GPU compare

Three GPUs.
One honest table.

By Bitpute · Updated 24 July 2026 · How we estimate · Sources · Editorial policy · Version history · Report an error

Line up any cards — RTX 5090 vs H100 vs A100 — on the specs that decide it, plus what actually matters: does it hold your model, how fast does it decode, and when does owning beat cloud. Winners highlighted, no marketing.

Fit-aware

Speed only counts if the model fits the VRAM.

Real decode

tok/s from memory bandwidth, per your model.

Buy vs cloud

Break-even months, not just sticker price.

Pick the cardslive

GPUs to compare

Set a column to “— none —” to compare just two.

Reference model — for fit & decode

All figures below use this model at Q4_K_M.

What matters most?

—

VRAM (GB)

Bandwidth (GB/s)

Decode (tok/s)

Decode tok/s is the memory-bandwidth ceiling on the model's active params at Q4_K_M — a clean upper bound for a single card, not a benchmarked figure. Break-even assumes 8 h/day vs the card's cloud rate.

How we compare

Fit first, then speed, then money.

Fit gates everything

A card that can't hold your model at Q4 gets no decode score — because 0 tok/s is the honest answer. VRAM is the first filter, always.

Speed from bandwidth

Decode is memory-bound: tok/s ≈ bandwidth ÷ active-weight bytes. That's why an A100 with more bandwidth beats a cheaper card even at the same VRAM.

Value is per-dollar

We show tok/s per $1k and per 100 W, plus buy-vs-cloud break-even — so a $25k H100 and a $2.3k 5090 get judged on what they return, not what they cost.

Run the full ownership cost →

Got a winner?