MIG Profile Comparison

MIG profiles represent different partitions of a physical NVIDIA A100 80GB GPU. Each profile gives users a slice of compute and memory resources while maintaining full isolation from other workloads running on the same GPU.

On Wulver, the following MIG profiles are supported:

• 10gb – 10 GB memory
• 20gb – 20 GB memory
• 40gb – 40 GB memory

The table below summarizes the hardware characteristics of each MIG profile available on Wulver, alongside the full NVIDIA A100 80 GB GPU. It lists memory capacity, compute resources, and other architectural limits so users can quickly compare performance and capability across profiles.

Specification	10gb	20gb	40gb	Full 80GB	Notes
Device Name	A100 MIG 10gb	A100 MIG 20gb	A100 MIG 40gb	A100-SXM4-80GB	GPU Profile
SU Usage factor	2	4	8	16	Service units
Global Memory	10.2 GB	20.9 GB	42.4 GB	85.2 GB	Raw hardware memory
Usable Memory	~9.5 GB	~20 GB	~40 GB	~80 GB	Available for applications
Multiprocessors (SMs)	14	28	42	108	Parallel compute units
Relative Compute Power	1x	2x	3x	7.7x	Performance scaling
Total Parallel Threads	28672	57344	86016	221184	SMs × threads/SMP
Memory Bus Width	640 bits	1280 bits	2560 bits	5120 bits	Memory bandwidth
Memory Bandwidth limits	~1.3 TB/s	~2.6 TB/s	~5.1 TB/s	~10.2 TB/s	Theoretical peak Bandwidth (B/s)=Memory Clock (Hz)×Bus Width (bits)÷8
L2 Cache Size	5 MB	10 MB	20 MB	41 MB	Fast memory cache
Async Engines	1	2	3	5	Concurrent operations
Max Threads per Block	1024	1024	1024	1024	CUDA block limit
Max Threads per SMP	2048	2048	2048	2048	Per multiprocessor
Memory Clock	1593 MHz	1593 MHz	1593 MHz	1593 MHz	Memory frequency
Clock Rate	1410 MHz	1410 MHz	1410 MHz	1410 MHz	GPU core frequency
Shared Memory/Block	49 KB	49 KB	49 KB	49 KB	Per CUDA block
Registers per Block	65536	65536	65536	65536	Per CUDA block
Constant Memory	64 KB	64 KB	64 KB	64 KB	Read-only memory
Warp Size	32	32	32	32	SIMD execution width
ECC Support	Yes	Yes	Yes	Yes	Error correction
Unified Memory	Yes	Yes	Yes	Yes	CPU-GPU memory sharing
Concurrent Kernels	Yes	Yes	Yes	Yes	Multiple kernel execution
Max Grid Dimensions	2³¹-1 × 65535 × 65535	2³¹-1 × 65535 × 65535	2³¹-1 × 65535 × 65535	2³¹-1 × 65535 × 65535	CUDA grid limits