LLM Roofline
NVIDIA Cloud · PCIE · Backend: vllm

Cloud NVIDIA L40S (48 GB) — LLM Sizing & Speed Matrix

48 GB cloud inference sweet spot for 32B Q8_0 or 70B IQ4_XS at sub-$1/hr.

Reference Pricing
~$0.89/hr Cloud
Usable Memory
47.4 GiB
Memory Bandwidth
864 GB/s (82% eff.)
FP16 Compute
362 TFLOPS
Roofline Ridge
337.2 FLOP/B
Power (Load / Idle)
350W / 0W

Buy or Rent Cloud NVIDIA L40S (48 GB)

Vendor & Cloud Links
Deploy on DigitalOcean GPU Droplets
DigitalOcean · On-Demand / Spot GPU
↗
Launch Pod on RunPod
RunPod · On-Demand / Spot GPU
↗
Deploy on Vast.ai Spot GPU
Vast.ai · On-Demand / Spot GPU
↗
Run Serverless / GPU on Novita AI
Novita AI · On-Demand / Spot GPU
↗

All 32 Open-Weight Models on Cloud NVIDIA L40S (48 GB) (Q4_K_M @ 32k)

Model Status Weights Total @ 32k Decode TTFT (4k) Max Safe Ctx
Llama-3.3 Nemotron Super 49B COMFORTABLE 28.7 GiB 36.7 GiB 18.9 tok/s 1.80s 88k
Mixtral 8x7B Instruct COMFORTABLE 26.9 GiB 32.2 GiB 50.8 tok/s 0.50s 32k
Qwen3 32B COMFORTABLE 18.9 GiB 28.3 GiB 24.5 tok/s 1.23s 40k
Qwen2.5 Coder 32B COMFORTABLE 18.9 GiB 28.3 GiB 24.5 tok/s 1.21s 32k
DeepSeek-R1 Distill Qwen 32B COMFORTABLE 18.9 GiB 28.3 GiB 24.5 tok/s 1.21s 108k
QwQ 32B Reasoning COMFORTABLE 18.9 GiB 28.3 GiB 24.5 tok/s 1.21s 40k
GLM-4 32B 0414 COMFORTABLE 18.6 GiB 22.1 GiB 32.2 tok/s 1.20s 32k
Qwen3 30B-A3B COMFORTABLE 17.6 GiB 21.6 GiB 118.5 tok/s 0.18s 40k
Gemma 3 27B IT COMFORTABLE 15.8 GiB 20.2 GiB 35.3 tok/s 1.01s 128k
Mistral Small 3.1 24B COMFORTABLE 13.8 GiB 20.3 GiB 35.1 tok/s 0.88s 128k
Codestral 22B COMFORTABLE 12.8 GiB 21.4 GiB 33.4 tok/s 0.84s 32k
Qwen3 14B COMFORTABLE 8.5 GiB 15.0 GiB 48.8 tok/s 0.57s 40k
Qwen2.5 14B Instruct COMFORTABLE 8.5 GiB 16.0 GiB 45.4 tok/s 0.58s 32k
DeepSeek-R1 Distill Qwen 14B COMFORTABLE 8.5 GiB 16.0 GiB 45.4 tok/s 0.58s 128k
Phi-4 14B COMFORTABLE 8.5 GiB 13.0 GiB 56.9 tok/s 0.57s 16k
Gemma 3 12B IT COMFORTABLE 7.0 GiB 10.6 GiB 70.7 tok/s 0.48s 128k
Qwen3 8B COMFORTABLE 4.7 GiB 10.5 GiB 71.5 tok/s 0.34s 40k
Llama 3.1 8B Instruct COMFORTABLE 4.6 GiB 9.9 GiB 76.5 tok/s 0.33s 128k
Qwen2.5 7B Instruct COMFORTABLE 4.4 GiB 7.4 GiB 107.7 tok/s 0.31s 32k
Gemma 3 4B IT COMFORTABLE 2.5 GiB 4.4 GiB 197.8 tok/s 0.19s 128k
DeepSeek-R1 671B OOM 386.4 GiB 390.3 GiB OOM — OOM
DeepSeek-V3 671B OOM 386.4 GiB 390.3 GiB OOM — OOM
Llama 3.1 405B Instruct OOM 233.8 GiB 252.5 GiB OOM — OOM
Llama 4 Maverick 400B-A17B OOM 231.5 GiB 239.0 GiB OOM — OOM
Qwen3 235B-A22B OOM 135.4 GiB 142.6 GiB OOM — OOM
Mixtral 8x22B Instruct OOM 81.2 GiB 89.8 GiB OOM — OOM
Command A 111B OOM 63.9 GiB 69.1 GiB OOM — OOM
Llama 4 Scout 109B-A17B OOM 62.8 GiB 70.2 GiB OOM — OOM
Qwen2.5 72B Instruct OOM 41.9 GiB 53.7 GiB OOM — 12k
DeepSeek-R1 Distill Llama 70B OOM 40.7 GiB 52.5 GiB OOM — 16k
Llama 3.3 70B Instruct OOM 40.7 GiB 52.5 GiB OOM — 16k
Llama 3.1 70B Instruct OOM 40.7 GiB 52.5 GiB OOM — 16k

Interactive Cloud NVIDIA L40S (48 GB) Simulator

49.9B Dense · GQA
864 GB/s · 362 TFLOPS
3. Weight Quantization 4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 88k
Memory Allocation Footprint 36.74 / 47.40 GiB usable
Weights: 28.74 GiB KV Cache: 6.13 GiB Overhead: 1.88 GiB Headroom: 10.66 GiB
Decode Speed
18.9
tok/s (batch=1)
Prefill Speed
2318
prompt tok/s
TTFT Latency
1.80s
for 4,096 tok
Active Params
49.9B
of 49.9B total
vLLM Production Serve Command 
vllm serve nvidia/Llama-3_3-Nemotron-Super-49B-v1 --max-model-len 32768 --quantization fp8
Context KV Cache Total VRAM Decode Status
4k 0.77 GiB 30.82 GiB 22.4 tok/s COMFORTABLE
8k 1.53 GiB 32.14 GiB 21.8 tok/s COMFORTABLE
16k 3.06 GiB 33.68 GiB 20.8 tok/s COMFORTABLE
32k 6.13 GiB 36.74 GiB 18.9 tok/s COMFORTABLE
64k 12.25 GiB 42.86 GiB 16.1 tok/s TIGHT
128k 24.50 GiB 55.11 GiB 12.4 tok/s OOM

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

USE SERVERLESS API
Own Hardware/mo
$0.00
226.1h active + power
Cloud GPU/mo
$231.42
@$0.89/hr spot/pod
Serverless API/mo
$13.50
API cheaper

Recommended Hardware & Cloud Deployment

Contextual to fit status

Hardware FAQ — Cloud NVIDIA L40S (48 GB)

How much usable VRAM / GPU memory does Cloud NVIDIA L40S (48 GB) actually provide for LLMs?

Cloud NVIDIA L40S (48 GB) provides 47.4 GiB usable VRAM after subtracting 0.6 GiB driver/display overhead.

What is the largest Q4_K_M model Cloud NVIDIA L40S (48 GB) can run at 32k context?

With Q4_K_M weights and FP16 KV cache at 32k context, the largest fitting model in our verified catalog on Cloud NVIDIA L40S (48 GB) is Llama-3.3 Nemotron Super 49B (49.9B total).