Arc Pro B70 32GB + 64GB RAM: LTX-2.3 22B, FP8/INT8 Support and Tested ComfyUI Workflow

Request for Intel Arc Pro B70 Support Guidance for LTX-2.3 22B, Including FP8 or Quantized Models

Hello Intel AI/llm-scaler team,

I have a home AI server with the following configuration:

Hardware and operating system

- GPU: Intel Arc Pro B70
- GPU memory: 32GB VRAM
- System memory: 64GB physical RAM
- Operating system: Ubuntu 26.04
- Intended application: ComfyUI
- Intended model: LTX-2.3 22B
- Intended use cases:
- Text-to-video
- Image-to-video
- Synchronized video and audio generation
- Video VAE decoding
- Audio decoding
- Spatial upscaling
- LoRA-based or distilled workflows where supported

I understand that Intel "llm-scaler" currently provides optimized AI and ComfyUI environments for Intel Arc Pro GPUs, including the Arc Pro B60 and B70.

However, I have not found clear confirmation that the complete LTX-2.3 22B pipeline has been tested successfully on a single Arc Pro B70.

I would appreciate clarification on the following questions.

1. Current LTX-2.3 support status

Is LTX-2.3 22B currently supported on a single Intel Arc Pro B70 with 32GB VRAM?

Please clarify whether the current Intel implementation supports:

- LTX-2.3 22B text-to-video
- LTX-2.3 22B image-to-video
- Synchronized audio generation
- Gemma 3 text encoding
- Video VAE decoding
- Audio VAE or audio decoder
- Spatial upscaling
- Distilled LTX-2.3 checkpoints
- LTX-2.3 LoRA models
- Long video generation or multi-stage generation

If LTX-2.3 is not yet fully supported, which parts of the pipeline currently work and which parts are known to fail?

Is the existing Intel support limited to an older LTX-2 or LTX-2 19B implementation?

2. FP8 model support

If the standard BF16 or FP16 version of LTX-2.3 22B cannot fit into 32GB VRAM, is there an FP8 version that can run on the Intel Arc Pro B70?

Please clarify the following:

1. Can the official LTX-2.3 FP8 checkpoint be loaded directly on Intel XPU?

2. Does Intel Arc Pro B70 support the required FP8 operations for this model through:

- PyTorch XPU
- oneDNN
- Intel Extension for PyTorch
- Intel Triton
- oneAPI
- llm-scaler

3. Is FP8 execution actually accelerated by the B70 XMX engines, or is the FP8 checkpoint internally converted to BF16 or FP16 during inference?

4. Does the FP8 version reduce only VRAM usage, or does it also improve generation speed?

5. Is FP8 supported for all major LTX-2.3 components, or only for the main transformer?

6. Can the following components use different precision settings independently?

- LTX-2.3 transformer
- Gemma 3 text encoder
- Video VAE
- Audio model
- Spatial upscaler

7. Is it recommended to run:

- Transformer in FP8
- Text encoder in BF16
- VAE in BF16 or FP32
- Audio components in BF16

8. Are there any known accuracy, stability, visual quality or audio quality problems when using FP8 on the Arc Pro B70?

9. Does the official FP8 checkpoint use a format compatible with Intel XPU, or does it contain NVIDIA-specific FP8 scaling or CUDA-specific kernels?

10. If direct FP8 loading is not supported, is there an official Intel conversion script for converting LTX-2.3 weights into an Intel-compatible FP8 format?

3. Other quantized model options

If FP8 is not currently supported or stable, are there other recommended quantized versions for the Arc Pro B70?

For example:

- INT8
- Weight-only INT8
- INT4
- BF16 with layer-by-layer CPU offloading
- Mixed BF16 and FP8
- Mixed BF16 and INT8
- GGUF or another converted format
- Quanto
- TorchAO
- Optimum Intel
- OpenVINO
- Intel Neural Compressor
- IPEX weight-only quantization

Which quantization method has actually been tested with LTX-2.3 on the Arc Pro B70?

Please provide the exact model format and quantization method that you recommend.

I would also like to know whether quantized LTX-2.3 models retain support for:

- Image-to-video conditioning
- Audio generation
- LoRA loading
- Spatial upscaling
- ComfyUI model patching
- CPU offloading
- Attention optimization

4. VRAM and system RAM requirements

The server has 32GB GPU memory and 64GB physical system memory.

Can the complete LTX-2.3 pipeline run with this configuration?

Please estimate the VRAM and system RAM requirements for the following configurations:

Configuration A

- Main transformer: BF16
- Text encoder: BF16
- VAE: BF16
- No CPU offloading

Configuration B

- Main transformer: FP8
- Text encoder: BF16
- VAE: BF16
- Partial CPU offloading

Configuration C

- Main transformer: INT8 or weight-only quantization
- Text encoder: BF16
- VAE: BF16
- Partial CPU offloading

Configuration D

- Main transformer: BF16
- Aggressive layer-by-layer CPU offloading
- 64GB system RAM

For a system with 64GB RAM, please clarify:

- Whether 64GB RAM is sufficient
- Whether 96GB or 128GB RAM is recommended
- How much swap space should be configured
- Whether swap on an NVMe SSD is acceptable
- Whether model loading may temporarily exceed the steady-state RAM usage
- Whether the system may run out of memory during VAE decoding or spatial upscaling
- Whether pinned memory or shared memory settings must be adjusted
- Whether Docker shared-memory size must be increased

If 64GB RAM is insufficient for the full pipeline, which specific component causes the highest system memory usage?

5. Expected performance

If LTX-2.3 can run on the Arc Pro B70, could you provide approximate tested performance numbers?

For example:

- Resolution
- Number of frames
- Video duration
- Number of inference steps
- Model precision
- Peak VRAM usage
- Peak system RAM usage
- Time required for text encoding
- Time required for diffusion or transformer inference
- Time required for VAE decoding
- Time required for audio generation
- Total generation time

Please indicate whether these results were measured with:

- Standard LTX-2.3 22B
- Distilled LTX-2.3
- FP8 LTX-2.3
- INT8 LTX-2.3
- CPU offloading enabled
- Spatial upscaling enabled

6. Required Intel software environment

Please provide the exact recommended software versions for the Arc Pro B70:

- Supported Ubuntu version
- Recommended Linux kernel version
- Intel GPU kernel driver
- Intel compute runtime
- Level Zero
- Intel GPU firmware
- oneAPI runtime
- oneDNN
- Intel Extension for PyTorch
- PyTorch XPU
- Intel Triton
- Python
- ComfyUI
- llm-scaler
- llm-scaler-omni Docker image tag

Is Ubuntu 26.04 currently supported?

If Ubuntu 26.04 has not yet been validated, should I use Ubuntu 24.04 LTS instead?

Please also clarify whether the B70 requires the upstream Linux "xe" driver or another specific Intel GPU driver configuration.

7. Recommended installation method

Which installation method is recommended?

Option 1

Intel "llm-scaler-omni" Docker container

Option 2

Native Ubuntu installation with PyTorch XPU

Option 3

Intel Extension for PyTorch environment

Option 4

OpenVINO-based inference

Option 5

A dedicated Intel ComfyUI container or installation script

Please provide the exact commands or official documentation for the recommended method.

If Docker is recommended, please provide:

- Exact container image
- Exact image tag
- Required device mappings
- Required group permissions
- Required environment variables
- Shared-memory setting
- Volume mappings
- Recommended model directory layout
- ComfyUI launch arguments
- XPU memory allocation settings

8. ComfyUI compatibility

Does LTX-2.3 require Intel-specific ComfyUI nodes or patches?

Please clarify whether I should use:

- Native ComfyUI LTXVideo nodes
- Lightricks ComfyUI-LTXVideo custom nodes
- An Intel-maintained fork
- An llm-scaler-specific ComfyUI build
- Intel-specific model-management patches
- Intel-specific attention implementations
- Intel-specific VAE patches

Is there an official workflow JSON that has been tested on the Arc Pro B70?

Ideally, I need tested workflows for:

- Text-to-video
- Image-to-video
- Video with synchronized audio
- Distilled LTX-2.3
- FP8 or quantized LTX-2.3
- Spatial upscaling

9. Known XPU problems

Are there currently any known issues involving:

- FP16 text encoding
- Gemma 3 token generation
- XPU kernel assertions
- BF16 attention
- FP8 matrix multiplication
- SDPA
- Intel Triton attention
- VAE decoding
- Audio decoding
- Layer normalization
- Rotary positional embeddings
- Quantized linear layers
- ComfyUI model offloading
- XPU memory fragmentation
- Host-to-device transfer
- Multi-stage model loading
- LoRA application

I have seen reports suggesting that FP16 text encoding may crash on Intel XPU and that BF16 or FP32 may be required.

Is this issue still present with the latest recommended Intel software stack?

10. Fallback solution

If the complete LTX-2.3 22B pipeline cannot currently run on the Arc Pro B70, what is the best officially supported alternative?

For example:

- Older LTX-2 19B
- Distilled LTX-2
- FP8 LTX-2.3 without audio
- Quantized LTX-2.3 with CPU offloading
- Lower-resolution LTX-2.3 workflow
- OpenVINO-converted LTX model
- Another Intel-optimized video-generation model

I would prefer an option that supports image-to-video generation and can use the B70’s 32GB VRAM effectively.

Requested information

Could you please provide:

1. A clear statement on whether LTX-2.3 22B is currently supported on the Arc Pro B70.

2. The exact model checkpoint or quantized checkpoint that should be used.

3. Confirmation of whether FP8 models work correctly on Intel XPU.

4. The exact driver and software versions.

5. The recommended Docker image and tag.

6. A tested ComfyUI workflow JSON.

7. Recommended precision settings for every model component.

8. Expected VRAM and system RAM requirements.

9. Recommended swap size.

10. Any required patches, environment variables or launch parameters.

11. Known limitations or unsupported features.

12. An alternative supported workflow if full LTX-2.3 support is not yet available.

Thank you.