A wide range of GPU options are available, allowing you to select the most optimal compute resource for your workload. If GPU resources are not requested, the workload will be executed on a CPU only node.
Vendor | Class | Generation | CUDA Cores | VRAM | Label |
NVIDIA | RTX A6000 | Ampere | 10,752 | 48 GB | RTX_A6000 |
NVIDIA | A100 NVLINK | Ampere | 6,912 | 40 GB | A100_NVLINK |
NVIDIA | Tesla V100 NVLINK | Volta | 5,120 | 16 GB | Tesla_V100_NVLINK |
NVIDIA | Tesla V100 | Volta | 5,120 | 16 GB | Tesla_V100 |
NVIDIA | RTX 6000 | Turing | 4,608 | 24 GB | Quadro_RTX_6000 |
NVIDIA | RTX 5000 | Turing | 3,072 | 16 GB | Quadro_RTX_5000 |
NVIDIA | RTX 4000 | Turing | 2,304 | 8 GB | Quadro_RTX_4000 |
NVIDIA | Tesla P100 | Pascal | 3,584 | 16 GB | Tesla_P100_NVLINK |
NVIDIA | Multi Purpose Pascal | Pascal | 2,000+ | 8 GB | NV_Pascal |
Each GPU includes a certain amount of host CPU and RAM, these are included at no additional fee. Allocating multiple GPUs to a single workload will increase the CPU and RAM allocation proportionally.
Class | vCPU | RAM | Great For |
RTX A6000 | 30 Epyc | 128 GB | Rendering, Neural Network training |
A100 NVLINK | 30 Epyc | 240 GB | Complex Deep Neural Network training, HPC |
Tesla V100 NVLINK | 4 Xeon Silver | 32 GB | Deep Neural Network training, HPC |
Tesla V100 | 3 | 20 GB | AI inference, Rendering, Batch processing, Hashcat |
RTX 6000 | 8 | 60 GB | Complex DNN Training, Rendering, Batch processing |
RTX 5000 | 8 | 60 GB | Machine learning, Rendering, Batch processing |
RTX 4000 | 3 | 16 GB | Machine learning, Rendering, Game streaming |
Tesla P100 NVLINK | 4 Xeon Silver | 32 GB | Entry level HPC, Rendering, Batch processing |
Multi Purpose Pascal | 1 | 8 GB | Transcoding, Rendering, Game streaming, Batch |
If a workload requests more peripheral compute resources (vCPU, RAM) than offered in a standard instance size, additional costs will incur.
Additional CPU and RAM is billed in increments of $0.07/hr for 1 vCPU + 8 GB RAM.
CPU Only nodes are available for tasks such as control-plane services, databases, ingresses and CPU rendering.
CPU Model | RAM per vCPU | Max CPU per Workload | Label |
Intel Xeon v1/v2 | 3 GB | 94 | xeon |
AMD Epyc Rome | 4 GB | 46 | epyc |
Workloads without GPU requests are always scheduled on CPU nodes. If a specific CPU model is not explicitly selected, the scheduler will automatically schedule workloads requesting few CPU cores on Epyc class CPUs, as these perform exceptionally well on single thread workloads.
A combination of resource requests and node affinity is used to select the type and amount of compute for your workload. CoreWeave Cloud relies only on these native Kubernetes methods for resource allocation, allowing maximum flexibilty.
spec:containers:- name: exampleresources:limits:cpu: 3memory: 16Ginvidia.com/gpu: 1affinity:nodeAffinity:requiredDuringSchedulingIgnoredDuringExecution:nodeSelectorTerms:- matchExpressions:- key: gpu.nvidia.com/classoperator: Invalues:- Tesla_V100
spec:containers:- name: exampleresources:limits:cpu: 15memory: 128Ginvidia.com/gpu: 4affinity:nodeAffinity:requiredDuringSchedulingIgnoredDuringExecution:nodeSelectorTerms:- matchExpressions:- key: gpu.nvidia.com/classoperator: Invalues:- Tesla_V100_NVLINK
spec:containers:- name: exampleresources:limits:cpu: 2memory: 16Ginvidia.com/gpu: 2affinity:nodeAffinity:requiredDuringSchedulingIgnoredDuringExecution:nodeSelectorTerms:- matchExpressions:- key: gpu.nvidia.com/classoperator: Invalues:- NV_Pascal
spec:containers:- name: exampleresources:limits:cpu: 16memory: 48Giaffinity:nodeAffinity:requiredDuringSchedulingIgnoredDuringExecution:nodeSelectorTerms:- matchExpressions:- key: cpu.coreweave.cloud/familyoperator: Invalues:- xeon
spec:containers:- name: exampleresources:limits:cpu: 1memory: 4Giaffinity:nodeAffinity:requiredDuringSchedulingIgnoredDuringExecution:nodeSelectorTerms:- matchExpressions:- key: cpu.coreweave.cloud/familyoperator: Invalues:- epyc
Kubernetes allows resources to be scheduled with requests
and limits.
When only limits
are specified, the requests
are set to the same amount as the limit.