Where to Buy an AMD EPYC 9654 Dedicated Server

Where to Buy an AMD EPYC 9654 Dedicated Server
Published on Jul 2, 2026 Updated on Jul 3, 2026

A workload that keeps demanding more cores eventually runs out of room on smaller processors. You can keep adding servers, but every one is another license, another OS to patch, another point of failure. Fewer, larger machines do the same work with less to manage.

That’s the case for an AMD EPYC 9654 dedicated server, 96 Genoa cores on single-tenant bare metal, enough to fold several smaller hosts into one. The catch is that all 96 draw on the same memory, so the build decides whether you get full throughput or stalled cores.

So buying one comes down to two questions. Does the workload actually scale to 96 cores, or would fewer cores handle it? And which provider pairs them with the memory, storage, and network to put every core to work, in the region you need?

This guide answers both, walking you through buying an AMD EPYC 9654 dedicated server. Before that, we cover the 9654’s specifications, workloads, comparisons, and what to look for in a server.

#AMD EPYC 9654 Specifications

With 96 cores and 192 threads, the EPYC 9654 is the high-core option in AMD’s Genoa line and is available on an AMD dedicated server. It pairs 384 MB of L3 with 12-channel DDR5-4800 and 128 PCIe 5.0 lanes, inside a 360 W TDP.

Spec Value
Architecture/codename Zen 4 “Genoa” (EPYC 9004 Series)
Cores/threads 96/192
Base clock 2.4 GHz
All-core boost 3.55 GHz
Max boost Up to 3.7 GHz
L3 cache 384 MB
Default TDP 360 W (cTDP 320-400 W)
Socket SP5, 1P/2P
Memory DDR5 ECC, 12 channels
Memory speed Up to 4800 MT/s (DDR5-4800)
Memory bandwidth ~461 GB/s per socket
Memory capacity Up to 6 TB per socket (architectural max)
PCIe PCIe 5.0, 128 lanes
Instruction sets AVX, AVX2, AVX-512
Security AMD Infinity Guard (SME, SEV, SEV-SNP)

All 96 cores share one 12-channel DDR5 controller. At 461 GB/s per socket, the total splits across all 96 sockets, so memory-bound loads can stall while compute-bound loads run flat out.

Powerful AMD Dedicated Servers

Deploy AMD-based bare metal servers with high-core processors, fast memory, and reliable storage. Perfect for compute-heavy workloads and instant setup.

#Best Workloads for AMD EPYC 9654 Dedicated Servers

AMD EPYC 9654 is built for workloads that use many cores at once: virtualization, AI and HPC, databases and analytics, and blockchain. Each scales the same way, by running more VMs, tasks, queries, or nodes at once.

#Virtualization and Kubernetes

One EPYC 9654 server holds 96 cores and 192 threads, enough to run virtual machines and Kubernetes nodes that would otherwise need several smaller servers. Running them on a single processor reduces rack space, license counts, and the overhead of managing additional hardware.

Because the VMs share one processor, the hardware itself has to keep them separate. SEV-SNP gives each guest its own memory encryption key, so a neighboring VM or a compromised hypervisor cannot read what another guest holds.

Each guest draws memory bandwidth from the 12 DDR5 channels and storage and network I/O from the 128 PCIe 5.0 lanes. You run out of memory well before the cores. A single server is also one failure away from taking everything down, so high availability still needs a second machine.

#AI and High-Performance Computing

Scientific simulations break a computation into pieces that compute concurrently. EPYC 9654 has 96 cores running CFD, finite-element analysis, molecular dynamics, and Monte Carlo sweeps through MPI or OpenMP. AVX-512 speeds the vector math inside each core. Those same vector units also run batch inference and scoring, so that light AI workloads won’t need a separate accelerator.

Memory-bound simulations gain less from the high core count than compute-bound ones, because the 461 GB/s is divided among all 96 cores. Large model training runs on GPUs instead, which attach over the 128 PCIe 5.0 lanes.

#Large Databases and Analytics

Analytics scales with the number of queries running at once, and the EPYC 9654 runs them across 96 cores. It holds large datasets entirely in memory, up to 6 TB per socket. The 384 MB L3 caches hot rows and indexes under a busy query mix.

The heaviest scans, though, are bandwidth-bound. A full-table sweep reads at the socket's 461 GB/s rate, regardless of how many cores are waiting. Its capacity and core count pay off when data spans terabytes and queries execute in parallel, while latency-bound OLTP runs on fewer, faster cores.

Registered ECC DDR5 automatically corrects single-bit errors, protecting in-memory records from silent corruption.

#Blockchain Infrastructure

A single node rarely needs 96 cores. An operator running many of them does. EPYC 9654 packs many small nodes onto one machine, with chain clients, RPC endpoints, and indexers running side by side. Public RPC throughput increases with core count, and parallel indexers can cover several chains at once.

Validators depend on a single-thread clock speed, so a higher-frequency processor serves them better than 96 slower cores. The core count is seldom the first limit anyway, since storage I/O, memory, and network throughput run out before it does. Size those for the fleet, and one EPYC 9654 runs the web3 infrastructure that once needed a rack of nodes.

#AMD EPYC 9654 vs Other CPUs

AMD EPYC 9654 has more cores than the 64-core 9554 and the 48-core 9454 in AMD’s standard Genoa line. All three run on the same SP5 platform, with DDR5-4800 across 12 channels, 128 PCIe 5.0 lanes, AVX-512, and 1P or 2P. Cores, clock speeds, cache, and power are where they differ, as the table shows.

9454 9554 9654
Cores/threads 48/96 64/128 96/192
Base clock 2.75 GHz 3.1 GHz 2.4 GHz
All-core boost 3.65 GHz 3.75 GHz 3.55 GHz
Max boost Up to 3.8 GHz Up to 3.75 GHz Up to 3.7 GHz
L3 cache 256 MB 256 MB 384 MB
Default TDP (cTDP) 290 W (240-300 W) 360 W (320-400 W) 360 W (320-400 W)

At the same 360 W as the 9554, EPYC 9654 trades clock speed for cores and cache. The 64-core 9554 runs at a 3.1 GHz base and 3.75 GHz all-core, higher than the 9654’s 2.4 GHz and 3.55 GHz. EPYC 9654 has 384 MB of L3 against 256 elsewhere, and 96 cores to the 9554’s 64.

When a job can be parallelized well, the EPYC 9654 runs all 96 cores. If per-core speed outweighs core count, the 9554's 64 cores clock higher. The 9454 draws the least power, 290 W, when the workload needs no more than 48 cores.

An EPYC dedicated server costs more as cores and power increase, from the 9454 up to the 9654.

Bergamo’s 9754 reaches 128 cores on a dense Zen 4c platform, at a lower per-core clock. Turin, the next Zen 5 generation, improves single-thread performance and uses DDR5-6400, at a higher price. Milan, the prior generation, lacks AVX-512 and runs DDR4-3200 with PCIe 4.0, leaving the 9654 as the Genoa step-up.

#What to Look for in an EPYC 9654 Dedicated Server

EPYC 9654 gives you 96 cores, but how many run at full output depends on the rest of the server. Memory population, NVMe endurance, the egress allowance, and single-tenant access each decide how much of that compute you get to use. Get them right, and the processor never waits on the surrounding hardware or on another tenant.

#DDR5 Memory Capacity

Capacity has to be sufficient for the data in active use, and 96 busy cores keep more of it live than fewer would. The base build comes with 768 GB of registered ECC DDR5, scaling to 1152 GB.

That settles the capacity. Bandwidth depends on how that memory is installed. The 12-channel controller reaches its full DDR5-4800 rate, around 461 GB/s, only with a matching DIMM in every channel.

Leave one empty, and the rate drops. That drop costs more on the 9654 than on a smaller processor, because 96 cores wait on the reduced bandwidth instead of 32. So no channel goes empty.

#NVMe Storage Performance

On a busy server, storage is a random-access problem. Databases, queues, and concurrent services scatter small reads and writes across the disk, the pattern that left spinning disks stalling on every seek. NVMe handles that pattern, with latency low enough that the disk is no longer the bottleneck.

The drives have requirements of their own:

  • Endurance: Write-heavy workloads churn through consumer NVMe, so a 9654 build needs datacenter-class drives rated for sustained writes.
  • Redundancy: With no second node to fail over to, a RAID array across the drives keeps the data when one fails.

EPYC 9654 ships with two 4 TB NVMe drives, expandable to 10, for a total of 160 TB with larger drives. That is room for arrays wide enough to keep that redundancy as capacity grows.

#Network Bandwidth and Egress

Outbound traffic is where hosting bills surprise people. Metered by the gigabyte, a busy public service or a blockchain node racks up egress charges fast. A large monthly allowance takes that cost off the table.

EPYC 9654 includes 100 TB of free egress a month, with inbound unmetered. The uplink runs at 3 Gbps, with up to 10 Gbps available through the bandwidth plans when a workload needs it. Past the allowance, overage is billed by the terabyte at a low rate, whereas metered clouds bill by the gigabyte.

Throughput is only half the story. Route quality, peering, and proximity to your users influence latency as much as the port speed. Standard DDoS protection filters attack traffic before it reaches the server.

#Single-Tenant Access

You get a physical machine of your own, with no other customer sharing it. All 96 cores, every memory channel, and the 128 lanes of PCIe 5.0 run your workloads, with no virtualization.

On a shared host, another tenant’s load competes for the same cores and cache, and the hypervisor spends cycles of its own. Single-tenant removes that contention, so the memory bandwidth, NVMe, and egress reach your workload in full.

Ensure the server is bare metal and single-tenant before you purchase. A “dedicated” label sometimes covers a VPS with reserved cores, which still shares hardware with other tenants.

#Bare Metal vs Cloud for AMD EPYC 9654 Workloads

Bare metal or cloud for a 9654 workload comes down to how steadily the machine runs. A workload that keeps the cores busy day and night belongs on bare metal. Intermittent load goes to the cloud, where you pay only for the hours you use.

The comparison is easy to misread, because cloud instances are sold in vCPUs. For common instance types, a vCPU is a hardware thread, so two of them share a single physical core. A 96-vCPU instance is closer to 48 physical cores, and matching the full 96 requires a node twice that size.

On bare metal, the EPYC 9654 has a single fixed price for the entire machine, whether flat-out or idle. The cloud instance is billed by the hour, so running it full-time results in a bill that grows with usage and is hard to forecast. When that gap grows wide, teams move steady work back onto bare metal. This is known as cloud repatriation.

Operations point the other way. The cloud comes with managed databases, load balancers, and autoscaling, so a small team ships without running the plumbing. With bare metal, you get root access and the full stack instead, which means you patch, monitor, and recover it yourself. An unmanaged server is the right call only when you want that control and have the in-house expertise to run it.

The cloud handles bursty, short-lived, or globally spread work, where the bill tracks the hours, and managed services cut the effort. An EPYC 9654 on steady, full-time load provides 96 dedicated cores at one fixed price.

#Where to Buy AMD EPYC 9654 Dedicated Servers

With the AMD EPYC 9654 chosen, the provider is the next decision to make. That decision sets how the server runs, how fast it deploys, and how much you pay for it.

Cherry Servers runs EPYC 9654 builds across four locations: Siauliai, Stockholm, Frankfurt in Europe, and Chicago in the US. Not every site keeps every model in stock, so check that your target location lists the EPYC 9654 before you build.

A pre-built EPYC 9654 deploys in roughly 12 minutes; a custom configuration is ready in 24 to 72 hours. Every build shows its full price before you order, with no setup cost or hidden add-ons.

You can run an EPYC 9654 by the hour for short projects or on a fixed term for production, with monthly or annual commitments. The longer the term, the lower the rate. Over 20 payment methods are accepted, including cards, bank transfer, and cryptocurrency.

Provisioning is API-first, with Terraform and Ansible providers, a command-line tool, and SDKs that slot into the pipelines you already run. Technical support is available 24/7 over chat, phone, and email, and every account comes with a dedicated point of contact at no cost.

DDoS protection is standard on every plan. A money-back guarantee covers the first 15 days, with a full refund if you cancel.

Price and configure an EPYC 9654 server on the plan page, then deploy on your own schedule.

#Conclusion

AMD EPYC 9654 combines 96 Zen 4 cores, 384 MB of L3 cache, and twelve channels of DDR5. That core count and cache suit virtual machine and container consolidation, parallel scientific computing, large in-memory analytics, and multi-chain node fleets.

Virtualized cloud splits those 96 cores among tenants, and shared memory channels add contention under load. A bare-metal server dedicates the full socket to one tenant, so cores and cache work without interference, at a fixed rate. Workloads that keep all 96 cores in use are the ones that justify EPYC 9654 on bare metal.

FAQs

How much memory bandwidth does the AMD EPYC 9654 have?

Around 461 GB/s, from twelve DDR5-4800 channels. That figure is the socket total, shared across all 96 cores. A workload that uses every core therefore has less bandwidth per core than the same workload would on a processor with fewer cores.

Does the EPYC 9654 support AVX-512?

Yes. EPYC 9654 runs AVX-512 on its Zen 4 cores, with the VNNI and bf16 extensions that accelerate inference and other vector math. Milan, the generation before Genoa, reached only AVX2.

What is the difference between the EPYC 9654 and 9654P?

Socket support. EPYC 9654 runs in single- or dual-socket configurations (1P/2P), while the 9654P is limited to one socket (1P). The rest is identical: 96 cores, 192 threads, a 2.4 GHz base clock, and 384 MB of L3 cache.

Bare Metal Servers - 12 Minute Deployment

Get 100% dedicated resources for high-performance workloads.

Share this article

Related Articles

Published on Jun 25, 2026 Updated on Jun 26, 2026

High-Performance Dedicated Servers: 6 Providers to Consider

This guide explores dedicated server options with high performance. We list some examples of the best high-performance dedicated servers to consider.

Read More
Published on Jun 21, 2026 Updated on Jun 22, 2026

How to Choose a 10Gbps Dedicated Server

Learn what a 10Gbps dedicated server really delivers. Compare hardware, bandwidth, pricing, providers, and performance tips to choose the right setup.

Read More
Published on Jun 18, 2026 Updated on Jun 19, 2026

Where to Buy an AMD Ryzen Dedicated Server

Discover the best AMD Ryzen dedicated servers for gaming, hosting, and development. Compare CPUs, features, providers, pricing, and performance.

Read More
No results found for ""
Recent Searches
Navigate
Go
ESC
Exit