Under the Inference
By Daniel Samson, CO-AUTHORED-BY: GPT 6 SOL · 2026-10-28
I recently bought an Nvidia Tesla V100 32GB on eBay. Not because I think a 2018 card is secretly the fastest way to run local AI, but because I looked at what I actually needed and what it would cost. The Tesla was £600. A Mac mini M5 Pro with 48GB was £2,299. The 32GB Nvidia 5090 I was comparing it with was £6,199.99.
That made the decision rather less complicated.
The comparison I cared about
Here are the numbers that put the purchase in perspective. The FP32 figures are theoretical single-precision performance, and a prediction of how fast a model will generate tokens.
Category | Tesla V100 32GB | Nvidia 5090 32GB | Mac mini M5 Pro 48GB |
|---|---|---|---|
RAM | 32GB HBM2 | 32GB GDDR7 | 48GB unified memory |
Single Precision Performance | Up to 15.7 TFLOPS FP32 | Around 105 TFLOPS FP32 | 8.3 TFlOPs |
Memory Bandwidth GB/S | Up to 900 | Up to 1,792 | up to 307 |
Token Per Second with Qwen3.8:27b @ 256K Context | ~8–15 t/s Extrapolated | ~35–40 t/s measured | ~12–22 t/s Extrapolated |
Cost | £600 | £6,199.99 | £2,299 |
Note: the T/S performance is base on the origional model. However there are fine-tuned versions that may produce 2x performance on the MacMini, and v100.
This isn’t a like-for-like shopping list. The Tesla and 5090 are cards; the Mac mini is a whole computer. The Mac also has more memory, while the 5090 is plainly a much faster GPU on paper. I’m not claiming the V100 wins a benchmark it has no business winning.
Old doesn’t mean useless
The V100 dates back to 2018, which sounds ancient if you follow GPU launches. But the question for me wasn’t whether it could keep pace with a 5090. It was whether I could run qwen3.8:28b on it and get what I wanted out of local inference. It can. For that job, it does just as well for me without requiring the other prices in that table.
There’s a temptation to shop for the biggest number before deciding what problem you’re solving. I’d rather spend £600 on a card that runs the model I want than spend £2,299 or £6,199.99 to find out I didn’t need to. The latest hardware is impressive. It just isn’t automatically the right purchase.