Build 06Placeholder
Add a used 24GB GPU later
Keep the CPU box, then add one used 24GB GPU if a measured run says the CPU is the limit.
A staged upgrade, not a first purchase. The P40 and the 3090 are the two 24GB classes this page will price when a listing is actually pulled.
- Price
- Placeholder
- Price pulled on
- Placeholder
- Source
- Placeholdername the sold listing or retailer page on the day it is pulled
- Idle
- Placeholder
- Load
- Placeholder
- Model
- Placeholder
- Quant
- Placeholder
- Context
- Placeholder
- Gen tok/s
- Placeholder
- Prompt tok/s
- Placeholder
Wall watts for the whole machine with the card installed, idle and under the timed run. Board-power stickers are not this row.
Same model, quant, and context as the CPU row it is compared with, or the comparison is unfair.
Don’t buy this if
- You have not measured the CPU build yet. The word in the title is later.
- The chassis cannot cool a passive card, and the card you are looking at is a P40.
- You need a monitor plugged into the card. A P40 will not do that.
- You are choosing P40 versus 3090 from a price you heard. This page gets a dated sold price when one is pulled, and not before.
Parts
| Part | Item | Price (USD) | Source | Notes |
|---|---|---|---|---|
| Host | Placeholderthe CPU build this card drops into, with a free slot and a PSU that can feed it | Placeholder | Placeholder | Power connector, slot space, and case airflow are the checklist. The chip name is not. |
| GPU | Placeholderone used 24GB card, P40 class or 3090 class, chosen after the CPU measurement | Placeholder | Placeholder | Record the exact board, not just the chip. Coolers and power plugs differ. |
| Power lead | Placeholderthe cable or adapter the PSU is missing, if any | Placeholder | Placeholder | Adapters are a real line item. A melted adapter is a bad benchmark. |
| Airflow | Placeholdera fan or duct if the card is a passive data-center cooler | Placeholder | Placeholder | A Tesla P40 has no fan of its own. A closed desktop will cook it. |
No affiliate links on this page yet. If one is added, the source cell will say so. See the affiliate disclosure.
Gotchas
- A Tesla P40 is a 24GB data-center card. It has no display output, and the cooler is passive. It wants server-style airflow. Confirm the runtime still supports the card before you pay.
- A 3090 is a 24GB consumer card with its own fans. It is easier to cool and harder on the PSU. Used cards have unknown thermal history. The listing photos matter.
- Compare the GPU run to a CPU run of the same model, quant, and context. A faster number on a smaller quant is a different experiment.
- Driver and runtime versions go in the benchmark row. A speed without the software version cannot be compared with the next one.
This is an upgrade path for the CPU-only Xeon or the 30B coding box. Frankenstein already has a Tesla P40, and the 18 Nov 2026 rows are in the benchmark table: about 6–9× the CPU generation rate on the same Qwen3 Q4_K_M models. The price cells below are still empty.
Twenty-four gigabytes is the shared fact that puts a P40 and a 3090 on the same page. They are not the same purchase. The P40 is a server part: no video output, passive heatsink, software support you have to check against the current runtime. The 3090 is a graphics card that happens to have the same memory size, with its own cooler and a power appetite you meter at the wall rather than copy off a review. Whichever card is bought, the row names that card. “24GB GPU” is not a model string.
The price cells stay placeholders until a specific listing is pulled. Sold prices and asking prices will not share a cell. The source line will say which one it was, and the date sits in “price pulled on.” A forum memory of what these cards cost last year is not a source.
The CPU rows stay next to the P40 rows. The difference is the reason to add a card. Idle watts at the wall are still unmetered, so the power cell on this page stays blank.
Models that point here
A judgment from the model card and the size, not a speed ranking. Memory and the measured-or-not speeds are on each model page.
- Gemma 3 27B · 27B
- gpt-oss-20b · 21B total, 3.6B active