Building an Efficient Password Recovery Workstation: NVIDIA RTX Passwords-per-Watt Benchmarks

 

This article opens the series of publications aimed to help experts specify and build effective and power-efficient workstations for brute-forcing passwords. Power consumption and power efficiency are two crucial parameters that are often overlooked in favor of sheer speed. When building a workstation with 24×7 workload, absolute performance numbers become arguably less important compared to performance per watt. We measured the speed and power consumption of seven video cards ranging from the NVIDIA Quadro T600 to NVIDIA RTX 3070 Ti and calculated their efficiency ratings.

The use of video cards for brute-forcing passwords is nothing new. This year alone, we posted GPU Acceleration On The Cheap: Using Affordable Video Cards to Break Passwords Faster followed by GPU Acceleration: Attacking Passwords with NVIDIA RTX Series Boards, the later discussing the price/performance rating of the last generations of GPU chips. Today, we are about to make a different rating, comparing video cards by their performance relative to power consumption. Interestingly, the results are non-linear and deviate significantly from the price/performance rating.

Why some GPUs are more power-hungry than others? There are multiple factors affecting the performance and power consumption of a given chip. The GPU architecture and manufacturing process are obvious, as is the number of Compute Units installed (or activated) in a given model. The more Compute Units a GPU has, the faster it works, and the more power it consumes.

Another factor contributing to the card’s performance is the frequency, which strongly affects the power consumption and power efficiency of the card. The increase in the number of Compute Units results in linear growth of power consumption, while the growth of the card’s frequency increases the card’s power consumption with diminishing returns. For example, an NVIDIA RTX 3090 Ti with its stock TDP of 480 W is only 13% faster than the same card limited to 300 W (a 1.6 times difference) as discovered by igor’sLAB: Cool flagship instead of fusion reactor: the GeForce RTX 3090 Ti turns the efficiency list upside down with a 300-watt throttle and beats the Radeons.

With multiple tests for every NVIDIA model, do we need yet another benchmark? We believe so. Our tests use NVIDIA’s compute units exclusively, placing a different type of load on the GPU compared to the typical FPS benchmarks. In addition to previously tested models, we are adding an NVIDIA RTX 3070 Ti Founders Edition. This chip is widely criticized for its high power consumption relative to performance. With its TDP of 290 W, reviewers call it an “Inefficient side-grade with high power consumption“. Let us see if this board is as inefficient as the others say.

The following boards participate in today’s benchmark, sorted by TDP:

All the boards except the RTX 2070 were tested on the same system based on Intel’s Alder Lake i9-12900K CPU, while the RTX 2070 was tested in a workstation based on the Intel Core i5-8500 CPU. Note that GPU-accelerated attacks put little load onto the CPU (approximately 3 to 6 per cent utilization per CPU core), which makes such tests CPU-agnostic.

Using Elcomsoft Distributed Password Recovery as a benchmark, we’ve seen the following results. The benchmarks demonstrate an almost linear growth of performance depending on the video card – with one exception.

This is the exception: the RTX 2070 benchmarked higher than expected in the SHA-256 test. Since the other tests were consistent, and since we had to use a different workstation to benchmark this card, we’d recommend taking this result with a grain of salt.

 

Our power efficiency charts are based on each board’s advertised TDP. We based the “performance” part on the WinZip/AES-256 benchmark using the latest and highly optimized version of Elcomsoft Distributed Password Recovery. While not perfect, this method allowed us to do a fair comparison.

The following chart shows the performance of each GPU relative to its power consumption (the “passwords per watt” rating is based on the WinZip/AES-256 benchmark alone):

The same chart sorted by efficiency (least efficient models on top):

Frankly, the results were quite unexpected. The slowest tested video card, the Quadro T600 model from NVIDIA’s professional range, has also become the most efficient, demonstrating the highest performance per watt. Is this due to the optimizations of the installed GPU or are professional cards targeting power efficiency by design? We don’t have enough data, but some reports suggest so. You can check out Igor’sLAB review of the NVIDIA A5000 server board to see if there is any truth in this speculation.

Interestingly, the NVIDIA RTX 3070 Ti, which is the fastest unit among the tested boards, is not quite an “inefficient outsider with high power consumption.” High power consumption? Quite so. Outsider? Not really: in our tasks, the card outperforms the NVIDIA RTX 3060 and 3050 models in terms of power efficiency, and is not far off the NVIDIA RTX 3060 Ti.

This publication is the first in the series of articles about the power consumption, heat dissipation and energy-efficiency of forensic workstations doing password recovery. Today’s article covers the power efficiency of today’s video cards relative to their performance. In subsequent articles we’ll talk about the waste heat, GPU cooling solutions, and the differences between consumer and server boards.

 By Oleg Afonin at 2022-07-08 18:01:43 Source ElcomSoft blog: