How much VRAM does the AMD Radeon RX 9070 XT have?

The AMD Radeon RX 9070 XT has 16 GB of GDDR6 with 640 GB/s memory bandwidth.

What is the AMD Radeon RX 9070 XT best for?

With 16 GB of VRAM, the AMD Radeon RX 9070 XT handles smaller models (7B–14B) at Q4–Q5 quantization, ideal for entry-level local LLM experimentation and lightweight inference.

What LLMs can the AMD Radeon RX 9070 XT run locally?

The AMD Radeon RX 9070 XT can run 46 of the 99 open-weight models tracked by CanItRun natively in VRAM at 8k context. Top options include: Qwen 3.8 27B at Q3_K_M, Ornith 1.5 35B-A3B (MoE) at Q2_K, Ornith 1.5 9B at Q8_0.

Can the AMD Radeon RX 9070 XT run Gemma 4 31B?

Yes. The AMD Radeon RX 9070 XT runs Gemma 4 31B natively in VRAM at Q2_K quantization, achieving approximately 37 tokens per second.

Can the AMD Radeon RX 9070 XT run Qwen 3.6 27B?

Yes. The AMD Radeon RX 9070 XT runs Qwen 3.6 27B natively in VRAM at Q3_K_M quantization, achieving approximately 35.9 tokens per second.

Can the AMD Radeon RX 9070 XT run Qwen3 8B?

Yes. The AMD Radeon RX 9070 XT runs Qwen3 8B natively in VRAM at Q8_0 quantization, achieving approximately 51.5 tokens per second.

AMD Radeon RX 9070 XT

The AMD Radeon RX 9070 XT has 16 GB VRAM and 640 GB/s memory bandwidth. It can run 46 of our 99 tracked models natively in VRAM at 8k context.

With 16 GB GDDR6, the AMD Radeon RX 9070 XT is a consumer-tier GPU that can run 46 models natively.

The AMD Radeon RX 9070 XT launched March 6, 2025 at a $599 MSRP as the top RDNA 4 gaming card: a full Navi 48 die (53.9 billion transistors, 356.5 mm2) with 64 compute units and 4,096 stream processors, 64MB of Infinity Cache, and 16GB of GDDR6 at 20 Gbps on a 256-bit bus, which is 640 GB/s. Total board power is 304 W. It is the same die as the 32GB Radeon AI PRO R9700 workstation card, so the two share a bandwidth figure and differ in VRAM capacity and price.

AMD Radeon RX 9070 XT: AMD's top RDNA 4 card, priced at $599 against the $749 RTX 5070 Ti and the $429 RTX 5060 Ti 16GB, the two NVIDIA cards that share its 16GB capacity. Its 640 GB/s lands between them: 29% less than the 5070 Ti's 896 GB/s and 43% more than the 5060 Ti's 448 GB/s, so on anything that fits in 16GB it decodes at about 75% of the 5070 Ti's speed and about 1.4 times the 5060 Ti's (Qwen3 14B at Q4_K_M: 48.1, 64.4 and 34.9 tok/s). It is the same Navi 48 die as the 32GB Radeon AI PRO R9700 workstation card, so the consumer version trades half the VRAM for less than half the price: $599 versus $1,299.

46 of this site's 99 tracked models fit natively at 8k context, the same set as every other 16GB card, because capacity rather than bandwidth decides what fits. The strongest are 27B-class dense models at Q3_K_M: Qwen 3.8 27B, Qwen 3.6 27B and UI-Mate 27B each total 15.15 GB against this card's 15.2 GB of usable VRAM and decode at 35.9 tok/s, while Q4_K_M needs 19.02 GB and offloads at 9.8 tok/s. 35B-class mixture-of-experts models such as Ornith 1.5 35B-A3B and the two Qwen 35B-A3B releases fit only at Q2_K (15.12 GB) but decode at 105.0 tok/s, because only about 3B parameters are read per token. Qwen3 14B holds Q4_K_M with room to spare (11.6 GB, 48.1 tok/s), GPT-OSS 20B fits at Q4_K_M (14.55 GB, 116.6 tok/s), and Llama 3.1 8B runs at Q8_0 (10.73 GB, 52.8 tok/s). 70B-class models are out of reach.

AMD's ROCm support tables list the RX 9070 XT (gfx1201) on Linux, limited to Ubuntu 24.04.4 and 22.04.5 or RHEL 10.1 and 9.7, and in the HIP SDK 7.2 table for Windows; RDNA 4 needs a recent ROCm (6.4.1 or newer), since older builds ship no gfx1201 kernels. llama.cpp's Vulkan backend needs no ROCm install at all. On the project's public Vulkan llama-bench scoreboard (Llama 2 7B Q4_0; driver not stated), this card generates 137.1 tok/s and reads a 512-token prompt at 5,036 tok/s. This site's decode model (88% of rated bandwidth plus a fixed cost per layer) predicts about 115 tok/s for that 3.56 GiB file, so the tok/s figures on this page are likely somewhat conservative for small dense models; the scoreboard's prompt-processing speed also shows a compute advantage that the bandwidth-only estimate does not model.

Vendor	AMD
Architecture	RDNA 4
VRAM	16 GB
Memory type	GDDR6
Memory bandwidth	640 GB/s
Compute backend	ROCM
Tier	Consumer
Released	2025
Models (native)	46 / 99
Models (offload)	12 / 99

Software: Supported by AMD's ROCm on Linux and by the HIP SDK on Windows (gfx1201). RDNA 4 needs a recent ROCm (6.4.1 or newer); older builds ship no gfx1201 kernels. llama.cpp's Vulkan backend also works on both operating systems with no ROCm install.

Need more than 16GB? Rent an RTX 3090 from $0.22/hr ↗

Prices as of Oct 13, 2026. Rates change; check RunPod for current pricing. Affiliate link: CanItRun may earn a commission. Doesn't affect the fit calculation above.

Popular models for this GPU

Qwen 3.5 35B-A3B (MoE)Qwen 3.6 35B-A3B (MoE)Ornith 1.5 35B-A3B (MoE)Nemotron 3 Nano 30B Gemma 4 31B

Models this GPU runs natively in VRAM (46)

Show 41 more

Models that fit with CPU offload (12)

These use system RAM for layers that don't fit in VRAM, so expect much slower inference.

Too large for this GPU (41)

Compare AMD Radeon RX 9070 XT with other GPUs

Continue reading

vram-guides13 min

Best LLMs for 16 GB VRAM GPUs (2026)

Frequently asked questions

How much VRAM does the AMD Radeon RX 9070 XT have?: The AMD Radeon RX 9070 XT has 16 GB of GDDR6 with 640 GB/s memory bandwidth.
What is the AMD Radeon RX 9070 XT best for?: With 16 GB of VRAM, the AMD Radeon RX 9070 XT handles smaller models (7B–14B) at Q4–Q5 quantization, ideal for entry-level local LLM experimentation and lightweight inference.
What LLMs can the AMD Radeon RX 9070 XT run locally?: The AMD Radeon RX 9070 XT can run 46 of the 99 open-weight models tracked by CanItRun natively in VRAM at 8k context. Top options include: Qwen 3.8 27B at Q3_K_M, Ornith 1.5 35B-A3B (MoE) at Q2_K, Ornith 1.5 9B at Q8_0.
Can the AMD Radeon RX 9070 XT run Gemma 4 31B?: Yes. The AMD Radeon RX 9070 XT runs Gemma 4 31B natively in VRAM at Q2_K quantization, achieving approximately 37 tokens per second.
Can the AMD Radeon RX 9070 XT run Qwen 3.6 27B?: Yes. The AMD Radeon RX 9070 XT runs Qwen 3.6 27B natively in VRAM at Q3_K_M quantization, achieving approximately 35.9 tokens per second.
Can the AMD Radeon RX 9070 XT run Qwen3 8B?: Yes. The AMD Radeon RX 9070 XT runs Qwen3 8B natively in VRAM at Q8_0 quantization, achieving approximately 51.5 tokens per second.