What Generation CPU Is Your Cloud Server? Generation Identification and Premium

Use CPUID, cache, and benchmarks to quantify old CPU premiums.

Updated 2026-09-18 · CloudWorth

CPU generationold CPUcloud server premiumbenchmark evidenceCloudWorthSteal TimeVPS oversellingVPS benchmark

Generational premium rate = single-thread score per dollar comparison; if below baseline, file a ticket to seek redress.

Three-Step Generation Forensics

To determine the hardware generation of a cloud server, don't trust that 'High-Performance Computing' marketing name in the console. Vendors can change the wording at any time, but they can't change CPUID. The position of this article is simple: generation can be proven with evidence, and the evidence can be directly turned into a number—how many single-thread points you get per dollar. If it's below the baseline for the same price, you have reason to open a ticket, or even request a downgrade or refund.

I've compressed the whole process into three steps, and the next five sections expand on each one:

1. Read the fingerprint—/proc/cpuinfo to get family/model/stepping, lscpu to check cache and frequency ranges, and pin down which microarchitecture and which die it is.
2. Check the tables—use the fingerprint to check the official SKU tables, including cloud-vendor custom models (such as frequency-locked variants like 7B12 and 7R32), and confirm the rated base clock/turbo clock/L3 cache.
3. Convert to price—measured single-thread score ÷ actual monthly fee, to get 'single-thread points per dollar'; compare it with the same-price baseline, and calculate the generation premium rate.

All three steps can be recalculated. What you get is not an impression, but a set of screenshots you can paste into a ticket. If you want a comparable pricing standard first, you can look at the configuration-to-price converter in CloudWorth /app, lock down the actual amount paid, and then discuss performance.

Before doing forensics, remember one rule: all commands must be run on the target machine, not locally. The generation differences between cloud servers only show up on that virtual machine.

CPUID and Cache Fingerprinting

The first step is a combination of lscpu and /proc/cpuinfo. These two commands reveal far more than most tutorials cover.

lscpu | grep -Ei 'model name|family|model|stepping|MHz|cache'
grep -m1 -E 'cpu family|model\s|stepping' /proc/cpuinfo

Focus on three fields. family + model + stepping uniquely identify a silicon die, and this is the strongest basis for determining the CPU generation—marketing names can be arbitrary, but these three numbers cannot. For example, family 6 model 85 is a Xeon from the Skylake-SP line, family 25 model 1 is an EPYC 7002 (Rome), and family 25 model 17 points to an EPYC 9004 (Genoa).

stepping is especially worth noting. Manufacturers release multiple steppings under the same model, and when cloud providers recycle old CPUs and put them back on the market, they often still have the old stepping. If your instance is advertised as “latest generation” but the stepping is the version from three years ago, that is the first red flag.

Cache is the second fingerprint. L3 size directly exposes positioning: in the E5 v3 era, single-socket L3 was commonly 20–30MB, EPYC 7002 is 128MB, and Genoa is 256MB or even 384MB (3D V-Cache variants are higher). If a “high-frequency” instance reports only 16MB L3, it is most likely consumer-grade or an old Xeon E5 series.

Frequency fingerprint is the third layer. The CPU MHz in lscpu is the current instantaneous value. To see the turbo ceiling, use lscpu | grep -i mhz together with sustained stress testing. Old CPUs lose turbo quickly, and under long loads their frequency drops more noticeably than that of newer generations.

Combining these three fingerprints, you can independently confirm which generation this machine belongs to without any cooperation from the vendor.

SKU Baseline Table Cross-Check

After you get the family/model/stepping, the next step is to cross-reference the table. Don't just look at Intel ARK or AMD's official site—cloud providers' custom SKUs are where the deepest traps are.

A custom SKU is a locked-frequency variant that a cloud provider buys from Intel/AMD. The model often has suffix digits, such as 7B12, 7R32, 8B13, etc. They often cannot be found in public databases. Their base frequency is lowered in exchange for better performance-per-watt and a lower purchase price, but they are still marketed externally under family names like “EPYC” and “Xeon Platinum.” You see the word “Platinum” and assume it's the latest generation, when it may actually be a locked-frequency version of the previous generation.

When cross-referencing, verify at least four things:

• Microarchitecture codename (Skylake, Rome, Ice Lake, Genoa…)
• Nominal base clock / boost clock
• L3 cache capacity
• Memory channels and supported generation (DDR4 or DDR5)

If any of the four does not match the “latest generation” description, write it down. For example, if an instance advertised as “new-generation high-clock” only supports DDR4 and has 32 MB of L3, then it is most likely Skylake-SP or Cascade Lake, not a product after Ice Lake.

By the way, let's address a common question: how to tell if a cloud provider recycles old CPUs. The criterion for recycling is not “old,” but “old and still sold at new prices.” Take the same EPYC 7002: if the monthly fee is only 60% of Genoa's, that's reasonable inventory clearance; if the price is the same or even higher, then it constitutes a generational premium. The purpose of cross-referencing is to translate “old” into “unreasonably expensive.”

Generation Premium Rate Formula

Once you've completed forensics and the comparisons, you can do the math. This is the core of the article, and the step that turns qualitative complaints into quantitative evidence.

Single-thread benchmark score = median obtained on the target machine using the same tool (sysbench 1 thread, Geekbench 6 single-core, or UnixBench single-core)

Single-thread score per yuan = single-thread benchmark score ÷ actual monthly fee paid

Generation premium rate = (same-price benchmark score per yuan − your score per yuan) ÷ same-price benchmark score per yuan × 100%

Where does the benchmark come from? Take currently available instances at the same price and with the same number of vCPUs, run the same set of commands, and take the median. Don't use the impressive scores from the vendor's official website as your benchmark; use your own measured values.

Here's a reproducible example: an instance costs 120 yuan/month and scores 420 on sysbench single-thread, which is 3.5 points per yuan. A new-generation instance at the same price scores 780 single-thread, which is 6.5 points per yuan. Premium rate = (6.5 − 3.5) ÷ 6.5 ≈ 46%. In other words, you're paying nearly 50% more for every point of performance.

Besides single-thread, it's best to also calculate a memory bandwidth score per yuan, measured with sysbench memory or STREAM. Older platforms are often DDR4 quad-channel, while newer platforms are DDR5 twelve-channel; the bandwidth gap is sometimes larger than the compute gap, and this is the first place cloud servers bottleneck when running databases.

Threshold recommendation: if the premium rate exceeds 30% and persists for more than three months, it's worth opening a ticket. Below 15% is in the normal inventory-clearing range, so don't bother. Write these two numbers into your record table, and your judgment will have consistency.

Benchmark Scores Don't Match the Advertised Specs

This section addresses the most common dispute: measured benchmark scores don't match the advertised configuration. The vendor says “from 2.5GHz,” but you measure sustained only 1.8GHz; it says “high clock speed,” but single-thread performance is worse than a neighboring instance at half the price.

First rule out problems on your side. During stress testing, do you have exclusive vCPU? Are neighbors stealing resources? It's recommended to run it like this:

sysbench cpu --threads=1 --time=60 run
sysbench cpu --threads=1 --time=60 run   # at least twice, take the median

Before running, confirm load is 0; after running, check whether the frequency in lscpu is capped. If two consecutive results are consistent and significantly below the level corresponding to the advertised specs, the problem is basically on the host side.

Mismatches commonly fall into three categories:

• Locked-frequency SKU: advertised turbo 3.5GHz, but actually locked at 2.2GHz, so benchmark scores naturally don't match.
• Oversubscription: physical cores are sliced too many ways, single-thread scores fluctuate wildly with neighbor load; a gap of more than 15% between two benchmark runs is a signal.
• Generation mismatch: the console says “Genoa,” but it's actually Rome; the single-thread score gap is over 30%.

When gathering evidence, be sure to keep three things: screenshots of two benchmark runs, the lscpu output, and the instance creation time. Miss one, and the ticket is easily dismissed with “performance fluctuations are normal.”

This also intersects with “Real benchmark scores vs advertised configuration”: the advertised specs are numbers on a procurement list, while benchmark scores are the actual delivery after you paid for it. If the two don't match, it's essentially delivery shrinkage, not you “asking too much.”

Renewal Price Hikes and Support Ticket Scripts

The final section is about money. The most classic cloud server trick is low first-year pricing, doubled renewal, but the same CPU generation. When you renew, the price goes up 60%, while the family/model/stepping in lscpu is exactly the same as last year—this is "old-customer lock-in": a price increase with no corresponding hardware upgrade.

Before renewing, make a four-step checklist:

1. Record the current family/model/stepping and save a screenshot.
2. Record the current single-thread score and the single-thread score per yuan.
3. Request or look up the post-renewal price, and calculate the new score per yuan.
4. If the new score per yuan is below the benchmark for on-sale instances at the same price, it constitutes a price hike with no upgrade.

The key to ticket wording is to stick to numbers, not emotions, and to state your request clearly. You can write:

Instance X was created in 2024-03. Its current CPUID is family 25 model 1 stepping 2. The sysbench single-thread median is 420 points. The actual monthly fee is 120 yuan, equivalent to 3.5 points per yuan. Your on-sale instances at the same price measure 6.5 points per yuan, giving a generation premium of about 46%. The renewal quote will rise to 190 yuan, with no hardware changes. Please provide: 1. the official SKU model and generation corresponding to this instance; 2. a plan to adjust the price according to actual delivered performance, or migration to a new-generation instance at the same price.

Giving two options in your request makes it easier for the other party to give an actionable response. If the ticket is rejected, organize the same numbers into a written record as evidence for a follow-up complaint. The generation premium is not a feeling; it is a percentage you can recalculate—and that is exactly why it is far more powerful than "I think it's slow."