As stated last year, the recommended ideal RAM capacity for gaming and AI workloads, was 96GB of system memory. Since then, a lot of things have changed, and a 96GB kit might no longer be a recommended option.
The memory market has changed dramatically. AI infrastructure is consuming enormous amounts of DRAM and HBM capacity, putting significant pressure on consumer DRAM supply and pricing. In 2026, 96GB kits can be difficult to find or carry a substantial price premium, making them considerably less attractive for a mainstream enthusiast build.
So, back to the lab trying to find the best solution given the circumstances.
32GB and 64GB configurations were compared using an RTX 4090 and local LLM workloads, while also examining the effect of DDR5 performance.
The testing system was the following:
The conclusion is straightforward: For a modern gaming and AI system, 64GB of DDR5 is the sweet spot and DDR5-6000 C40 offers an excellent balance of capacity, performance, and practicality.
As expected, 32GB of memory can’t handle a large LLM with 70B parameters. The RAM utilization is well over 95% and it makes the system slow. On the other hand, the 64GB system had ~15GB of RAM left, which allows the users to perform other tasks simultaneously.
On the tokens generated time, there’s not a big difference noticed. The 64GB 6000C40 kit is slightly ahead without that meaning that is practically the best. For this specific graph, the fastest vs the slowest runs were selected between the 2 kits.
Although, on average of 3 runs the results are the same. The 5600C40 and the 6000C40 kits needed, almost, the same time to complete the prompts. The 6000C40 kit was slightly faster by 0.5 seconds but that’s within the margin of error.
Overall, 64GB provided enough memory capacity to run substantially larger local AI workloads without pushing the system into severe memory pressure. That distinction matters.
But why local LLMs? Why not use an online API like ChatGPT or others? Privacy. Data that users import to any online service could be exposed to the public. By running an LLM on a machine that you control, you don’t have to connect to the internet and control its behavior to the maximum. Local LLMs’ benefits don’t stop at privacy only. Cost, customization, and context length are important factors that benefit the local LLMs.
For the testing the Meta's Llama 3.1 70B Instruct model was used, running locally through Ollama. The 70B model contains approximately 70 billion parameters and supports a 128K context window. This isn't intended to suggest that every AI user should run a 70B model. Although larger models generally have more capacity to represent knowledge and relationships and can perform better on demanding reasoning, coding, analysis, and instruction-following tasks.
An RTX 4090 VGA provides 24GB of dedicated VRAM. That's a tremendous amount of graphics memory, but it isn't necessarily enough for larger local AI models. Once a model exceeds the available VRAM, part of the workload has to reside in system memory (RAM). That's where system RAM becomes much more than a resource for Windows, games, and background applications.
For the primary performance workload, the Llama3.1 8B was selected. It is substantially smaller than the 24GB of VRAM available on the RTX4090, meaning system memory becomes a less important part of the workload.
Rather than asking the model to perform an artificial benchmark such as generating thousands of words, tasks representative of everyday AI usage was used.
These included activities such as planning, summarization, analysis, and structured content generation.
The following prompt was used: “I’m planning a 10-day trip to Japan covering Tokyo, Kyoto, Osaka, and Hiroshima. Create a detailed day-by-day itinerary. For each day, include the main attractions, the order in which I should visit them, approximate travel time between locations, suggested restaurants or food areas, and practical tips. Explain the reasoning behind the itinerary and identify any potential scheduling problems. At the end, provide a concise packing list and a summary of the total time spent traveling versus sightseeing.”
The PC enthusiast market is changing.
They provide enough capacity for modern gaming, multitasking, and meaningful local AI workloads while avoiding the cost and availability problems increasingly associated with larger kits.
Combined with DDR5-6000 C40 specs, they provide a strong balance between capacity and memory performance.
記事内の製品
Stay up to date with CORSAIR. Get our latest News, Guides, and Product Updates in your Google feeds.
Add CORSAIR as a preferred sourceThanks for helping us improve this article.
JOIN OUR OFFICIAL CORSAIR COMMUNITIES
Join our official CORSAIR Communities! Whether you're new or old to PC Building, have questions about our products, or want to chat about the latest PC, tech, and gaming trends, our community is the place for you.