Sui node requirements in 2026: full node and validator hardware

As of October 2026, Sui’s documentation suggests 8 physical cores (16 vCPUs), 128 GB of RAM and a 4 TB NVMe drive for a full node, and 24 physical cores (48 vCPUs), 128 GB, 4 TB of NVMe and 1 Gbps for a validator. Pruning drives disk use: 2.5 TB for a pruned RPC node, 16 TB unpruned (November 2025).

Below: the official figures, what pruning, snapshots and the switch to gRPC change, the ports, and a server for each role. Hardware and operations only, not staking or financial advice.

The short version

Two caveats frame the rest. Without active pruning, Sui warns, a node “can fill 7 TB or more”. And the formal snapshot used to start a node measured about 175 GB on October 15, 2026, by our count, not the “~30GB” the documentation still quotes.

Official Sui full node and validator requirements

Both roles run the same sui-node binary with different configuration files. A validator takes part in consensus; a full node receives certified checkpoints over state sync, re-executes them and serves reads, “a read-only view of the network state”.

• Full node. The full node configuration page lists a “suggested minimum” of “8 physical cores / 16 vCPUs”, 128 GB of RAM and a “4 TB NVMe drive”; the exchange integration guide repeats it. Neither gives a bandwidth figure.
• Validator. The validator deployment page asks for “24 physical cores (or 48 virtual cores)”, 128 GB, “4 TB NVME” and a 1 Gbps network; the Sui for Node Operators guide in the MystenLabs/sui repository matches it line for line.

Neither page gives a recommended tier, a clock speed or a CPU model. Sui recommends Linux, with Ubuntu and Debian supported and macOS “only recommended for development”; Mysten Labs ships linux/amd64 binaries and Docker images, and the docs accept bare metal, virtual machines or Kubernetes. Read the CPU figures as physical cores: a 16-core, 32-thread processor clears the full node figure twice over but misses the validator’s 24 physical or 48 virtual cores.

How much disk a Sui node needs: pruning decides

Two settings in authority-store-pruning-config decide disk use: num-epochs-to-retain for old object versions, and num-epochs-to-retain-for-checkpoints for past checkpoints with their transactions, effects and events. Sui epochs last 24 hours, so each value is roughly a number of days. The documented profiles:

• The template does not prune transactions. The shipped fullnode-template.yaml sets num-epochs-to-retain: 1 but omits num-epochs-to-retain-for-checkpoints, which defaults to none in the node’s source: transaction history is never pruned. Values below 2 are rounded up to 2. Sui advises setting pruning before the first start.
• Space comes back slowly. Pruning deletes logically and RocksDB compaction frees the space later. If usage has not dropped after 3 to 5 days, Sui says to check the configuration and set periodic-compaction-threshold-days: 1.

Sui’s storage page gives Mainnet sizes as of November 2025: about 250–400 GB for a validator; 2.5 TB for a pruned full node with indexes, 1.5 TB of it in an indexes/ directory that grows with throughput; and 16 TB unpruned. Its growth examples for an unpruned node are about 10 GB a day at around 18 transactions per second and 40 GB a day at around 183. For scale, Sui’s Bigtable-based archival service held about 18.1 TB of Mainnet history on July 27, 2026.

Our reading: 4 TB is comfortable for a validator and adequate for a pruned RPC node with a short window. But the 2.5 TB figure predates the 2026 JSON-RPC shutdown and has not been updated, so monitor usage and size up for a long retention window.

Snapshots: formal or RocksDB

Sui says not to sync from genesis, which “takes a very long time and consumes a lot of resources (including likely filling up your disk)”. Start from a snapshot instead:

• Formal snapshots hold only the live object set at the end of an epoch, verified against a commitment signed by that epoch’s validator committee. With no indexes and no history, they suit validators and state sync nodes and are “not suitable” for an RPC node that does historical lookups. Mysten Labs serves them free through sui-tool download-formal-snapshot --no-sign-request, from a bucket “currently only in North America”.
• RocksDB snapshots copy a node’s database, indexes included, and are “equal in size to the database of the full node that produced them”. They come only from Requester Pays buckets on Google Cloud Storage or S3, so you pay the egress.

The storage page puts a formal snapshot at “~30GB for recent Mainnet epochs” (November 2025). Today it is much larger: the 2,886 files listed in the Mainnet manifest for epoch 1280, written on October 15, 2026, add up to about 175 GB by our count, and epoch 1221, from August 17, to about 173 GB. At a sustained 1 Gbps, 175 GB takes about 23 minutes; on a 10 Gbps port, the bucket becomes the limit.

Other discrepancies, with dates: the storage page gives about 1.6 TB for database snapshots in its table and 2.5 TB in its text, both for November 2025 (2.5 TB takes about 5.6 hours at 1 Gbps). On retention, the snapshot page says 90 epochs and the data stores page about 60 on Mainnet; the live manifest listed 60, epochs 1221 to 1280. And the JSON-RPC migration guide ended “JSON-RPC snapshots for full nodes” in August 2026, so check what a RocksDB bucket holds before paying for it.

Archival fallback, and why full history is a cloud job

The Mainnet seed peers “only keep a few epochs of history”, so a node that falls further behind needs the archive, read through state-archive-read-config. On Mainnet that archive is an S3 Requester Pays bucket, and the docs say to “run the node on EC2 with an instance profile”: environment variables, ECS task roles and IRSA “are not loaded by this configuration”. The free checkpoints.mainnet.sui.io endpoint keeps only 30 days, and state sync “explicitly rejects” it.

Off AWS, then, the documented recovery for a lagging Mainnet node is the runbook’s: restore a recent snapshot rather than wait for catch-up. Keep the node updated, watch its checkpoint lag and script the restore. Full history is a different product: Sui’s Archival Store and Service runs on Google Cloud Bigtable, and for years of history that service or a data provider fits better than a single server.

Ports, bandwidth and network tuning

Full node defaults:

Validator ports:

The docs and the GitHub guide both say to open 8080–8084 “including the specific protocol”; the GitHub copy also lists UDP 8081 and 8082 for Narwhal interfaces, which the docs no longer show. The validator page adds larger Linux network buffers, because the defaults “are too small”:

# /etc/sysctl.d/100-sui-node.conf, then apply with: sudo sysctl --system
net.core.rmem_max = 104857600
net.core.wmem_max = 104857600
net.ipv4.tcp_rmem = 8192 262144 104857600
net.ipv4.tcp_wmem = 8192 262144 104857600

Sui specifies 1 Gbps for validators, nothing for full nodes, and no monthly traffic for either. On an unmetered 10 Gbps port, snapshot downloads and public API traffic do not appear on the bill.

RPC nodes in October 2026: gRPC replaces JSON-RPC

The API for application-facing nodes changed this year. From the JSON-RPC migration guide:

A new RPC node therefore needs rpc.enable-indexing: true. The first index build “takes time depending on your hardware” and the node “might not serve other traffic” meanwhile, so plan a rolling cutover. The ListTransactions, ListEvents and ListCheckpoints APIs need those indexes; data beyond your retention window returns NOT_FOUND, and the Archival Service covers it. Put TLS, a firewall and rate limits in front, and do not run production on fullnode.mainnet.sui.io, which Sui says is rate-limited and meant for development.

Which PowerDedi server for a Sui node

Every configuration below has a 10 Gbps unmetered port and month-to-month billing. Running nodes and validators is permitted under our acceptable use policy; mining is not.

You choose the disk layout when ordering or reinstalling; our software RAID guide explains the options.

• Validators: keep the mirror. RAID 1, the default on two-drive servers, still leaves far more space than a validator uses, and a failed drive then does not take it offline. That matters: under Sui’s tallying rule, other validators can cut the rewards of one that performs badly.
• RPC nodes: mirror or stripe. The mirrored 3.84 TB fits Sui’s 2.5 TB figure with room to spare. Striping doubles the space, but a failed drive means a new snapshot restore, and since a formal snapshot carries no history, your local query window restarts from that epoch.
• 3.84 versus 4 TB. 3.84 TB is the usual capacity of an enterprise “4 TB class” drive, 4% under the figure as written.

What does not fit: our 32 and 64 GB servers (Ryzen 5 7600, Ryzen 7 7700, Ryzen 9 9900X, Xeon E-2488) are below the 128 GB both roles ask for; 16-core servers are below the validator’s 24 physical cores; and the storage servers’ capacity is on hard drives, not NVMe. The EPYC 4585PX ($179/mo) is the 9950X’s alternative with ECC memory and the same disks. For full history, the 9965 has the space, but building an unpruned Mainnet node off AWS runs into the archive constraint above; Sui’s Bigtable-based service or a data provider is usually the better route.

Sui’s docs set no latency target or region for either role; put an RPC node near the applications that query it. The 9950X3D is offered in Amsterdam, Frankfurt, New York and Los Angeles, the 9454P adds Singapore, and the 9355P and 9555P are in Amsterdam, Frankfurt, New York and Tokyo. For another high-throughput chain, see our Solana validator and RPC hardware guide.

Running a validator: requirements beyond hardware

• Stake. SIP-39, marked Final, replaces the old 30 million SUI minimum with voting power on a 10,000-unit scale: 3 to join, seven epochs of probation below 2, removal below 1. When it was written in August 2024, one unit cost about 800,000 SUI, so about 2.4 million to join; the amount moves with total stake. The validator page says SIP-39 “is rolling out”, and its setup steps still mention 30M SUI.
• Keys. protocol.key (BLS12-381) for transactions, plus Ed25519 account, network and worker keys. An operation capability can be delegated to another address for gas price quotes and reports, so the account key can stay offline.
• Peers. The default and recommended max-concurrent-connections: 0 lets other validators state-sync from you but no full node, so serve read traffic from a separate full node.
• Upgrades. Mainnet releases ship every two weeks on Wednesdays; the latest is mainnet-v1.81.1 of October 7, 2026, with protocol version 138. Validators enact protocol upgrades by stake, and a full node that misses one “cannot connect to the network”. Urgent security fixes can arrive out of band as signed binaries before the source is published.

How to set up a Sui full node: outline

These steps use only commands from Sui’s documentation; adjust the paths to your fullnode.yaml.

# 1. Install sui-node and sui-tool, then get the Mainnet template and genesis blob
curl -fLO https://github.com/MystenLabs/sui/raw/mainnet/crates/sui-config/data/fullnode-template.yaml
cp fullnode-template.yaml fullnode.yaml
curl -fLO https://github.com/MystenLabs/sui-genesis/raw/main/mainnet/genesis.blob

# 2. Before the first start, set db-path, genesis-file-location, the Mainnet
#    seed-peers from the docs and the settings below

# 3. Restore the latest formal snapshot from the free bucket into db-path
sui-tool download-formal-snapshot --latest --genesis /opt/sui/config/genesis.blob --network mainnet --path /opt/sui/db --num-parallel-downloads 50 --no-sign-request

# 4. Start the node and watch its checkpoint height
sui-node --config-path fullnode.yaml
curl -s http://localhost:9184/metrics | grep highest_synced_checkpoint

# fullnode.yaml settings for an RPC node (retention values from the full node page)
metrics-address: "127.0.0.1:9184"
rpc:
  enable-indexing: true
authority-store-pruning-config:
  num-epochs-to-retain: 14
  num-epochs-to-retain-for-checkpoints: 14
  periodic-compaction-threshold-days: 1

The template already has metrics-address and pruning entries, so edit them rather than adding duplicates; for a node without RPC traffic, use 0 and 2 and leave indexing off. Run the node under systemd as an unprivileged user, let only 8084/UDP and your TLS proxy through the firewall, and follow our first-hour hardening checklist. Current stock is on our blockchain node hosting page.

Common questions

Is 64 GB of RAM enough for a Sui full node? Not by the documentation: 128 GB is the suggested minimum for full nodes and validators alike, and Sui publishes no lower tier.

Can I run a Sui validator on a 16-core Ryzen? Not to the letter. The validator figure is 24 physical cores or 48 virtual cores, and a 16-core, 32-thread CPU meets neither.

Do I need my own node to build on Sui? Not for development. For production, Sui’s exchange integration guide says to run your own full node or use a provider whose contract meets your throughput and uptime needs, not the rate-limited public endpoints.

Sources

• Sui Full Node Configuration — Sui Documentation
• Validator Deployment and Configuration and Validator Node Management — Sui Documentation
• Storage — Sui Documentation (Mainnet figures as of November 2025)
• Data Management — Sui Documentation
• Database Snapshots and Available Data Stores — Sui Documentation
• Sui Archive Data and Running the Archival Store and Service — Sui Documentation (storage measured July 27, 2026)
• JSON-RPC Migration Guide — Sui Documentation
• Updating a Full Node and Exchange Integration — Sui Documentation
• Sui for Node Operators, fullnode-template.yaml and node.rs — MystenLabs/sui on GitHub
• mainnet-v1.81.1 — MystenLabs/sui releases on GitHub (October 7, 2026)
• SIP-39: Lowering the Validator Set Barrier to Entry — Sui Improvement Proposals on GitHub
• Mainnet formal snapshot manifests — Mysten Labs (available epochs, and file sizes for epochs 1221 and 1280 summed by PowerDedi on October 15, 2026)