G-Power
A new formula can be developed and utilized to objectively compare and evaluate the performance of GPU devices in absolute terms. In GPGPU (General-Purpose Computing on Graphics Processing Units), the performance unit of a GPU can be expressed as G-Power or GP. By distinguishing between performance metrics that are effective for machine learning and graphic rendering tasks, we can use two separate formulas to compare representative GPU devices like the RTX 4090 and A100.
Performance of the RTX 4090
FP16 Performance: 142 TeraFLOPS
FP32 Performance: 35.6 TeraFLOPS
Number of CUDA Cores: 16,384
Number of Tensor Cores: 1,216
Memory Bandwidth: 936.2 GB/s
VRAM: 24 GB
Power Consumption: 450 W
Performance of the A100
FP16 Performance: 312 TeraFLOPS
FP32 Performance: 19.5 TeraFLOPS
Number of CUDA Cores: 6,912
Number of Tensor Cores: 6,912
Memory Bandwidth: 1,555 GB/s
VRAM: 40 GB
Power Consumption: 400 W
Machine Learning Performance Comparison Formula
In machine learning, FP16 performance and the number of Tensor cores are crucial. Therefore, a performance comparison formula has been developed based on these factors. The formula includes all the important metrics for machine learning tasks:
Comparison Metrics
FP16 (Half Precision) FLOPS
Number of Tensor Cores
Memory Bandwidth (GB/s)
VRAM (GB)
Power Consumption (W)
Comparison Formula
Where α, β, γ, δ, ϵ are weights that can be adjusted based on the importance of each metric according to user preference. In this case, they are set as follows:
Weight Settings
α = 0.3 (FP16)
β = 0.3 (FP32)
γ = 0.2 (Tensor Core)
δ = 0.1 (Memory Bandwidth)
ϵ = 0.1 (VRAM)
ML Performance Index Calculation
Machine Learning Task Efficiency Results:
ML Performance Index of RTX 4090: 0.87
ML Performance Index of A100: 4.10
Graphic Rendering Performance Comparison Formula
To comparatively evaluate GPUs more specialized in graphic rendering, a formula was developed considering key performance metrics for game rendering. The key metrics are as follows:
FP32 Performance (TeraFLOPS): One of the most critical performance indicators in game rendering.
Number of CUDA Cores: Indicates parallel processing capability.
Memory Bandwidth (GB/s): Affects texture and data transfer speeds.
VRAM (GB): Important for handling large textures and high-resolution gaming.
Power Consumption (W): Necessary to consider efficiency.
The formula encompassing all crucial metrics in graphic rendering is as follows:
Where α, β, γ, δ represent weights indicating the importance of each performance metric. Since FP32 and CUDA core count are most important in game rendering, they are given higher weights:
Weight Settings
α = 0.4 (FP32)
β = 0.3 (CUDA Core)
γ = 0.2 (Memory Bandwidth)
δ = 0.1 (VRAM)
RTX 4090 Performance
FP32 Performance: 35.6 TeraFLOPS
Number of CUDA Cores: 16,384
Memory Bandwidth: 936.2 GB/s
VRAM: 24 GB
Power Consumption: 450 W
A100 Performance
FP32 Performance: 19.5 TeraFLOPS
Number of CUDA Cores: 6,912
Memory Bandwidth: 1,555 GB/s
VRAM: 40 GB
Power Consumption: 400 W
Game Rendering Performance Index Calculation
Game Rendering Performance Comparison Results:
Game Performance Index of RTX 4090: 11.38
Game Performance Index of A100: 5.99
These formulas are designed to objectively and quantitatively evaluate the performance of each GPU. By reflecting the key metrics optimized for both machine learning and graphic rendering tasks, these analyses can significantly aid in selecting the optimal GPU for various scenarios, ultimately helping to calculate G-Power.
Last updated
