_to_bits(suffix)[:STORAGE_SUFFIX_BITS]    # 1 + 60 + 187 = 248
    return _bits_to_bytes(bits)

def get_tree_key_for_storage_slot(address: Address32, storage_key: int):
    if storage_key < CODE_OFFSET - HEADER_STORAGE_OFFSET:   # storage_key < 64
        return get_tree_key_for_header(address, HEADER_STORAGE_OFFSET + storage_key)
    tree_index = storage_key // STEM_SUBTREE_WIDTH
    sub_index  = storage_key %  STEM_SUBTREE_WIDTH
    return storage_stem(address, tree_index) + bytes([sub_index])

Slots within one STEM_SUBTREE_WIDTH range share a stem, except for slots 0..63, which are in the header. Adjacent slots, common in mappings and arrays, group together.

Fork

Described in EIP-7612.

Access events

Partitioned Binary Tree (PBT) adopts EIP-4762’s access-event framework with two required modifications:

1. Content-addressed code. EIP-4762 keys code-chunk access events per account at (CODE_OFFSET + i) // STEM_SUBTREE_WIDTH. PBT moves chunks ≥ 128 to a leaf shared by all contracts with the same code_hash. Access events for overflow code MUST be keyed by the (zone, stem, sub-index) tree-key, not by (address, chunk), so that a shared chunk is charged once per block regardless of which contract triggers the access and so the witness contains one copy. Header chunks (0–127) remain per-account.

2. Branch-cost calibration. WITNESS_BRANCH_COST (1900) is calibrated in EIP-4762 to ≈13.2 gas/byte assuming a ~144-byte average branch for the EIP-7864 depth profile. PBT’s branches are deeper, so we need to adjust the values.

Rationale

This EIP defines a binary tree that starts empty. Only new state changes are stored in it. The MPT continues to exist but is frozen, setting up a later hard fork that migrates MPT data into the binary tree (EIP-7748).

Single tree with zones

A single key/value tree is simpler to work with than a tree of tries: database access, caching, syncing, and proof code all operate on one abstraction, and witness gas rules are clearer. State spreads uniformly, so one contract’s millions of slots are not concentrated in one place, which helps state sync and blunts unbalanced-tree-filling attacks.

Zones add structure without giving up that uniformity. Each zone is a self-contained subtree at a known depth, so a node can sync, prove, or expire one category without touching the rest. Because no leaf stores a storage_root, an account’s nonce in the account zone and one of its slots in the storage zone are independent writes. The root recomputes in one bottom-up pass and the two branches meet near the root. This admits parallelism across zones, across accounts within a zone, and across stems within an account.

Storage layout

Storage takes the upper half of the tree because it is the largest category and the most frequently proven, so it should branch earliest and shallowest. The 60-bit address prefix gives each account its own storage bucket at depth 61, which is the unit later expiry and partial-statefulness schemes can prune or sync. The 187-bit suffix is bound to the address and tree_index so a prefix collision cannot correlate two accounts’ storage layouts.

Content-addressed code

Keying overflow code by code_hash rather than by account lets all contracts with identical bytecode share leaves. Most deployed contracts repeat a small number of templates, so this removes a large amount of duplicate code from the state. The first 128 chunks stay in the account header, keyed per account, so they expire with the account and need no reference counting. Only chunks beyond ~4 KB are shared.

SNARK friendliness and post-quantum security

The design avoids complex branching: no extension nodes mid-branch, no RLP. This keeps circuit implementations simple and lowers proving overhead. The dominant factor is the merkelization hash, which should be efficient in and out of circuit. The choice is open, with candidates:

1. BLAKE3: good native performance, reasonable in-circuit, well-studied, currently used in the reference implementation.
2. Keccak: already in Ethereum, well-studied, less efficient to prove.
3. Poseidon2: strong in-circuit performance, security analysis ongoing through the Ethereum Foundation (EF) cryptography initiative, needs extra specification for field encoding.

Because the tree depends only on a hash function and not on elliptic curves, it stays secure under the post-quantum assumptions that NIST’s 2030 guidance and Verkle’s curve-based stack do not. Progress in proving systems suggests pre-state and post-state proofs can be generated fast enough, which was Verkle’s main advantage.

Arity-2

Binary tries minimize witness size. In an N-element tree with k children per node, the average branch is roughly 32 * (k-1) * log(N) / log(k) bytes, minimized at k = 2. For N = 2**24:

Tree depth

The tree design attempts to be as simple as possible considering both out-of-circuit and circuit implementations, while satisfying our throughput constraints on proving hashes.

The proposed design avoids a full 248-bit depth as it would happen in a Sparse Merkle Tree (SMT). This approach helps reduce the hashing load in proving systems, which is currently a throughput bottleneck on commodity hardware.

Moreover, we could push this further trying to introduce extension nodes in the middle of paths as done in an MPT, but this also adds complexity that might not be worth it considering the tree should be quite balanced.

State expiry

Per-account and per-bucket expiry is a natural operation on the zone topology. The storage bucket at depth 61 roots one account’s storage in the common case. Record its hash and prune below it. The account header stem expires the account’s core data, hot storage, and initial code in one step. Content-addressed code needs reference counting or deferral to a state sweep, since its leaves may be shared. Resurrection re-attaches a subtree consistent with the recorded commitment. The mechanism itself is left to a separate EIP.

Backwards Compatibility

The main breaking changes are:

1. Gas costs for code chunk access can affect application economics. This can be mitigated by raising the gas limit alongside this EIP.
2. The tree structure change makes in-EVM proofs of historical state stop working.

The change is invisible to the EVM. Contracts address storage by 256-bit slot numbers through SLOAD and SSTORE and never see tree keys. Key derivation runs inside the client, below the EVM, exactly as the MPT already hashes slot keys and addresses. No contract, Solidity, or Yul code changes.

Test Cases

The hash function is not fixed, so digests cannot be pinned. The deterministic parts of the derivation are given as vectors. H(...)[:n] is the high n bits.

Account header, BASIC_DATA of address A:

stem    = 0x0 || H(A)[:244]      (4-bit zone || 244-bit hash)
sub_idx = 0x00
widths  = 4 + 244 + 8 = 256

Storage slot storage_key = 5 of address A (in the header, since 5 < 64):

sub_idx = HEADER_STORAGE_OFFSET + 5 = 69 (0x45)
key     = 0x0 || H(A)[:244] || 0x45

Storage slot storage_key = 1000 (in the storage zone, since 1000 >= 64):

tree_index = 1000 // 256 = 3
sub_idx    = 1000 %  256 = 232 (0xE8)
stem       = 1 || H(A)[:60] || H(A || 3)[:187]      (high bit = 1)
widths     = 1 + 60 + 187 + 8 = 256

Code chunk chunk_id = 5 (in the header, since 5 < 128):

sub_idx = CODE_OFFSET + 5 = 133 (0x85)
key     = 0x0 || H(A)[:244] || 0x85

Code chunk chunk_id = 300 of bytecode with hash C (overflow, since 300 >= 128):

overflow   = 300 - 128 = 172
tree_index = 172 // 256 = 0
sub_idx    = 172 %  256 = 172 (0xAC)
stem       = 0x1 || H(C || 0)[:244]
widths     = 4 + 244 + 8 = 256

Security Considerations

A collision means two distinct items derive the same stem.

Storage prefix collisions are not consensus-fatal. At 10 billion accounts, about 43 address pairs share a 60-bit prefix and so share a bucket. Their leaves interleave below depth 61 but stay distinct, because the suffix hashes the address with tree_index. An attacker who finds a prefix collision cannot overwrite the other account’s storage. The only effect is reduced parallelism at that bucket.

Content-addressed code. Two contracts with identical bytecode share code-zone leaves by design, which is deduplication, not a collision. Two distinct bytecodes mapping to the same stem would need a 244-bit collision on H(code_hash || tree_index), which is infeasible.

Sub-index. The sub-index is storage_key % 256 or the analogous code arithmetic, a direct mapping rather than a hash. Two distinct keys share a sub-index only if they also share a tree_index, in which case they are the same item, so no collision is possible between distinct items.

The hash function selection is the dominant security parameter and remains open. The collision bounds above assume a hash with uniform 256-bit output truncated to the stated widths.

Copyright

Copyright and related rights waived via CC0.

Citation

Please cite this document as:

Vitalik Buterin (@vbuterin), Guillaume Ballet (@gballet), Dankrad Feist (@dankrad), Ignacio Hagopian (@jsign), Kevaundray Wedderburn (@kevaundray), Tanishq Jasoria (@tanishqjasoria), Gajinder Singh (@g11tech), Danno Ferrin (@shemnon), Piper Merriam (@pipermerriam), Gottfried Herold (@GottfriedHerold), Wei Han Ng (@weiihann), Carlos Perez (@CPerezz), "EIP-8297: Partitioned Binary Tree [DRAFT]," Ethereum Improvement Proposals, no. 8297, June 2026. Available: https://eips.ethereum.org/EIPS/eip-8297.