For LoRA fine-tuning of a 7B parameter model in fp16, VRAM requirements typically land between 10 and 14GB depending on batch size, sequence length, and whether gradient checkpointing is enabled. The RTX 5080 16GB covers this with headroom for a reasonable batch size without gradient checkpointing compromise. For QLoRA 4-bit fine-tuning at 13B scale, VRAM requirements with quantised weights and LoRA adapters typically sit in the 12 to 15GB range — again within 16GB with workable batch sizes. Where 16GB becomes constrained at 13B is in full-precision LoRA without 4-bit quantisation, where the model weights alone approach 26GB in fp16. If your training method requires full precision at 13B or larger, the G6 Cobalt with 32GB VRAM is the correct machine. Call Kevin and describe your model size and training method before ordering if you are uncertain.

For most production Blender scenes, 16GB is well-matched. The VRAM ceiling in Blender is reached when the total texture memory, geometry data, and Cycles rendering buffers exceed available VRAM — at which point Blender falls back to CPU rendering, which is significantly slower. 16GB covers scenes with 8K textures, high-polygon geometry, displacement, and HDRIs that represent most architectural visualisation and product rendering environments. Where 16GB becomes the constraint is in extreme production environments with many 8K or 16K texture sets loaded simultaneously, or in very large scene caches with hundreds of instanced high-poly objects. If your Blender scenes have been causing VRAM overflow errors on a 12GB card and you are aware of the specific scene complexity involved, 16GB resolves most of those cases. Call Kevin and describe the scene type and the error you are hitting for a specific answer.

The Titanium uses an RTX 5080 with 16GB VRAM and 64GB system RAM. The G6 Cobalt uses an RTX 5090 with 32GB VRAM and 64GB system RAM. The practical difference is in the workloads that require more than 16GB VRAM: full-precision LoRA fine-tuning at 13B parameter scale, Blender scenes with extreme texture and geometry requirements, and multi-stream compositing in Resolve Fusion or Nuke. For researchers who have confirmed they need 32GB VRAM for their specific training method, or 3D artists who know their scenes consistently exceed 16GB, the Cobalt is the correct machine. For most researchers and 3D artists who have outgrown 12GB and need to reach 16GB, the Titanium resolves the immediate constraint at a lower price point.

Yes, for most professional architectural GPU rendering workflows. In V-Ray GPU and Redshift, the RTX 5080's raw throughput and 16GB VRAM handle complex architectural scenes — exterior environments with large terrain, interior spaces with multiple light sources and 4K to 8K material textures, and product visualisation at production quality — without requiring scene simplification or texture downscaling for most projects. Where the RTX 5090 32GB becomes relevant for architecture is in extremely large scenes where the total texture footprint exceeds 16GB, or in studios running multiple simultaneous renders in a single GPU session. For most architectural visualisation studios and individual practitioners, the RTX 5080 16GB provides the throughput and VRAM needed for commercial production work at this specification.

Build time is 3 to 5 working days from order confirmation, including the 24-hour stress test at sustained GPU training load. Delivery to UK mainland addresses is free and fully tracked. Call Kevin on 01902 714533 before ordering if you have a specific date in mind — for example, before a research project start date or a production deadline.