1000
150
97
Quick estimate
Start with just a model name. We fill the rest, then let you tune every assumption.
✓ In catalog
Popular models: Llama 3.1, Mistral, Qwen 2.5, Gemma 2 — type to autocomplete
Stored in this browser only — never sent to our servers.
Based on your configuration — ISL 1000, OSL 150, FP16 KV cache, 97 concurrent users.
GPUs required0× NVIDIA H200 SXMConfigure workload below to see results↻ see math
How we got 0
total memory = 20 GB
usable / GPU = 72 GB (90% of 80)
⌈20 ÷ 72⌉ = 1 GPU
peak 3× → range up to 2
↻ flip backusable / GPU = 72 GB (90% of 80)
⌈20 ÷ 72⌉ = 1 GPU
peak 3× → range up to 2
Weight memory 0GBNemotron-Mini-4B-Instruct↻ see math
Weight memory
params × bytes/param
8B × 2 (BF16)
= 16 GB
↻ flip back8B × 2 (BF16)
= 16 GB
KV cache / req 0MB1150 tokens/req · 97 users↻ see math
KV cache / request
2 × layers × kv_heads ×
head_dim × bytes × tokens
2×32×8×128×2 = 128 KB/tok
× 150 tokens = 19 MB
↻ flip backhead_dim × bytes × tokens
2×32×8×128×2 = 128 KB/tok
× 150 tokens = 19 MB
MONTHLY COST
CLOUD
$0.0K/mo
AWS · $0K over 5yr
SELF-HOSTED
$0.0K/mo
5yr amort · $0K total
Self-hosted saves $0.0K/mo
↻ see mathHow we calculated this
Cloud:
0 GPUs × $1.22/gpu-hr × 730 hrs
= $0.0K/mo
Self-hosted:
$0K ÷ 60 months
= $0.0K/mo
(hardware amortization only)
5-year totals:
Cloud: $0K
Hardware: $0K
↻ flip back0 GPUs × $1.22/gpu-hr × 730 hrs
= $0.0K/mo
Self-hosted:
$0K ÷ 60 months
= $0.0K/mo
(hardware amortization only)
5-year totals:
Cloud: $0K
Hardware: $0K
⚠️ Self-hosted excludes: power (~$X/mo), cooling, staff, networking. Typical full TCO adds 40–80% to this number.
Want to change assumptions?