CPU
16 core minimum
Required instruction sets plus a passing result on the production host.
Browse CPU results →Consensus performance depends on the complete system. Start with a benchmark-proven CPU, then give state and application data fast, isolated storage.
16 core minimum
Required instruction sets plus a passing result on the production host.
Browse CPU results →32 GB minimum
Use reliable, high-speed memory. ECC is preferred for production validator hosts where supported.
2 TB+ NVMe
Use an independent OS disk plus a separate, high-speed NVMe drive of at least 2 TB for node data.
View NVMe options →1 Gbps minimum
Use a stable, low-latency symmetric connection with enough sustained headroom for peak validator traffic.
Results reported in issue #6046 for Celestia's 128 MB / 6s benchmark. Readiness combines all six consensus operations and weights them by their reference-time contribution.
| Result | Manufacturer | Processor model | Cores | Threads | Base | Boost | iWorkload-weighted against the reference timings in Celestia's cpu_requirements benchmark. |
|---|---|---|---|---|---|---|---|
| ✓ Compatible | AMD | Ryzen 9 7950X | 16 | 32 | 4.5 GHz | 5.7 GHz | 1.7× faster |
| ✓ Compatible | AMD | Ryzen 9 7950X3D | 16 | 32 | 4.2 GHz | 5.7 GHz | 1.9× faster |
| ✓ Compatible | AMD | Ryzen 9 9900X | 12 | 24 | 4.4 GHz | 5.6 GHz | 2.4× faster |
| ✓ Compatible | AMD | Ryzen 7 PRO 8700GE | 8 | 16 | 3.65 GHz | 5.1 GHz | 1.7× faster |
| ✓ Compatible | AMD | Ryzen 7 3700X | 8 | 16 | 3.6 GHz | 4.4 GHz | 1.2× faster |
| ✓ Compatible | AMD | Ryzen 9 3900X | 12 | 24 | 3.8 GHz | 4.6 GHz | 1.1× faster |
| ✓ Compatible | AMD | Ryzen 5 7600 | 6 | 12 | 3.8 GHz | 5.1 GHz | 2.2× faster |
| ✓ Compatible | AMD | Ryzen 7 7700X | 8 | 16 | 4.5 GHz | 5.4 GHz | — |
| ✓ Compatible | AMD | Ryzen 9 7900X | 12 | 24 | 4.7 GHz | 5.6 GHz | — |
| ✓ Compatible | AMD | Ryzen 9 7900 | 12 | 24 | 3.7 GHz | 5.4 GHz | 1.7× faster |
| ✓ Compatible | AMD | EPYC 4344P | 8 | 16 | 3.8 GHz | 5.3 GHz | 2.2× faster |
| ✓ Compatible | AMD | EPYC 4464P | 12 | 24 | 3.7 GHz | 5.4 GHz | 2.0× faster |
| ✓ Compatible | AMD | EPYC 4545P | 16 | 32 | 3.0 GHz | 5.4 GHz | 2.2× faster |
| ✓ Compatible | AMD | EPYC 4584PX | 16 | 32 | 4.2 GHz | 5.7 GHz | 2.1× faster |
| ✓ Compatible | AMD | EPYC 7313P | 16 | 32 | 3.0 GHz | 3.7 GHz | 1.6× fastermixed runs |
| ✓ Compatible | AMD | EPYC 7452 | 32 | 64 | 2.35 GHz | 3.35 GHz | 1.5× faster |
| ✓ Compatible | AMD | EPYC 7543P 32C/64T | 32 | 64 | 2.8 GHz | 3.7 GHz | 1.6× faster |
| ✓ Compatible | AMD | EPYC 9124 | 16 | 32 | 3.0 GHz | 3.7 GHz | 1.5× faster |
| ✓ Compatible | AMD | EPYC 9454P | 48 | 96 | 2.75 GHz | 3.8 GHz | 1.7× faster |
| ✓ Compatible | AMD | EPYC 9555P | 64 | 128 | 3.2 GHz | 4.4 GHz | 2.4× faster |
| ✓ Compatible | Intel | 13th Gen Core i5-13500 | 14 | 20 | 2.5 GHz P | 4.8 GHz | 1.8× faster |
| ✓ Compatible | Intel | Core i7-1270P | 12 | 16 | 2.2 GHz P | 4.8 GHz | 1.5× faster |
| ✓ Compatible | Intel | Xeon E-2386G | 6 | 12 | 3.5 GHz | 5.1 GHz | 1.6× faster |
| × Below target | AMD | Ryzen 9 5950X | 16 | 32 | 3.4 GHz | 4.9 GHz | 1.3× fastermixed runs |
| × Below target | AMD | EPYC 7443P 24-Core | 24 | 48 | 2.85 GHz | 4.0 GHz | 1.3× faster |
| × Below target | Intel | Xeon Platinum 8474C | 48 | 96 | 2.5 GHz | 3.8 GHz | 0.8× fastermixed runs |
| × Below target | Intel | Xeon Gold 6254 | 18 | 36 | 3.1 GHz | 4.0 GHz | 0.6× faster |
| × Below target | Intel | Xeon Silver 4210 | 10 | 20 | 2.2 GHz | 3.2 GHz | — |
| × Below target | Intel | Xeon E5-1620 | 4 | 8 | 3.6 GHz | 3.8 GHz | 0.5× faster |
| × Below target | Intel | Dual Xeon Silver 4116 | 24 | 48 | 2.1 GHz | 3.0 GHz | 0.4× faster |
How readiness is calculated. Decentra uses a workload-weighted aggregate: total reference time divided by total measured time across PrepareProposal, ProcessProposal, FinalizeBlock, ProposeBlock, EncodeBlock, and DecodeBlock. This gives slower, longer operations more influence than a simple average. A 1.9× result means the reported workload completed about 1.9 times faster overall; it does not mean every operation was 1.9× faster. “Mixed runs” marks materially different reports for the same model. A CPU can score above 1× and still miss an individual deadline, so test the exact production host. Core counts and advertised base/boost clocks come from AMD and Intel specifications; hybrid Intel base clocks refer to P-cores.
Use these as practical starting options. Prefer enterprise endurance and power-loss protection for production, and validate sustained random-write behavior on the exact server.
| Manufacturer | Drive | Class | Capacity | Interface | Claimed random write |
|---|---|---|---|---|---|
| Samsung | 9100 PRO | Consumer | 2 TB+ | PCIe 5.0 | 2,600K IOPS |
| Samsung | 980 PRO | Consumer | 2 TB | PCIe 4.0 | 1,000K IOPS |
| Samsung | PM9A1 | OEM | 2 TB | PCIe 4.0 | 850K IOPS |
| Micron | 9550 MAX | Enterprise | 3.2–6.4 TB | PCIe 5.0 | 540–640K IOPS |
| Samsung | PM9A3 | Enterprise | 3.84 TB | PCIe 4.0 | 180K IOPS |
| Kioxia | KXG80ZNV2T04 | OEM | 2.05 TB | PCIe 4.0 | 620K IOPS |
Selection note: Headline sequential speed is not enough. Favor high endurance, sufficient cooling, and stable sustained random-write performance. Qualify every drive on your own workload before production.
Fibre runs alongside celestia-appd, stores and serves blob shards, and has its own networking, memory, and storage behavior. Validator CPUs must also support GFNI and SHA-NI and pass the benchmark.
Separating Fibre and application data prevents growth in one service from exhausting storage needed by the other.
Set --home or FIBRE_HOME to a path on a separate physical disk whenever possible.
Default connection caps can consume roughly 27 GiB at peak. 64 GB+ system RAM provides more practical headroom.
Application and signer gRPC links are not TLS-protected. Keep them host-local or on a trusted private network.
Make TCP port 7980 publicly reachable. Fibre's client connection uses automatic TLS 1.3.
Fibre retains shards for a limited window; it is not permanent blob storage. Capacity requirements vary with stake, traffic, and the network's shard-retention parameter.
The official tool runs 20 iterations of consensus steps using a 128 MB block, compares median operation times against reference targets, and returns a readiness assessment for the exact host being tested.
git clone https://github.com/celestiaorg/celestia-app.gitcd celestia-appgo run ./tools/cpu_requirements # ORdocker run ghcr.io/celestiaorg/128mb-6s-benchmark:v0.0.1