32 GB GDDR7 Blackwell. Comfortable for Q4_K_M and Q5 with long context. Needs a 1000 W-class PSU and a case that fits a 3.5-slot card. Verify the 5090 32 GB SKU.
Who it is for
- CUDA users who want Qwen3.8 27B at Q4/Q5 with 128k on one card.
- People already on a high-TGP desktop who rejected 24 GB as too tight.
Who should skip it
- Cost-first homelabs — a used 3090 24 GB is the cheaper Q4 64k path.
- Anyone without a ~1000 W PSU and 3.5-slot clearance.
What to compare
- 32 GB is the CUDA size that covers measured ~26 GB at 128k for this 27B, which a 24 GB card does not.
- Blackwell can use NVFP4 GGUF/runtime paths that Ada/Ampere cannot. Still not a 2.4T Max box.
- Power, length, and 12V-2x6 cabling dominate the build. Confirm PSU and case before the GPU.
Verify on the Amazon listing
- Confirm RTX 5090 32 GB GDDR7, not a 5080 16 GB.
- Confirm seller, board partner, and that the listing is the card, not a bundle SKU.
Honest limitation
This is the expensive CUDA rung. It still will not load Qwen3.8 Max (~2.4T). BF16 27B (~56 GB) needs more than one 32 GB card or unified memory.
Related picks in this catalog
Affiliate disclosure: As an Amazon Associate I earn from qualifying purchases.
Verify live price
Check price on Amazon
Paid link, placed here for readers who compared the notes first.