ARE TOO MANY CORES SLOWING DOWN YOUR GPU?

Surprisingly, and very specifically in 3DMark Time Spy, the answer appears to be yes. During extensive tests conducted at our OSHKO LAB research center, we observed a Ryzen 9 GPU performance bottleneck when pairing the RTX 5090 with AMD’s high-core-count 9950X and 9950X3D CPUs.

Despite their flagship specs, these chips consistently produced lower Time Spy scores than the Ryzen 7 9700X and 9800X3D, processors positioned a tier or two below in the lineup.

Notably, this anomaly did not occur in other benchmarks like Port Royal or Speed Way, making the Ryzen 9 GPU performance bottleneck both real and benchmark-specific.

This observation contradicted our expectations based on theoretical specs and immediately caught our attention. It became clear that a performance bottleneck might be at play. Was it merely statistical variance, an external bottleneck, or the symptom of a deeper CPU to GPU resource management quirk?

THE BUILDS & SETTINGS

The test rigs were OSHKO Creator Pro systems equipped with:

• ASUS GeForce RTX 5090 32GB
• 128GB DDR5 5600 MHz RAM
• 1TB PCIe 990 PRO M.2 SSD

THE PROCESSORS – TECHNICAL SPECIFICATIONS

The four CPUs share a modern architecture but differ in configuration. For full specifications, you can visit the AMD Ryzen processor page.

• Ryzen 7 9700X and 9800X3D feature 8 cores / 16 threads, while Ryzen 9 9950X3D and 9950X double that with 16 cores / 32 threads.
• Boost clocks: 9700X up to 5.5 GHz, 9800X3D 5.2 GHz, and the Ryzen 9s reach 5.7 GHz.
• “3D” variants (9800X3D and 9950X3D) stand out with massive L3 cache, 96 MB and 128 MB respectively, versus 32 MB on the 9700X and 64 MB on the 5950X.
• Power draw: Ryzen 9s show a 170 W TDP, versus 120 W for the 9800X3D and just 65 W for the 9700X, underscoring the entry‑level model’s efficiency.
• Pricing: 9700X is most affordable at $400, followed by the 9800X3D at $850, the 9950X at $900 and the more high‑end 9950X3D at $1,000.

Click on the table to enlarge the image.

INITIAL OBSERVATION: A SURPRISING ANOMALY

Running our 3DMark Time Spy benchmarks, the outcomes were unexpected. Intuition says the higher end parts (with double the cores) should at least match, if not surpass, their lower tier siblings.

Click on the table to enlarge the image.

ANALYSIS AND INITIAL HYPOTHESIS

Faced with this situation, we launched an extensive investigation. Monitoring data clearly showed the RTX 5090 was utilized at only about 70–80 % of its capacity with the Ryzen 9 9950X and 9950X3D, whereas it hit 100 % with the other CPUs.

To interpret these results, we consulted domain specialists to help us get a better understanding of the situation. A shared hypothesis emerged: the high number of active cores on the Ryzen 9s might trigger scheduling conflicts or inefficiencies in task assignment, preventing the GPU from being fully leveraged.

VALIDATION METHOD: REDUCING CORES TO IMPROVE PERFORMANCE?

To test our hypothesis, we ran multiple passes with half the physical cores disabled on the Ryzen 9 9950X and 9950X3D.

We examined three scenarios:

• 8 cores even split (4 active per CCD)
• 8 cores on CCD 0 only
• 8 cores on CCD 1 only

PROCEDURE

These configurations were applied via ASUS advanced BIOS settings (valid for Prime, ProArt, and TUF Gaming AMD 800 series motherboards):

• Enter BIOS at boot by pressing Delete.
• Switch to Advanced Mode with F7.
• Navigate to Advanced → AMD Overclocking.
• Accept AMD’s terms (select “Accept”).
• Go to Manual CPU Overclocking → CPU Core Count Control.
• In CCD 00 Bit Map Down Core Control and CCD 01 Bit Map Down Core Control, set cores active (1) or inactive (0).
• Adjust the bits for the desired setup (e.g., disable all cores on CCD 01 to keep only those on CCD 00).
• Apply changes via Bit Map Down Core Apply Changes.

COMPARISON BETWEEN THE DIFFERENT CORE CONFIGURATIONS (CCD)

Comparing the three 8 core modes (even split, CCD 00 only, CCD 01 only) shows very little variance on the non 3D 9950X (≤ 1 %) and modest gaps on the 9950X3D (max ≃ 3.3 %). On the non 3D chip, an even 4 per CCD split is most efficient; on the 3D variant, CCD 00 alone wins, likely the one housing the 3D V Cache.

Still, these gaps are small and could stem from measurement variability or benchmark precision; drawing firm conclusions about one CCD being “better” would be risky.

What is clear, however, is that fewer total cores markedly improve the GPU bound Time Spy benchmark. This suggests that too many active cores can disrupt the GPU pipeline, and focusing the load on a single CCD remains a sound strategy.

Click on the table to enlarge the image.

Click on the table to enlarge the image.

INITIAL RESULTS VS POST ADJUSTMENT

The Time Spy figures reveal an interesting trend: reducing active cores improves GPU usage. On the 16 core Ryzen 9s, certain layouts cause poor resource distribution or thread saturation, throttling graphics performance. Cutting back to just 8 active cores boosted scores significantly, +19.7 % on the Ryzen 9 9950X and +22.8 % on the 9950X3D.

Click on the table to enlarge the image.

CONCLUSION AND PRACTICAL IMPLICATIONS

This experiment highlights an often overlooked fact: more cores don’t always equal better performance, especially when GPU utilization is the priority. In our case, disabling some cores on the Ryzen 9 helped eliminate what we identified as a Ryzen 9 GPU performance bottleneck, restoring optimal GPU performance in Time Spy. For heavily parallel tasks (3D rendering, editing, compilation), keeping all 16 cores active is still advisable.

If you see disappointing benchmark or gaming results with a high end CPU, trying an 8 core configuration may be worthwhile. Yet this fix is highly situational. It worked in our test environment but may not suit other setups or workloads. Proceed cautiously before tweaking your BIOS and be sure to back up your settings.

Finally, a question remains: why does this phenomenon appear only in 3DMark Time Spy and not in Speed Way or Port Royal? The exact cause is still unclear and deserves deeper investigation.

A second point also warrants study: what level of bottleneck would we see with other GPUs of the same generation, like the RTX 5080 or 5070 Ti, compared with the 5090? Exploring this could shed more light on the CPU configuration’s real world impact as hardware evolves.

Feel free to share your own experiences or ask about your setup. The OSHKO team is here to help you get the most out of your hardware.