Best CPUs for Monero Mining in 2026: RandomX Hashrate Compared

“Which CPU should I buy to mine Monero?” is one of the most repeated questions in mining forums, usually followed by a screenshot of a hashrate that looks too low. The answer depends less on clock speed than most people expect. RandomX, Monero’s proof-of-work, cares mostly about cache size, real cores and memory channels. This guide explains what matters, compares real benchmark numbers and looks ahead to RandomX v2.

Why RandomX is a CPU algorithm

RandomX was introduced in Monero’s November 2019 hard fork to keep mining on ordinary computers. Each hash runs a randomly generated program on a virtual machine that resembles a general-purpose CPU: integer and floating-point math, branches and lots of random memory reads. Specialized chips gain little because they would need to rebuild most of a CPU anyway.

Two memory modes exist, according to the RandomX README:

• Fast mode needs about 2,080 MiB of shared memory for the dataset. This is what miners use.
• Light mode needs only 256 MiB but is much slower. It is meant for verifying blocks, not mining.

What makes a CPU fast at RandomX

1. L3 cache: about 2 MB per thread

XMRig’s optimization guide says each mining thread needs about 256 KB of L2 and 2 MB of L3. Your useful thread count is roughly your L3 size divided by 2 MB. That is why a Ryzen 7 5700X (8 cores, 32 MB L3) can run 16 threads at full speed, while an older chip with 8 MB of L3 is limited to about 4.

2. Real performance cores

Hyper-threading helps only while there is enough cache to share. Intel’s hybrid designs mix performance cores with smaller efficiency cores, which contribute less per thread.

3. Memory channels

RandomX reads the 2 GB dataset constantly. The same guide puts the ceiling at roughly 4,000–6,000 H/s per DDR4 channel. Dual-channel memory is mandatory on desktops, and server platforms with 8 or 12 channels pull far ahead.

4. Hardware AES and a modern instruction set

All recent x86 and ARMv8 chips have hardware AES. Newer AMD cores also benefit from XMRig optimizations; 6.26.0 added VAES-512 for Zen 5.

RandomX hashrate by CPU

The table below uses the best results submitted to the XMRig benchmark (RandomX, multi-thread) as of October 2026. These are top submissions, often from tuned systems with huge pages, the MSR mod and good RAM. Expect a stock system to land somewhat lower.

Desktop CPUs

Workstation and server CPUs

Apple and ARM

A few patterns stand out:

• 16-core Ryzen 9 chips are the desktop sweet spot. Each generation adds a little, but a 5950X still delivers over 80% of a 9950X.
• 3D V-Cache does not pay off. The 7800X3D scores below the plain 7700X, and the X3D variants of the 16-core chips are no faster than the regular ones. Once each thread has its 2 MB of L3, extra cache adds little, and X3D parts tend to run at lower clocks.
• Intel is competitive only at the top end. The i9-13900KS and 14900K are respectable, but they draw a lot of power to get there.
• Apple chips are efficient but not fast for the price, and you cannot add memory channels to a MacBook.

Note: Treat any single benchmark number with care. Sample counts differ a lot between CPUs, and a result from an overclocked system with exotic RAM is not what you will see on an off-the-shelf PC.

Best CPUs by budget and goal

Tip: The cheapest hashrate is the hardware you already own. Buying a CPU just to mine Monero rarely pays back on mining income alone; see our profitability breakdown in the mining category. Many miners mine on a machine they need anyway, such as a home server.

Power efficiency matters more than peak H/s

Your income is proportional to hashrate, but your cost is proportional to watts. Desktop chips, especially recent Ryzen 9 and Intel K-series parts, ship with power limits set for peak benchmark scores rather than efficiency. Lowering the power limit (PPT on AMD, PL1/PL2 on Intel) or using AMD’s Eco Mode usually reduces power draw proportionally more than it reduces hashrate. How much depends on the chip and the silicon lottery, so measure it yourself: use a cheap plug-in power meter at the wall rather than software readouts, step the limit down, and stop when hashes per watt stop improving.

RandomX v2: what changes for your CPU

Monero’s next hard fork, which also brings FCMP++, is planned to switch the network to RandomX v2. XMRig 6.26.0 already supports it, and it runs on the FCMP++ beta stressnet. No mainnet date has been set.

According to benchmark notes published by P2Pool developer SChernykh, v2:

• Raises the program size from 256 to 384 instructions per hash.
• Adds extra AES mixing and prefetches the dataset further ahead.
• Does about 53% more work per hash than v1 overall.

The practical effect for miners:

So raw H/s numbers will usually fall after the fork, typically to 75–100% of today’s values, but each hash represents more computation. Because everyone’s hashrate shifts at the same time, your share of the network should change only slightly, depending on how your CPU fares relative to the average. The author also noted that later code tweaks improved v2 results, so treat these numbers as early.

Key point: After RandomX v2 activates, benchmark tables like the one above will need to be redone. Comparing a v1 number with a v2 number is meaningless.

GPUs, ASICs and botnets

• GPUs can run RandomX but are far less efficient per watt than CPUs. Very few Monero miners use them.
• ASICs for RandomX have not appeared in any meaningful form, which is what the algorithm was designed for.
• Botnets and cloud abuse are a real part of RandomX hashrate, because any computer can mine it. This is one reason Monero’s community encourages home miners to join decentralized options like P2Pool.

Bottom line

If you are buying, a 16-core Ryzen 9 gives the most hashrate on a normal desktop, and a used 5950X gives the best value. Fill both memory channels, enable huge pages and tune power limits. Before spending anything, check whether your electricity price makes mining worthwhile, and plan to update your miner for RandomX v2 when the fork date is announced.