Compare 2–4 GPUs side by side
Specs, bars, radar and workload tabs, and how to share a comparison.
On this page · 6 sections
Compare GPUs puts two to four cards next to each other — from the spec sheet down to how each one serves the same model. It answers the question a spec table alone cannot: for this workload, which card is faster, which is cheaper per token and which needs fewer GPUs.
Choosing the cards
Start from the GPU catalog — tick up to four rows and press Compare — or add cards on the compare page itself with Add a GPU, which searches the catalog. Each card gets a fixed colour that it keeps across all four views; remove a card with its ✕. The selection lives in the address, /gpus/compare?gpus=h100-sxm5-80gb,h200-sxm-141gb, so Copy link shares exactly this comparison. The view parameter opens a specific tab.
Spec table
Every catalog field in four groups — Memory, Compute, Connectivity, Metadata — with one column per GPU. The leader of each numeric row is marked best (higher is better, except TDP, where lower wins; ties are not marked), and a Spread column shows how far apart the cards are. Only rows that differ hides the rows where all cards match, which for two cards of one family leaves the few that matter. On a phone the label column stays fixed while the GPU columns swipe.
Normalized bars
Each metric as a bar per GPU, scaled to the leader: VRAM and memory bandwidth; BF16, FP8 and FP4 TFLOPS and BF16 TFLOPS per watt; NVLink bandwidth, the largest NVLink domain and intra-node bandwidth. Read it for proportions — a card at 50 % of the leader's bandwidth decodes at roughly half its speed on a bandwidth-bound model, all else equal. A missing bar means the hardware has no such path (no FP4, no NVLink).
Radar
The same comparison as a shape: six axes — VRAM, memory bandwidth, BF16, FP8, NVLink bandwidth and BF16 TFLOPS per watt — each as a percentage of the best card. A balanced shape is a general-purpose card; a spike shows what a card is built for. It is the quickest view for slides and the least precise one; use the bars or the table for numbers.
Per workload
The tab that turns specs into sizing. Pick a catalog model, a precision, a context and a user count; the page runs the Inference calculator on exactly the compared cards and shows, per GPU, how many cards are needed, memory use with its fit band, tok/s per user with TPOT, TTFT, total throughput, $/hr and $/1M tokens. The runtime is fixed at vLLM with an FP16 KV cache and a 1K → 256 request, so the cards are compared on equal terms; open the calculator for other runtimes or traffic. Notes flag a precision that is not native on one of the cards and the per-request meaning of TTFT.
The workload is part of the link too — model, dtype, ctx and users — so "H100 versus H200 for Llama 3.1 70B at 32K and 16 users" is one URL. Prices are current market medians; a card without a market price shows a catalog estimate or no cost.
Reading a comparison
- Compare $/1M tokens before $/hr: a dearer card that needs fewer GPUs or runs faster is often cheaper per token.
- A card that needs twice the GPUs also needs tensor parallelism; check its interconnect before you count it as equal.
- Spec leads do not always become speed leads: decode follows bandwidth, prefill follows compute, and precision support decides which TFLOPS figure applies.