Installing the NVIDIA GeForce RTX 4070: First Impressions and Hash Cracking Benchmarks

Computer Hardware

9/22/2025

Motivation for Installing a GPU

Upgrading my HP Elite Tower 600 G9 with a dedicated Graphics Processing Unit (GPU) was driven by a passion for advancing personal capabilities in experimentation with Artificial Intelligence (AI) and cybersecurity research. This enhancement is not only a technical improvement but a strategic investment for harnessing the power of modern local AI workloads and brute-force password simulations (as an ethical hacker).

Motivation for Local LLM Experimentation

Running Large Language Models (LLMs) locally empowers greater control over data, enables deeper model customization, and fosters real-time experimentation without relying on expensive or restrictive cloud services, or pay-per-use OpenAI APIs. The ability to directly evaluate different LLM architectures and prompt engineering strategies on personal hardware translates to enhanced learning outcomes and impactful research.

Consumer GPUs such as the NVIDIA GeForce RTX 4070 provide sufficient VRAM and compute throughput, making them suitable for inference and model fine-tuning tasks that were previously the domain of costly datacenter infrastructure.

Rationale in Cybersecurity Research

Modern cybersecurity work demands large-scale, parallel brute-force simulations to evaluate password security mechanisms and test resilience against evolving threats. With a high-end GPU, tools like Hashcat and John the Ripper become much more effective, enabling the testing of billions of password combinations per second. This capability supports proactive defense, advanced vulnerability analysis, and the ability to stay ahead of potential attackers by simulating realistic adversarial scenarios.

Personal and Professional Growth

Investing in high-performance local hardware embodies my commitment to continuous learning and hands-on exploration in AI and cybersecurity. For me, true understanding comes from practical experience, experimenting, building, and troubleshooting directly, rather than relying solely on theoretical knowledge. This approach allows me to internalize complex concepts and remain adaptable in a fast-evolving technological landscape, ensuring that my expertise is grounded in real-world application.

Hardware Starting Point

The upgrade journey began with an HP Elite Tower 600 G9. The motherboard includes a PCIe Gen4 x16 slot, which enables seamless integration of powerful GPUs like those from NVIDIA, essential for local LLM experimentation and brute-force attacks simulation.

• HP Elite Tower 600 G9, equipped with a 64GB DDR5 SDRAM and a 2TB NVMe SSD
• Intel 12th/13th Gen Core CPUs, compatible with high-performance graphics cards via a PCIe Gen4 x16 slot.
• Initial system configuration with a 400W power supply unit (PSU), adequate for standard tasks but insufficient for power-hungry GPUs and enhanced cooling needs.

Upgrade Goals and Modifications

To fully realize the potential of local LLM inference and intensive brute-force password cracking, the following upgrades are targeted:

• Installation of an NVIDIA RTX GPU using the motherboard’s PCIe Gen4 x16 slot for optimal bandwidth and stability
• Upgrade to an HP 550W power supply to meet the significant power requirements of modern GPUs while maintaining operational safety and reliability
• Enhanced cooling with a stronger fan and improved airflow solutions, critical for sustaining performance during prolonged high-load simulations and AI compute sessions

Selection of the NVIDIA Model

The RTX 4070 was selected for its suitable balance of price, power efficiency, and strong support for running modern Large Language Models (LLMs) such as Llama 2 and Llama 3 at both 7B and 13B parameter sizes with quantized weights. It delivers fast inference for real-world text generation, provides code assistance and cybersecurity automation tasks.

With 12GB GDDR6X memory and dedicated Tensor Cores, this GPU allows efficient execution of high-performance AI workloads, including local inference, smaller model fine-tuning, and the use of advanced toolchains like llama.cpp, Ollama, and TensorRT-LLM.

Cooling

For thermal management and system stability, the Thermalright Peerless Assassin 90 SE was chosen. Upgrading the cooling solution is critical, as running sustained AI inferencing or brute-force operations places a high continuous demand on the hardware, generating significant heat that must be efficiently dissipated to prevent thermal throttling and ensure optimal system longevity.

Upgrade Process of the Hardware

Upgrading the hardware requires careful preparation and organization. It starts by powering down the system and unplugging the power cable. I ensure that my workspace is tidy, well-lit, and that all required tools are readily accessible.

To minimize the risk of static discharge, which can irreparably damage sensitive electronics, I work on a non-carpeted surface and regularly ground myself by touching the metal chassis of the PC. This avoids building up an electrical charge between me and the components.

Here’s the sequence I followed:

• Removed the old power supply and installed an HP 550W PSU (M86264-001)
• Mounted the new cooler unit (Thermalright PA90 SE), being sure to apply fresh thermal paste to the CPU
• Installed the NVIDIA RTX 4070 GPU

The full upgrade took around 3,5 hours. This included extra time for adapting the cooler mount to the CPU and carefully routing cables for optimal airflow and neatness. I recommend photographing each step. This makes it much easier to retrace steps or troubleshoot if anything doesn’t work as expected.

By preparing thoroughly, following best practices for handling components, and documenting progress, I set myself up for a smoother, more successful upgrade experience.

Test Environment

What is Hashcat (GPU)?

Hashcat is an advanced, open-source password recovery tool specifically engineered to leverage the massive parallel computing power of modern GPUs. By distributing brute-force or dictionary attack calculations across thousands of GPU cores, Hashcat can test billions of password candidates per second.

This makes Hashcat exceptionally effective and fast for assessing password security, recovering lost credentials, and conducting penetration testing on password-protected hashes like NTLM, SHA-family, and bcrypt.

Hashcat supports various attack modes, including brute-force, mask, dictionary, hybrid, and rule-based, offering flexibility for different password recovery challenges and security audit scenarios.

What is John the Ripper (CPU)?

John the Ripper (often called "John") is another highly reputable password cracking tool. While it can utilize GPU acceleration (in its "jumbo" builds), John is traditionally CPU-focused and known for its powerful built-in wordlists, smart hybrid attacks, and ability to work across a huge range of hash formats.

John is favored in many forensics and compliance projects for its automation, customizability, and unique incremental attack modes, which can make it very effective even on systems with limited or no GPU resources. Like Hashcat, John supports brute-force, dictionary, and combinator attacks, as well as specialized techniques for certain hash types.

In real life, a wordlist such as rockyou.txt containing millions of real-world passwords would be used to associate hashes with passwords that are commonly used. This archive was sourced from the RockYou data breach in 2009. When performing password security testing, tools like Hashcat and John the Ripper use these wordlists by systematically trying each password in the list against a hash until a match is found.

Instead of using wordlists with common passwords, I created my own "PINs" to simplify and better control the simulation.

Benchmarking GPU with Hashcat against CPU with John the Ripper

To demonstrate the difference modern hardware can make in password security research, I generated 100.000 unique NTLM hashes using a short Python script.

The association between MD4.new and NTLM in the Python code is that NTLM (NT LAN Manager) was an authentication protocol used by Microsoft systems and its core password hashing algorithm is MD4.

Each hash corresponded to a numeric password in the format of exactly nine digits, ranging from "000000000" to "999999999". (NTLM is an unsalted cryptographic representation of user passwords by Microsoft that has actually been deprecated due its weaknesses.)

This setup ensured a brute-force mask pattern would require a candidate space of exactly 1.000.000.000 possibilities (ten to the ninth power), simulating realistic worst-case enumeration for PIN-based credentials.

For the brute-force trial, I leveraged two tools:

• Hashcat (utilizing an NVIDIA GeForce RTX 4070 GPU)
• John the Ripper (running on my workstation's CPU)

The command to measure the time for hash cracking with Hashcat on the GPU in Git Bash is:

$ time ./hashcat.exe -m 1000 -a 3 --quiet ntlm_9digit_hashes.txt ?d?d?d?d?d?d?d?d?d

The command to measure the time for hash cracking with John the Ripper on the CPU with Powershell is:

Measure-Command { .\john.exe --format=nt --incremental=Digits hashcat-7.1.2\ntlm_9digit_hashes.txt }

Results of Benchmark Tests

• Hashcat (GPU, RTX 4070): 1,79 seconds to process and crack all 100.000 nine-digit NTLM hashes.
• John the Ripper (CPU): 18,96 seconds for the same workload and search space.

These numbers highlight the advantage a modern, consumer-grade GPU brings to large-scale password cracking tasks. Hashcat, tapping into thousands of GPU cores in parallel, completed the scan in less than two seconds, over ten times faster than my CPU-based attempt using John the Ripper. (The latter is optimized for CPU, as it is running natively. I tried to use more cores/threads for parallel processing, but it did not produce a gain apparently because of the overhead of inefficient workload distribution.)

This experiment underscores not only the need for strong, non-numeric passwords but also the critical importance of understanding the real-world speed at which even huge hash-sets can be brute-forced with today's commodity hardware.

Importance of Running the Latest Official NVIDIA GPU Drivers

To achieve the best performance for hash cracking simulations with the NVIDIA GeForce RTX 4070, it is crucial to use the latest official NVIDIA GPU drivers. NVIDIA's current Studio drivers fully support CUDA 12.x, which ensures maximum compatibility and speed in GPU-accelerated applications like Hashcat.

CUDA stands for "Compute Unified Device Architecture." It is a proprietary parallel computing platform and programming model developed by NVIDIA that enables software to use NVIDIA GPUs for general-purpose (non-graphics) processing tasks.

Before benchmarking, I need to ensure that CUDA is detected by Hashcat. A message "Successfully initialized the NVIDIA main driver CUDA runtime library." should appear after running

hashcat -l

Monitoring the GPU Usage with MSI Afterburner

For monitoring and tuning, MSI Afterburner is a highly effective utility. It allows real-time tracking of GPU metrics such as core load, temperature, clocks, power usage, and fan speed. With its on-screen display feature, it is possible to track the performance metrics live during hashcat operations. One can also adjust power and thermal limits, and even fine-tune clock speeds and voltages for stable, higher sustained throughput. By observing stability and utilization in MSI Afterburner, adjustments can be made on the fly to avoid throttling or unnecessary power draw during extended cracking runs. Here is a snapshot of a 16-digit hash run (9-digit hash run gives just a small peak):

Summary

In summary, investing in powerful local hardware has been essential for transforming curiosity and theoretical knowledge into real, hands-on expertise in cybersecurity and AI. The careful upgrade process of my workstation, highlighted by the installation of a new power supply, an advanced cooling solution, and the NVIDIA GeForce RTX 4070, laid the foundation for meaningful experimentation and learning.

The first round of benchmark tests confirmed the significant advantage of GPU-accelerated password cracking. Hashcat on the RTX 4070 completed the task of brute-forcing 100.000 NTLM 9-digit numeric hashes in under two seconds, while John the Ripper on CPU required almost twenty seconds for the same challenge.

These results demonstrate how hardware selection and the right tools can impact both research efficiency and practical defensive capabilities. This project underscores the value of continuous, practice-oriented learning, anchoring technical exploration in real hardware and real-world use cases.

Go back