How good are ARM servers like Ampere Altra for RandomX?

ARM servers like Ampere Altra can run RandomX, Monero’s mining algorithm, but they are usually a poor purchase if mining is your main goal. You get many CPU cores, yet Altra’s small shared cache limits how well those cores turn into mining speed. If you already own one, a careful test is more useful than judging it by core count.

Why Altra’s cache matters so much

RandomX is designed to favor CPUs, the main processors in your computer or server. It does not reward core count alone: it also needs fast access to memory.

Each mining thread—one worker running on your CPU—needs a roughly 2 MB working area. For strong performance, you generally want about 2 MB of L3 cache per active thread. L3 cache is a small, fast memory area inside the processor.

A standard Ampere Altra processor has up to 80 cores, with 1 MB of private L2 cache per core and 32 MB of shared L3 cache. Altra Max offers up to 128 cores but has 16 MB of shared L3.

That creates a mismatch for RandomX:

• You have many cores available to do calculations.
• You do not have enough shared L3 to comfortably support one mining thread per core.
• Adding threads can increase memory traffic without adding much useful mining speed.

The 2 MB guideline is not a hard limit. Private caches and memory performance also matter. But it explains why an 80-core Altra should not be treated like an 80-thread mining powerhouse.

ARM support is not the main problem

ARM is a processor architecture—the basic instruction system a CPU uses. RandomX supports 64-bit ARM, and native mining software such as XMRig can run on suitable Linux ARM servers.

Use an up-to-date ARM64 build with RandomX’s optimized execution support. An outdated or poorly configured build can make the hardware look worse than it really is.

RandomX’s fast mode uses a 2,080 MB dataset, plus a 256 MB cache and working memory for mining threads. The dataset is normally shared by threads; you do not need a separate 2 GB copy for every core. Some configurations use extra copies to keep memory close to each processor.

Altra servers generally have ample RAM capacity. Their challenge is feeding many active cores efficiently, not simply fitting the dataset into memory.

How you should test one

Start with a modest thread count, then increase it in steps. Do not assume using every core gives you the best result.

Track three things:

• Hashrate: how many mining calculations your server completes each second.
• Wall power: electricity used by the whole server, including fans and memory.
• Hashes per watt: how much mining work you get for that power.

Enable huge pages if your operating system supports them. These are larger memory blocks that can reduce memory-management overhead. On multi-processor servers, keep workers and their memory on the same processor where possible.

Compare your measured results with similarly priced AMD Ryzen or EPYC systems. Their larger caches often make them stronger RandomX choices.

The bottom line

An existing Altra server is worth testing with native software, but buying or renting one specifically for Monero mining is usually hard to justify. Check electricity costs and a live mining calculator before spending money.

For browser mining, native benchmark results are not a reliable earnings estimate: browser software has different execution and memory limits. Also check your hosting provider’s rules before mining on rented hardware.