Two Cards in the Same PC — How Far Apart Are They? RX 9060 XT vs RTX 3090 on Speed and Power Cost (Part 2 of the RX 9060 XT Series)

September 27, 2026

Last time, I confirmed that a Radeon RX 9060 XT installed in my desktop PC is visible to local-LLM software without installing any special extra software. I didn’t measure speed in that article and left it for the next one.

The same PC also has an RTX 3090 in it. Inside the same machine, with the same model and the same measurement method, how far does this new 16GB card get when the two cards are lined up side by side?

These are measurements taken as of September 2026.

The previous article in this series is here.

Overview of the Hardware Used

Here’s the desktop PC used for the measurements. It’s the same setup as last time, with two GPUs inside one machine.

CPU	AMD Ryzen 9 3950X (16 cores)
Memory	62GB
GPU	RTX 3090 (slot wired directly to the CPU), ASUS DUAL Radeon RX 9060 XT 16GB (chipset-connected slot)
OS	Ubuntu 24.04.4 LTS (kernel 7.0.0-30-generic)
Display output	Connected to the RTX 3090 (no monitor attached to the RX 9060 XT)

GPU Used: ASUS DUAL Radeon RX 9060 XT 16GB (DUAL-RX9060XT-16G)

It’s a plain-looking card with two fans side by side. I measured it exactly as I installed it last time, without pulling it out or reseating it.

As an Amazon Associate we earn from qualifying purchases.

What I Measured

The speed of a local LLM comes in two different kinds: how fast the AI reads the text you send it, and how fast the AI writes out its answer. In the previous article, I lined up write speed and power draw for 4 cards.

This time, I pull the read speed, which I hadn’t shown yet, out of that same measurement, and look at the following 3 things alongside it.

• Does the gap between the two cards open up the same way for read speed and write speed?
• When the card keeps writing for a long time, does it hold that speed?
• Per unit of electricity cost, which one writes more?

I compared only the two cards in the same PC: the RTX 3090 and the RX 9060 XT. Lining up GPUs that sit in different machines would let differences in software versions and connections leak into the numbers.

How I Measured It

Here’s how I measured. Both cards used the same machine, the same software, and the same weight files.

Date measured	September 6, 2026
Models	qwen3:8b (4.87GB) and qwen3:14b (8.64GB). Both in a compressed format called Q4_K_M
Benchmark tool	llama-bench, which comes with llama.cpp (build b9859)
Backend	Vulkan for both cards. Neither NVIDIA’s CUDA nor AMD’s ROCm was used
Read speed	Speed when reading an input of 512 tokens
Write speed	Speed when writing 128 tokens. Launched separately from the read measurement
Runs	4 runs each. The 1st run includes loading the model, so it was dropped; the value is the median of the remaining 3
Power	Measured separately from speed. Generation was run 20 times in a row, sampled every 0.25 seconds. Median of the stretch after discarding the ramp-up and the tail
Electricity cost	Calculated at ¥31 per kWh

A token is the small unit that text gets chopped into; for Japanese text, one character is roughly 1–2 tokens.

One note about the number of runs. I took these measurements on September 6, and at the time I measured 4 times and dropped the 1st. Since then I’ve changed the method to measuring 5 times, dropping the highest and the lowest, and taking the median of the remaining 3 points.

So on September 23, with the same machine, the same software, and the same weights, I re-measured all 8 combinations above, 5 times each.

GPU	Model	What was measured	September 6 [tok/s]	September 23 [tok/s]	Difference
RTX 3090	qwen3:8b	Read speed	4,628.4	4,588.4	−0.9%
RTX 3090	qwen3:8b	Write speed	119.9	120.2	+0.3%
RTX 3090	qwen3:14b	Read speed	2,772.8	2,738.2	−1.2%
RTX 3090	qwen3:14b	Write speed	77.5	77.3	−0.2%
RX 9060 XT	qwen3:8b	Read speed	2,125.5	2,118.5	−0.3%
RX 9060 XT	qwen3:8b	Write speed	54.5	54.6	+0.1%
RX 9060 XT	qwen3:14b	Read speed	1,182.5	1,173.3	−0.8%
RX 9060 XT	qwen3:14b	Write speed	32.1	32.1	±0%

The largest shift was 1.2%. The numbers below use the September 6 set as-is, since it lines up with the power measurements taken the same day.

Does the Gap Change Between Read Speed and Write Speed?

Here are the results for the two cards.

GPU	Model	Read speed [tok/s]	Write speed [tok/s]
RTX 3090 24GB	qwen3:8b	4,628.4	119.9
RX 9060 XT 16GB	qwen3:8b	2,125.5	54.5
RTX 3090 24GB	qwen3:14b	2,772.8	77.5
RX 9060 XT 16GB	qwen3:14b	1,182.5	32.1

The RX 9060 XT comes in at roughly 40% of the RTX 3090. What caught my eye here was that this ratio was almost the same for read speed and for write speed.

Model	Read speed ratio	Write speed ratio
qwen3:8b	45.9%	45.5%
qwen3:14b	42.6%	41.5%

The two speeds are normally decided by different parts of the card. Read speed depends mainly on raw compute power; write speed depends mainly on how fast data can be read out of memory. Even so, the ratios matched, so between these two cards, the order and the size of the gap stayed the same whichever use you look at.

It doesn’t look like a case of picking one card for short questions and the other for having it read long documents.

Does the Speed Hold Up When It Keeps Writing?

The numbers above come from short measurements that finish in a few seconds each. In real use, you have it write for much longer. So I ran generation 20 times in a row and looked at the speed over that stretch as well.

GPU	Model	Short measurement [tok/s]	Running continuously [tok/s]	Length of the run	Difference
RTX 3090 24GB	qwen3:8b	119.9	119.1	23 seconds	−0.7%
RX 9060 XT 16GB	qwen3:8b	54.5	54.5	50 seconds	−0.1%
RTX 3090 24GB	qwen3:14b	77.5	61.4	45 seconds	−20.8%
RX 9060 XT 16GB	qwen3:14b	32.1	31.8	85 seconds	−0.9%

The only one that dropped was the RTX 3090 on 14B, by 20%. The RX 9060 XT stayed within 1% on both models.

That difference narrows the distance between the two cards. In the short measurement the RX 9060 XT was at 41.5% of the RTX 3090, but when writing continuously it was at 51.9%.

Why Did Only the RTX 3090 on 14B Drop?

Looking at the power log, the RTX 3090 was pinned at 344–346W while generating. The minimum of 178.9W is from the ramp-up; after that it stays almost flat.

To see what this number is running into, I compared it with the card’s rating. NVIDIA’s published graphics card power for the GeForce RTX 3090 is 350W. 344–346W is one step short of that. It most likely hit the ceiling and couldn’t go any higher.

If this comes from hitting that ceiling, the same card shouldn’t show it on the lighter 8B. And in fact, the drop on 8B stayed at 0.7%. It happening only on 14B doesn’t contradict this reading.

That said, I haven’t separated whether it is being held back by the power ceiling or by heat. I didn’t record temperature or clock speed in this measurement, so this part is unconfirmed.

As a side note, I still have a record from August, when I measured a different job (image generation) on the same RTX 3090 in the same machine. In that one I captured the value the GPU itself reports as the reason it slowed down, and both the power limit and the thermal limit were flagged at the same time. Power was 345–352W, and in some runs it stopped right at 350W.

So framing it as “power or heat?" is probably the wrong question to begin with. Both can be in effect at once. Since that record is from a different job in a different month, I can’t apply it directly to this 14B result, but next time I measure I plan to capture the reason values as well.

How Much Can It Write per Unit of Electricity Cost?

Power draw while generating was about 345W for the RTX 3090 and 159W for the RX 9060 XT. At ¥31 per kWh, here’s how much each can write for ¥1.

GPU	Model	Per ¥1, short measurement [tok]	Per ¥1, running continuously [tok]
RTX 3090 24GB	qwen3:8b	40,465	40,198
RX 9060 XT 16GB	qwen3:8b	39,834	39,813
RTX 3090 24GB	qwen3:14b	26,114	20,685
RX 9060 XT 16GB	qwen3:14b	23,460	23,248

In the short measurement the RTX 3090 comes out ahead, but when writing 14B continuously the two swap places. Speed drops by 20% while power doesn’t go down, so the amount per ¥1 shrinks.

The continuous-run figures are calculated from the measured speed and the measured power: amount per ¥1 = write speed × 3600 ÷ (power ÷ 1000 × ¥31). This is GPU power only; the CPU and power-supply losses aren’t included.

Power draw while doing nothing was another place with a big gap.

GPU	Idle power [W]	Monthly cost if left on 24 hours a day
RTX 3090 24GB	15.4	About ¥344
RX 9060 XT 16GB	3.0	About ¥67

These two are not under the same conditions. As I wrote last time, the RTX 3090 is handling the monitor and the desktop display, and that load is included. There’s no monitor attached to the RX 9060 XT. These aren’t numbers to rank side by side. Treat each as a value for its own particular state.

In the previous article I gave the RX 9060 XT’s idle power as 3.5W. That was the median of 30 samples taken on a different day; this time it’s 3.0W from 60 samples. Think of it as a number that moves within about that range.

How Much Do You Lose by Standardizing on Vulkan?

This time I measured both cards on Vulkan. If the backend isn’t the same, you can’t tell whether a difference comes from the GPU or from the software. But the RTX 3090 also has CUDA, a faster backend provided by NVIDIA.

For how big that gap is, I have a record from a different day, measured on an RTX 3060 connected externally to a different machine. It’s a comparison of the same card and the same models run on CUDA and on Vulkan.

Model	CUDA [tok/s]	Vulkan [tok/s]	Difference
phi4-mini	99.8	93.8	−6.0%
nemotron-3-nano:4b	92.7	88.2	−4.9%

This measurement was done on Windows, with an RTX 3060 connected externally. The environment differs from this desktop PC, so these aren’t values you can add to or subtract from the RTX 3090 numbers in the tables above. Treat them as a way to get a feel for the range: on NVIDIA cards, switching the backend moves things by a few percent.

The AMD side also has a separate backend, ROCm. I haven’t installed it, so I don’t know how the RX 9060 XT’s numbers would move. I also can’t assume that because it was a few percent on NVIDIA, it’ll be a few percent on AMD too. The underlying mechanism is different, so it isn’t the same story.

What I Haven’t Confirmed in This Article

• What happens to the RX 9060 XT’s speed with ROCm installed
• What happens when running a large model that doesn’t fit in 16GB. This time I only measured sizes that fit entirely on both cards
• The RX 9060 XT is in a chipset-connected slot, not one wired directly to the CPU. I haven’t yet separated out what that slows down
• Whether the RTX 3090’s drop on 14B was caused by the power ceiling or by heat
• What happens when both cards are used at the same time

Summary — What I Learned Measuring Two Cards in the Same PC

• Against the RTX 3090 in the same PC, the RX 9060 XT came in at a little over 40% for both read speed and write speed. The gap opened up the same way whichever use you look at
• A 2.4x gap in the short measurement narrows to 1.9x when writing 14B continuously, because the RTX 3090’s speed drops by 20%
• While it was dropping, the RTX 3090 was pinned at 345W. The RX 9060 XT stayed within 1%, holding the same speed to the end at 159W
• For the amount written per ¥1, the RTX 3090 comes out ahead in the short measurement. When writing 14B continuously, they swap places
• Power while doing nothing was 3.0W, about ¥67 a month if left on. The RTX 3090 is about ¥344, but that value is with it handling the display, so the conditions aren’t the same

A fast card doesn’t necessarily stay fast. If you only look at numbers measured over a short time, you can’t see this.

How to Decide Whether the RX 9060 XT Is for You

I’ve split what can be said from these measurements by type of use. Where I didn’t measure something, I say so.

If this is how you use it	What these measurements can tell you
You use models of around 8B to 14B, and electricity cost matters more to you than waiting time	3.0W idle, about ¥67 a month if left on. Write speed is a little over 40% of the RTX 3090, and power is less than half
You have it write long text continuously	Speed didn’t drop even after 85 seconds of continuous writing. When running continuously, the gap narrows to 1.9x
You want answers as fast as possible	The RTX 3090 stays faster. For short exchanges, the gap was 2.4x
You want to run large models that don’t fit in 16GB	Not measured this time. This article can’t help you decide
You want to use AMD’s ROCm	I haven’t installed it, so I don’t know. These are Vulkan values

As a first step, the reliable move is to check the file size of the model you plan to use and confirm first whether it fits in 16GB. If you stay within what fits, the numbers here work directly as a guide.

Next time, I’ll measure what the chipset-connected slot this card sits in slows down, splitting it into model loading and generation.

The figures in this article were measured as of September 2026 with the configuration above. Results will change as GPU generations and software versions change.

Gear Used in This Test

The graphics card measured in this test.