Why Hashrate and TFLOPS Are the Wrong Metrics for Mining and AI Hardware

The two headline specs everyone reads are gross theoretical ceilings. Neither predicts what a machine earns or how much real work it does. Here is the math, with live numbers, and the two figures of merit that replace them.
Two numbers dominate hardware buying decisions and both are close to useless on their own. For a Bitcoin miner it is hashrate, in terahash per second. For a GPU it is FP16 throughput, in teraflops. Each is a gross, theoretical ceiling, and each hides the variable that decides the outcome. Optimize for either headline and you will routinely buy the wrong machine.

Mining: profit is an identity, and hashrate is one term in it
Daily mining profit is not a mystery. It is a simple identity:

profit_per_day = income_per_day - power_cost_per_day
income_per_day = hashrate x network_revenue_per_hash x coin_price
power_cost_per_day = (watts / 1000) x 24 x electricity_rate

Hashrate appears once, inside income. Power is a separate term, subtracted. The figure of merit that ties them together is efficiency:

efficiency (J/TH) = watts / hashrate

A worked example from a live catalog, at an electricity rate of 0.07 USD/kWh:

Machine Hashrate Power Efficiency Daily profit
Whatsminer M79S 1350 TH/s 20000 W 14.81 J/TH ~ +$15.1
Antminer S23 Hyd 3U 1160 TH/s 11020 W 9.50 J/TH ~ +$23.3

The M79S has about 16 percent more hashrate and earns about 35 percent less profit. The reason is entirely in the efficiency column: 20,000 watts against 11,020 for a smaller hashrate lead. Rank the catalog by hashrate and the M79S is near the top; rank it by daily profit and it drops behind machines producing far less hash. Efficiency, not hashrate, tracks the outcome.
It sharpens across generations:

Machine Hashrate Power Efficiency Daily profit
Antminer S23 318 TH/s 3498 W 11.00 J/TH ~ +$5.6
Antminer S19 XP Hyd 512 TH/s 10600 W 20.70 J/TH ~ +$0.7

The 318 TH/s machine out-earns the 512 TH/s machine by roughly eight to one, because it does more work per joule. Push back one more generation to S19-class hardware near 21 to 22 J/TH and, at the same rate, daily profit goes negative while hashrate still reads in the hundreds of terahash.

There is a second term that swings the sign of the whole equation harder than any hardware choice: electricity_rate. Hold the machine fixed and vary only the rate, and profit crosses from negative to positive somewhere between a residential tariff (0.13 to 0.22 USD/kWh across much of the US) and an industrial one (0.04 to 0.08 USD/kWh). The same box loses money at home and earns at a hosting facility. Any honest ROI model treats the rate as a first-class input, not an afterthought, which is why a live profit calculator that lets you set your own rate is the only kind worth using. The mining numbers above come from one such live tool that ranks 285 machines this way at whatever rate you enter.

AI: TFLOPS is a peak, real inference is a measurement
FP16 teraflops is the theoretical peak floating-point rate of the chip. Three things make it a poor predictor of real performance:

1. Vendors frequently quote FP16 with 2:4 structured sparsity, which doubles the printed number relative to the dense math most workloads run.
2. Inference is usually memory-bandwidth-bound, not compute-bound. The card spends its time moving weights and activations, so peak FLOP throughput is not the binding constraint.
3. The software stack, kernels, drivers, and framework support, determines how much of the theoretical hardware you can reach at all.

The fix is to measure, not to read the datasheet. Run real workloads, language-model token generation, diffusion image generation, and vision, then normalize every card to a common baseline. With the RTX 3090 fixed at an index of 100:

GPU FP16 TFLOPS Inference index (RTX 3090 = 100)
RTX 3090 35.6 100
RTX 4090 165.2 133
RTX 5090 419.1 207
A100 40GB 312.0 152

The RTX 4090 has about 4.6x the paper teraflops of the 3090 and delivers about 1.33x the measured inference. The 5090 has nearly 12x the paper number and delivers about 2x. The theoretical ratio and the measured ratio are different quantities, and only the measured one shows up in tokens per second.

Training is a separate axis again. The A100 posts strong training throughput, roughly 1,396 images per second on a standard benchmark against the 3090's 905, yet sits around 152 on the inference index. A card's rank on one workload does not carry to another, so a single scalar, teraflops or hashrate, cannot summarize hardware worth.

This is exactly what a workload-indexed benchmark is for. MillionMiner's GPU and AI benchmark comparison indexes 78 cards on measured LLM, image, and vision performance against the RTX 3090 at 100, with VRAM and training throughput alongside and a head-to-head that reports the measured gap rather than the datasheet one.