One processor. 16 cores at up to 5.7 GHz.

Every server we sell runs on the AMD EPYC 4584PX, with DDR5 memory and NVMe storage. This page is what that means, using AMD's figures and nothing we made up.

Boost clock
5.7GHz
L3 cache
128MB
Cores / threads
16/ 32

Figures from AMD's EPYC 4584PX page.

Two compute dies and an I/O die under one lid. The 16 cores sit on two dies of 8. Memory, disks and the network are reached through the third.

AMD builds this class of processor from separate pieces of silicon on one package. The 16 cores sit on two compute dies of 8 each. A third die, the I/O die, holds the memory controller and the PCI Express lanes, and the compute dies reach memory, disks and the network through it.

What the drawing shows

  1. Compute die 1 and 2. 8 cores each, 16 in total, 32 threads.
  2. L3. 128 MB in total across the package, per AMD. AMD's page gives the total, so the drawing does not split it.
  3. I/O die. 2 channels of DDR5 (up to 5200 MT/s) and PCIe 5.0 x28.

That split is why our core diagrams elsewhere on the site draw the processor as two groups of 8. When a VDS is allocated whole cores, they come from these two dies.

Schematic of the AMD EPYC 4584PX package: two compute dies with 8 cores each, 16 cores and 32 threads in total, sharing 128 MB of L3 cache across the package, connected to one I/O die that provides 2 DDR5 memory channels (up to 5200 MT/s) and PCIe 5.0 x28.AM5 package, schematicCompute die 1L3Compute die 2L3I/O dieDDR52 channelsPCIe 5.0x28 lanes

The package, as a schematic

Schematic, not a die photograph: positions and sizes are illustrative. Figures from AMD.

The full datasheet. Every field AMD publishes for the EPYC 4584PX, as AMD publishes it.

Processor
AMD EPYC 4584PX
Family
AMD EPYC 4004 series
Cores
16
Threads
32
Base clock
4.2 GHz
Boost clock
Up to 5.7 GHz
L3 cache
128 MB
Default TDP
120 W
Socket
AM5
Socket count
1P
PCI Express
PCIe 5.0 x28
Memory type
DDR5
Memory channels
2
Memory speed
Up to 5200 MT/s

Derived from the rows above: 2 threads per core (32 / 16), and a boost clock 36% above base (5.7 / 4.2).

Where your data waits. Each step from the core holds more and sits further away.

Order only, not to scale. Figures from AMD's product page.

  1. Cores16 cores, 32 threads

    4.2 GHz base, up to 5.7 GHz boost. Where your code runs.

  2. L3 cache128 MB

    On the processor, shared by its cores. Data the cores keep asking for is served from here instead of from memory.

  3. MemoryDDR5, 2 channels

    Up to 5200 MT/s. Everything your processes have loaded lives here.

  4. StorageNVMe

    Drives attached over PCI Express. The processor provides PCIe 5.0 x28. Everything that has to survive a restart lives here.

Why one fast core still matters. Extra cores run more work side by side. Only a faster core makes any one piece of it finish sooner.

The same step, at two clocks

A lot of server work happens one step at a time on a single thread: an event loop answering requests, a database executing a query, a compiler working through one file. Extra cores let you run more of those side by side, but they do not make any one of them finish sooner. The speed of one core does.

So the clock of that core sets how much work fits in each step. That is why we chose a processor that boosts to 5.7 GHz, and why, for single-thread-bound work, the clock of one core often matters more than how many cores there are.

At the 4.2 GHz base clock3.00 steps in the window

At up to 5.7 GHz boost4.07 steps in the window

One thread doing the same step of work over and over (a request, a query, a loop iteration), drawn to scale by clock speed alone: at 5.7 GHz a clock-bound step takes 4.2 / 5.7 of the time it takes at 4.2 GHz, about 26% less, so 1.36 steps fit where one did. This is arithmetic from AMD's two published clocks, not a benchmark. Real steps also wait on memory, disk and the network, and the boost clock is a ceiling the processor reaches when conditions allow, not a constant.

Event loops and request handlers

Node.js runs your JavaScript on one thread per process, and a PHP or Python worker handles one request at a time. A faster core shortens the queue that forms behind a slow handler.

Databases and caches

Redis executes commands one after another on its main thread, and a single SQL query usually runs on one core unless the planner splits it. A faster core finishes each one sooner.

Build steps and game server loops

Compiling one source file runs on one core, however many the machine has. The same holds for a game server's main loop, which advances in fixed ticks that must each finish before the next is due.

Memory, disks and network, in plain words. What a program works on, what survives a restart, and how traffic arrives.

DDR5 memory

Memory is where a running program keeps what it is working on: open sessions, a cache, the database's hot rows. DDR5 is the current generation of server memory, and this processor drives it on 2 channels at up to 5200 MT/s.

How much of it you get depends on the plan: from 2 GB on the smallest VPS to 64 GB on the largest VDS, and whatever you ask for on bare metal.

Type
DDR5
Channels
2
Speed
Up to 5200 MT/s

NVMe storage

NVMe drives talk to the processor over PCI Express, the same fast lanes a graphics card uses, instead of through the older SATA interface built for spinning disks. For a server that means databases, uploads and logs are read and written without an interface designed for a slower kind of drive in the way. NVMe, SSD and HDD compared goes further.

The processor provides 28 lanes of PCIe 5.0 in total. We do not publish drive speeds or IOPS, because they depend on the drive model and we have not published one.

Interface
NVMe over PCI Express
Processor lanes
PCIe 5.0 x28
Disk per plan
20 to 1500 GB

A 25 Gbit/s network card in every host

Every packet for your server arrives through this card, and Prism's XDP/eBPF program runs in its driver, so attack traffic is dropped before the operating system's network stack sees it. On a VPS or VDS your plan has a 1 Gbit/s port on that card. On bare metal the whole card is part of the machine you rent.

Host network card
25 Gbit/s
VPS and VDS port
1 Gbit/s
Filter
Prism, XDP/eBPF in the card's driver
NIC
25 Gbit/shost network card
Driver: Prism, XDP/eBPF
  • VPS or VDS1 Gbit/s port
  • Bare metalThe whole card

The same machine, rented three ways. The hardware does not change between products. How much of it is yours does.

vCPUs scheduled across shared cores
4 of 16 cores allocated to you
All 16 cores are yours

AMD's published figures for the AMD EPYC 4584PX, the same under every plan.

On two compute dies
16cores
2 per core
32threads
Base clock
4.2GHz
Boost clock, when conditions allow
5.7GHz
L3 cache on the package
128MB
Default TDP
120W

About the hardware

Is every plan on this processor?

Yes. Every VPS and VDS plan runs on the AMD EPYC 4584PX with DDR5 memory and NVMe storage, in Mumbai, Singapore and Limburg, and it is the standard bare metal build. If you need other bare metal hardware, we can source it on request, depending on availability.

Will my server always run at the boost clock?

No. 5.7 GHz is the top of the range AMD publishes, reached when conditions allow. 4.2 GHz is the base clock. We do not publish a sustained figure because we have not measured one we can stand behind.

How many cores do I get?

A VPS gets 1 to 12 vCPU scheduled on shared cores. A VDS gets 2 to 16 cores allocated to it. Bare metal gets all 16 cores and 32 threads.

Why no benchmark numbers on this page?

Because we would have to run them on our own machines under conditions you cannot check. Every figure here is AMD's published specification, linked to AMD's page, or arithmetic from those figures that the page explains.

What is the difference between cores and threads?

Each of the 16 cores can run 2 threads at once, which is how the processor reaches 32 threads. Two threads on one core share that core's execution resources, so a thread is not a second core.

Pick how much of the machine you need.

Every plan runs on the AMD EPYC 4584PX in Mumbai, Singapore and Limburg.