LLM Roofline
32 Verified Models 28 Local & Cloud GPUs MoE · MLA · UMA Tuned Limits

Real-Physics LLM VRAM, Token/s & Buy-vs-Cloud Sizing

Stop estimating speed by dividing file size by memory bandwidth. Model exact MoE active expert bandwidth, MLA / GQA / sliding-window KV cache growth, compute-bound prefill TTFT, OS unified-memory limits, and 36-month hardware vs cloud TCO directly in your browser.

Client-Side Deterministic Engine

Interactive VRAM, Bandwidth Roofline & TCO Simulator

30.5B (3.3B active) · GQA
400 GB/s · 28.4 TFLOPS
3. Weight Quantization 4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 40k
Memory Allocation Footprint 21.40 / 49.92 GiB usable
Weights: 17.57 GiB KV Cache: 3.00 GiB Overhead: 0.83 GiB Headroom: 28.52 GiB
Decode Speed
50.8
tok/s (batch=1)
Prefill Speed
2006
prompt tok/s
TTFT Latency
2.08s
for 4,096 tok
Active Params
3.3B
of 30.5B total
llama.cpp Server Command 
llama-server -hf Qwen/Qwen3-30B-A3B:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080
Context KV Cache Total VRAM Decode Status
4k 0.38 GiB 18.63 GiB 109.5 tok/s COMFORTABLE
8k 0.75 GiB 19.15 GiB 94.0 tok/s COMFORTABLE
16k 1.50 GiB 19.90 GiB 73.3 tok/s COMFORTABLE
32k 3.00 GiB 21.40 GiB 50.8 tok/s COMFORTABLE

Buy-vs-Cloud-vs-API Monthly TCO Break-Even

USE SERVERLESS API
Own Hardware/mo
$66.40
88.9h active + power
Cloud GPU/mo
$90.97
@$0.89/hr spot/pod
Serverless API/mo
$13.50
Payback: 221.3 mo

Recommended Hardware & Cloud Deployment

Contextual to fit status

Popular Model Sizing Hubs

Verified against Hugging Face config.json and live OpenRouter pricing.

View all 32 models →

Qwen3 235B-A22B

GQA
235.1B total · 22B active
API: $0.14/$0.8 per 1M

Qwen3 32B

GQA
32.8B total · Dense
API: $0.08/$0.28 per 1M

Qwen3 30B-A3B

GQA
30.5B total · 3.3B active
API: $0.12/$0.5 per 1M

Qwen3 14B

GQA
14.8B total · Dense
API: $0.12/$0.24 per 1M

Qwen2.5 Coder 32B

GQA
32.8B total · Dense
API: $0.66/$1 per 1M

DeepSeek-R1 671B

MLA
671B total · 37B active
API: $0.5/$2.15 per 1M

DeepSeek-V3 671B

MLA
671B total · 37B active
API: $0.29/$1.14 per 1M

DeepSeek-R1 Distill Llama 70B

GQA
70.6B total · Dense
API: $0.2/$0.6 per 1M

DeepSeek-R1 Distill Qwen 32B

GQA
32.8B total · Dense
API: $0.1/$0.3 per 1M

Llama 4 Maverick 400B-A17B

GQA
402B total · 17B active
API: $0.35/$1 per 1M

Llama 4 Scout 109B-A17B

GQA
109B total · 17B active
API: $0.1/$0.3 per 1M

Llama 3.3 70B Instruct

GQA
70.6B total · Dense
API: $0.71/$0.71 per 1M

Hardware Roofline Profiles

Apple Silicon UMA, Strix Halo Mini-PCs, Desktop GPUs, and SXM Cloud GPUs.

View all 28 nodes →

Apple Mac Mini M4 Pro (48 GB)

273 GB/s 37.44 GiB safe
~$1,799 · 95W TDP

Apple MacBook Pro M3 Max (64 GB)

400 GB/s 49.92 GiB safe
~$3,499 · 115W TDP

Apple Mac Studio M4 Max (128 GB)

546 GB/s 99.84 GiB safe
~$3,499 · 140W TDP

Apple Mac Studio M3 Ultra (256 GB)

819 GB/s 199.68 GiB safe
~$5,599 · 230W TDP

Ryzen AI Max+ 395 Strix Halo (128 GB)

256 GB/s 96 GiB safe
~$1,899 · 140W TDP

Ryzen AI Max+ 395 Strix Halo (64 GB)

256 GB/s 48 GiB safe
~$1,499 · 130W TDP

NVIDIA RTX 5090 (32 GB)

1792 GB/s 31 GiB safe
~$2,499 · 575W TDP

NVIDIA RTX 5070 Ti (16 GB)

896 GB/s 15 GiB safe
~$849 · 300W TDP

NVIDIA RTX 4090 (24 GB)

1008 GB/s 23 GiB safe
~$1,999 · 450W TDP

2x NVIDIA RTX 3090 Rig (48 GB)

936 GB/s 46.65 GiB safe
~$1,950 · 680W TDP

Cloud NVIDIA H200 SXM (141 GB)

4800 GB/s 140.4 GiB safe
Cloud ~$3.49/hr

Cloud NVIDIA H100 SXM (80 GB)

3350 GB/s 79.4 GiB safe
Cloud ~$2.39/hr