Loading the fit, utilization assumptions, and cited price data.
Fit is always evaluated when model sizing is available. GPU rent, paid idle capacity, API crossover, and measured optimization appear only when their required evidence exists. When the API wins, the page says so.
Model
meta-llama/Llama-3.3-70B-Instruct
GPU
NVIDIA H100 SXM 80 GB
Requests/day
50,000
GPU count
1
Quantization
FP16
Tokens/request
500 in, 500 out
Does not fit
The model and KV cache need more VRAM than the selected GPU count provides. Cost is intentionally hidden for a configuration that cannot run.
Model
meta-llama/Llama-3.3-70B-Instruct
Hardware
1 x NVIDIA H100 SXM 80 GB, FP16
Weights and KV cache stay separate so the fit is reproducible.
Hardware recommendation
Each row uses an engine GPU fit when present; the remaining catalog GPUs apply this result's required memory to published VRAM. GPU count rises only enough to clear that memory requirement.
Engine fit suggestion
Memory fit: 7x NVIDIA A10 24 GB at 90% vLLM GPU memory utilization.
On-demand rates are preferred for each GPU; when none is captured, the lowest dated captured tier is used. The complete rentable GPU catalog is ranked only by monthly rent over the contract's 730.5-hour month, rounded to cents. Equal displayed prices share a rank. A non-fitting result has no engine cost or replica count, so each amount is a one-deployment rental floor; alternative throughput is not re-estimated, and memory fit does not establish tensor-parallel topology support.
145.2 GB required, 151.2 GB available inside the engine memory budget.
Current result memory total, Ampere, 768 GB/s memory bandwidth
Monthly rental floor
$818.16
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 10% utilization scenario, the provider still bills the full monthly amount.
Open rental source145.2 GB required, 172.8 GB available inside the engine memory budget.
Current result memory total, Ampere, 768 GB/s memory bandwidth
Monthly rental floor
$964.26
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 10% utilization scenario, the provider still bills the full monthly amount.
Open rental source145.2 GB required, 172.8 GB available inside the engine memory budget.
Current result memory total, Ampere, 696 GB/s memory bandwidth
Monthly rental floor
$1,022.70
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 10% utilization scenario, the provider still bills the full monthly amount.
Open rental sourceOnly result-triggered recommendations appear here.
FP16 weights use 141.1 GB and bring total memory to 145.2 GB; INT4, NVFP4, or MXFP4 would lower weights to 35.3 GB and total memory to 39.4 GB, which fits inside 72 GB.
The current vLLM command uses tensor parallel size 1 and provides 72 GB, while 145.2 GB requires 3 same-GPU workers for 216 GB; raise --tensor-parallel-size to 3 if the model topology supports it.
A physics bound, not an expected throughput result.
24 output tok/sCeiling
CEILING at 100% unit utilisation, not expected throughput. Decode is bounded by memory bandwidth and prefill by dense compute. Ideal tensor-parallel scaling is assumed for both. Real decode and prefill throughput are lower.
Decode and prefill ceilings derived from GPU memory bandwidth, dense compute, and model parameters, as of 2026-08-08
The exact command attached to this result.
vllm serve 'meta-llama/Llama-3.3-70B-Instruct' --dtype float16 --tensor-parallel-size 1 --max-model-len 1000 --gpu-memory-utilization 0.9Every published number keeps its source and date.
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
RunInfra Engine parameter-only KV-cache heuristic
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
Does not fit
The model and KV cache need more VRAM than the selected GPU count provides. Cost is intentionally hidden for a configuration that cannot run.
Model
meta-llama/Llama-3.3-70B-Instruct
Hardware
1 x NVIDIA H100 SXM 80 GB, FP16
Weights and KV cache stay separate so the fit is reproducible.
Hardware recommendation
Each row uses an engine GPU fit when present; the remaining catalog GPUs apply this result's required memory to published VRAM. GPU count rises only enough to clear that memory requirement.
Engine fit suggestion
Memory fit: 7x NVIDIA A10 24 GB at 90% vLLM GPU memory utilization.
On-demand rates are preferred for each GPU; when none is captured, the lowest dated captured tier is used. The complete rentable GPU catalog is ranked only by monthly rent over the contract's 730.5-hour month, rounded to cents. Equal displayed prices share a rank. A non-fitting result has no engine cost or replica count, so each amount is a one-deployment rental floor; alternative throughput is not re-estimated, and memory fit does not establish tensor-parallel topology support.
145.2 GB required, 151.2 GB available inside the engine memory budget.
Current result memory total, Ampere, 768 GB/s memory bandwidth
Monthly rental floor
$818.16
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 30% utilization scenario, the provider still bills the full monthly amount.
Open rental source145.2 GB required, 172.8 GB available inside the engine memory budget.
Current result memory total, Ampere, 768 GB/s memory bandwidth
Monthly rental floor
$964.26
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 30% utilization scenario, the provider still bills the full monthly amount.
Open rental source145.2 GB required, 172.8 GB available inside the engine memory budget.
Current result memory total, Ampere, 696 GB/s memory bandwidth
Monthly rental floor
$1,022.70
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 30% utilization scenario, the provider still bills the full monthly amount.
Open rental sourceOnly result-triggered recommendations appear here.
FP16 weights use 141.1 GB and bring total memory to 145.2 GB; INT4, NVFP4, or MXFP4 would lower weights to 35.3 GB and total memory to 39.4 GB, which fits inside 72 GB.
The current vLLM command uses tensor parallel size 1 and provides 72 GB, while 145.2 GB requires 3 same-GPU workers for 216 GB; raise --tensor-parallel-size to 3 if the model topology supports it.
A physics bound, not an expected throughput result.
24 output tok/sCeiling
CEILING at 100% unit utilisation, not expected throughput. Decode is bounded by memory bandwidth and prefill by dense compute. Ideal tensor-parallel scaling is assumed for both. Real decode and prefill throughput are lower.
Decode and prefill ceilings derived from GPU memory bandwidth, dense compute, and model parameters, as of 2026-08-08
The exact command attached to this result.
vllm serve 'meta-llama/Llama-3.3-70B-Instruct' --dtype float16 --tensor-parallel-size 1 --max-model-len 1000 --gpu-memory-utilization 0.9Every published number keeps its source and date.
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
RunInfra Engine parameter-only KV-cache heuristic
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
Does not fit
The model and KV cache need more VRAM than the selected GPU count provides. Cost is intentionally hidden for a configuration that cannot run.
Model
meta-llama/Llama-3.3-70B-Instruct
Hardware
1 x NVIDIA H100 SXM 80 GB, FP16
Weights and KV cache stay separate so the fit is reproducible.
Hardware recommendation
Each row uses an engine GPU fit when present; the remaining catalog GPUs apply this result's required memory to published VRAM. GPU count rises only enough to clear that memory requirement.
Engine fit suggestion
Memory fit: 7x NVIDIA A10 24 GB at 90% vLLM GPU memory utilization.
On-demand rates are preferred for each GPU; when none is captured, the lowest dated captured tier is used. The complete rentable GPU catalog is ranked only by monthly rent over the contract's 730.5-hour month, rounded to cents. Equal displayed prices share a rank. A non-fitting result has no engine cost or replica count, so each amount is a one-deployment rental floor; alternative throughput is not re-estimated, and memory fit does not establish tensor-parallel topology support.
145.2 GB required, 151.2 GB available inside the engine memory budget.
Current result memory total, Ampere, 768 GB/s memory bandwidth
Monthly rental floor
$818.16
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 60% utilization scenario, the provider still bills the full monthly amount.
Open rental source145.2 GB required, 172.8 GB available inside the engine memory budget.
Current result memory total, Ampere, 768 GB/s memory bandwidth
Monthly rental floor
$964.26
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 60% utilization scenario, the provider still bills the full monthly amount.
Open rental source145.2 GB required, 172.8 GB available inside the engine memory budget.
Current result memory total, Ampere, 696 GB/s memory bandwidth
Monthly rental floor
$1,022.70
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 60% utilization scenario, the provider still bills the full monthly amount.
Open rental sourceOnly result-triggered recommendations appear here.
FP16 weights use 141.1 GB and bring total memory to 145.2 GB; INT4, NVFP4, or MXFP4 would lower weights to 35.3 GB and total memory to 39.4 GB, which fits inside 72 GB.
The current vLLM command uses tensor parallel size 1 and provides 72 GB, while 145.2 GB requires 3 same-GPU workers for 216 GB; raise --tensor-parallel-size to 3 if the model topology supports it.
A physics bound, not an expected throughput result.
24 output tok/sCeiling
CEILING at 100% unit utilisation, not expected throughput. Decode is bounded by memory bandwidth and prefill by dense compute. Ideal tensor-parallel scaling is assumed for both. Real decode and prefill throughput are lower.
Decode and prefill ceilings derived from GPU memory bandwidth, dense compute, and model parameters, as of 2026-08-08
The exact command attached to this result.
vllm serve 'meta-llama/Llama-3.3-70B-Instruct' --dtype float16 --tensor-parallel-size 1 --max-model-len 1000 --gpu-memory-utilization 0.9Every published number keeps its source and date.
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
RunInfra Engine parameter-only KV-cache heuristic
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
Does not fit
The model and KV cache need more VRAM than the selected GPU count provides. Cost is intentionally hidden for a configuration that cannot run.
Model
meta-llama/Llama-3.3-70B-Instruct
Hardware
1 x NVIDIA H100 SXM 80 GB, FP16
Weights and KV cache stay separate so the fit is reproducible.
Hardware recommendation
Each row uses an engine GPU fit when present; the remaining catalog GPUs apply this result's required memory to published VRAM. GPU count rises only enough to clear that memory requirement.
Engine fit suggestion
Memory fit: 7x NVIDIA A10 24 GB at 90% vLLM GPU memory utilization.
On-demand rates are preferred for each GPU; when none is captured, the lowest dated captured tier is used. The complete rentable GPU catalog is ranked only by monthly rent over the contract's 730.5-hour month, rounded to cents. Equal displayed prices share a rank. A non-fitting result has no engine cost or replica count, so each amount is a one-deployment rental floor; alternative throughput is not re-estimated, and memory fit does not establish tensor-parallel topology support.
145.2 GB required, 151.2 GB available inside the engine memory budget.
Current result memory total, Ampere, 768 GB/s memory bandwidth
Monthly rental floor
$818.16
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 90% utilization scenario, the provider still bills the full monthly amount.
Open rental source145.2 GB required, 172.8 GB available inside the engine memory budget.
Current result memory total, Ampere, 768 GB/s memory bandwidth
Monthly rental floor
$964.26
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 90% utilization scenario, the provider still bills the full monthly amount.
Open rental source145.2 GB required, 172.8 GB available inside the engine memory budget.
Current result memory total, Ampere, 696 GB/s memory bandwidth
Monthly rental floor
$1,022.70
RunPod, Community Cloud, on demand
Captured 2026-08-08
At the selected 90% utilization scenario, the provider still bills the full monthly amount.
Open rental sourceOnly result-triggered recommendations appear here.
FP16 weights use 141.1 GB and bring total memory to 145.2 GB; INT4, NVFP4, or MXFP4 would lower weights to 35.3 GB and total memory to 39.4 GB, which fits inside 72 GB.
The current vLLM command uses tensor parallel size 1 and provides 72 GB, while 145.2 GB requires 3 same-GPU workers for 216 GB; raise --tensor-parallel-size to 3 if the model topology supports it.
A physics bound, not an expected throughput result.
24 output tok/sCeiling
CEILING at 100% unit utilisation, not expected throughput. Decode is bounded by memory bandwidth and prefill by dense compute. Ideal tensor-parallel scaling is assumed for both. Real decode and prefill throughput are lower.
Decode and prefill ceilings derived from GPU memory bandwidth, dense compute, and model parameters, as of 2026-08-08
The exact command attached to this result.
vllm serve 'meta-llama/Llama-3.3-70B-Instruct' --dtype float16 --tensor-parallel-size 1 --max-model-len 1000 --gpu-memory-utilization 0.9Every published number keeps its source and date.
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
RunInfra Engine parameter-only KV-cache heuristic
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
As of 2026-08-08
Embed code appears only when the calculator can publish a sourced cost for a fitting configuration.
Badge unavailable for this configuration
This configuration does not fit the selected hardware, so the calculator does not publish a cost or embed code.
RunInfraby RightNow© 2026 RunInfra. All rights reserved.
