NIM GPU resource bundles¶

This page shows how standard bundle labels map to GPU node shape, aggregate FP16 TFLOPS, VRAM, and system RAM on each supported managed Kubernetes platform, with model-level coverage in NIM GPU bundle validation.

EKS (AWS):

バンドル	GPU Type	GPU Count	FP16 TFLOPS	VRAM (GB)	Total FP16 TFLOPS	Total VRAM (GB)	RAM (GB)
S	T4	1	65	16	65	16	10
M	A10	1	125	24	125	24	32
L	L40S	1	181	48	181	48	32
XL	A10	4	125	24	500	96	192
XXL	L40S	4	181	48	724	192	384
3XL	L40S	8	181	48	1448	384	1536
4XL	A100 SXM	8	312	80	2496	640	1152

AKS (Azure):

バンドル	GPU Type	GPU Count	FP16 TFLOPS	VRAM (GB)	Total FP16 TFLOPS	Total VRAM (GB)	RAM (GB)
M	A10	1	125	24	125	24	440
L	A10	2	125	24	250	48	880
XL	A100 PCIe	1	312	80	312	80	220
2XL	A100 PCIe	2	312	80	624	160	440
3XL	A100 PCIe	4	312	80	1248	320	900
4XL	A100 SXM	8	312	80	2496	640	1900

OCP (Oracle):

バンドル	GPU Type	GPU Count	FP16 TFLOPS	VRAM (GB)	Total FP16 TFLOPS	Total VRAM (GB)	RAM (GB)
XL	A100 SXM	1	312	80	312	80	80
2XL	A100 SXM	2	312	80	624	160	160
3XL	A100 SXM	4	312	80	1248	320	320
4XL	A100 SXM	8	312	80	2496	640	640

OKE H100 (Oracle):

バンドル	GPU Type	GPU Count	FP16 TFLOPS	VRAM (GB)	Total FP16 TFLOPS	Total VRAM (GB)	RAM (GB)
XL	H100 SXM	1	1000	80	1000	80	80
2XL	H100 SXM	2	1000	80	2000	160	160
3XL	H100 SXM	4	1000	80	4000	320	320
4XL	H100 SXM	8	1000	80	8000	640	640