AI development is pushing past the initial realm of the large AI providers and hyperscale data centers and into the enterprise. Many enterprise developers are leveraging pretrained foundation models with proprietary data to retrain, optimize, and/or build on the open-source models. However, security concerns combined with the cost of on-premises data center infrastructure and/or cloud resources are driving a need for an alternative solution. This is forging a new opportunity for workstations, systems that can bridge the performance gap between PCs and servers with a more palatable cost and flexible form factor.
Disclosure: My company, Tirias Research, has consulted for AMD, Nvidia, Micron and other companies mentioned in this article.
To address this, AMD has released its new Ryzen Threadripper 9000 Series, the Threadripper 9970X and Threadripper 9980X. This latest generation of workstation processors empowers desktop and deskside workstations to fully utilize the latest advancements in large language models, STEM (science, technology, engineering, and mathematics) applications, and game development. With 32 cores and 64 threads, the Ryzen Threadripper 9970X demonstrates power and reliability across all applications. Whereas the 9980X with 64 cores and 128 threads is the powerhouse of the series, offering thermal and power efficiency even when running exceedingly CPU-intensive simulations. As AI continues to advance and the demand for local AI rises for individuals, developers will benefit from investing in AI-ready developer systems.
The Threadripper 9970X is likely a go-to choice for developers considering its moderate price point of $2,499 and broad applicability. Alternatively, the Threadripper 9980X can utilize the most stressful multi-threaded applications at a higher price of $4,999. Tirias Research built a test unit with an approximate cost of $4,900 for the Threadripper 9970X and $7,400 for the Threadripper 9980X system, excluding the GPU and PC case. Users who need to upgrade their case and GPU to properly support the Threadripper system may need to invest an additional $1,300 to $3,000.
We put these processors to the test to measure performance in frame generation, token generation, power utilization, and thermal efficiency. Each processor was tested through a series of CPU benchmarks, local Llama3 projects, and CPU-intensive video games. HWiNFO and MSI Afterburner were used to record any metrics not included in the testing applications.
Benchmark performance
As anticipated, the 64-core/128-thread Threadripper 9980X outperformed the 32-core/64-thread Threadripper 9970X on CPU-intensive benchmarks, including Cinebench 2024 multi-core, 3Dmark’s Time Spy, and PugetBench for After Effects and Premiere Pro. The Threadripper 9980X nearly doubled the performance of the Threadripper 9970X on the V-Ray 6 test, demonstrating the platform’s scalability with an application that effectively leverages the higher core/thread count.
However, the Threadripper 9970X performed similarly to the 9980X on single-core tests, such as Cinebench and GeekBench. It even outperformed the 9980X on benchmarks like the PugetBench for Photoshop and GeekBench AI using both ONNX and OpenVINO. This demonstrated that the lower core-count Threadripper 9970X isn’t just a lower-cost option, but should be seen as a better choice for applications optimized for a lower core/thread count.
STEM and AI applications
As seen in the benchmarks, the Threadripper 9980X excels in the most demanding CPU applications, while the Threadripper 9970X offers an efficient alternative, providing comparable performance and efficiency in applications that do not utilize all 64 cores/128 threads of the 9980X. AI applications, in particular, excel when supported by a powerful AI-enabled GPU and extensive virtual memory. When testing Blender for the CPU, the Threadripper 9980X outperformed the Threadripper 9970X with nearly double the score in each test. With Matlab R2025a, the 9970X produced near equivalent scores to the 9980X and a slight improvement in graphics performance.
One of the most significant benefits of the high-core Threadripper processors is the ability to run large language models locally, thereby avoiding the costs associated with AI cloud platforms. Both Threadrippers were tested for their ability to run AI and LLMs locally with the Llama-3.1-8B-Instruct model using PyTorch. This is a new benchmarking metric developed by Tirias Research, specifically designed to evaluate the latest workstation hardware capabilities in conjunction with local AI applications. Each processor was tested at bfloat16 and float16 quantization for 128-token and 512-token context and generation lengths, in batch sizes of 1 and 4, using 32 and 64 threads to measure throughput in tokens per second, average complete response latency in seconds, and average first token latency in seconds. Larger batch sizes and context lengths contribute to longer total response latency, as well as a slight decrease in throughput. Both Threadrippers maintained similar performance and latency during every test, with a slight performance increase when using float16 instead of bfloat16. Generating between 5 and 7 tokens per second demonstrates that each Threadripper is capable of generating multiple words per second and can stream text at a normal reading speed even without a GPU for acceleration. With the added ability of handling multiple high-end GPUs, the Threadripper platforms are ideal for local AI development.
Workstation vs. the cloud
While Threadripper workstations come with upfront costs, developer access to a local LLM development environment is increasingly cost-efficient when compared to cloud-based services over thousands of work hours. As an example, the estimated upfront investment investment including the case and GPU is $9,700 for the Threadripper 9980X system and $7,200 for the Threadripper 9970X system. This combined with the 350W power draw by both processors results in an added cost of $0.0525 in electricity per hour assuming a rate of $0.15 per kWh. We compared this to the average cost A100 instances from ThunderCompute, Lambda, and AWS. ThunderCompute’s On-demand A100 (40GB) service offers the most competitive pricing at $0.66 per GPU hour. While Lambda’s Nvidia A100 (40GB) service is priced at $1.29 per GPU/Hour, and AWS’ A100 p4d.24xlarge service is priced the highest at $2.74 per GPU/Hour. “While the A100 cloud instances are a cost-effective option, an engineering team could see a cost benefit from a Threadripper workstation is just a few quarters with the additional flexibility of running non-AI workloads and the security of maintaining all data locally,” according to Tirias Research Analyst Caiden McGregor.
Gaming performance
Because developers spend extensive hours working on high-end computer systems, most developers share a common interest in video games. While these Threadripper workstations were not designed for gaming, we thought it would be interesting to at least put them through the gaming wringer. To verify that the Threadripper can play as hard as it can work, the systems were tested using a catalog of modern games with CPU-intensive requirements. MSI Afterburner and HWiNFO were used to record frames per second, CPU temperature, thermal limit, active core count, and package power. All games were configured to minimize graphics processes and maximize the CPU-demanding features. Additionally, built-in performance tests were utilized when available. To prevent bottlenecking in the GPU, all texture-related settings were set to minimum with a constant 1920×1080 resolution. Whereas CPU-related settings, such as crowd or traffic density, were set to maximum. Both processors demonstrated powerful results, with the 9980X processor promoting cooler temperatures and the highest power efficiency, while the 9970X maintained comparable or superior frame generation across all games. Even when leveraging the industry’s highest performing GPU, the CPU has a powerful impact on core gaming functionality. However, most games are not designed for a 64-core system, which can result in higher latency and lower performance efficiency with the 9980X.
Additional Impressions
Each Threadripper processor comes with a calibrated torque wrench making it a straightforward installation with the compatible ASUS TRX50 motherboard. An important note is that these Threadripper workstations are not your average workstations in terms of power and heat dissipation. As a result, a liquid cooler and high-quality thermal paste is required to support the high performance of the Threadripper processors.
Conclusions
The Ryzen Threadripper processors have been king of the hill since the initial introduction of the product family in 2017. The latest Ryzen Threadripper 9000 series processors based on the Zen 5 processor architecture continue to set the standard by demonstrating remarkable performance in all metrics. According to Tirias Research Analyst Damian McGregor, “the Threadripper 9980X processor is a beast that excels at high-core and thread-intensive tasks, whereas the Threadripper 9970X processor is the efficient little sibling that performs at equivalent or greater performance for less thread-intensive applications.”
Both Threadripper processors demonstrate excellent performance for developer applications, including Blender, Matlab, Adobe, and Llama3. While neither processor is optimized for gaming, the robust power of the Threadripper 9000 series paired with a high-performance GPU like the Nvidia RTX 5090 will guarantee steady framerates at even the highest performance settings.
In the end, the Threadripper 9000 series workstations are powerful solutions for top-end gaming, local AI, and STEM projects. Developers will receive optimum performance from the Threadripper 9980X processor for high-core demanding applications, while the Threadripper 9970X processor will provide a balanced experience at a lower price point. Additionally, both Threadrippers will provide developers with higher flexibility and increased data security at a reasonable value compared to cloud-based solutions.