Table of Contents
- Table of Contents
- Phase 3: Blitz
- GPU Breakdown
- A-Tier Dream Scenario (Max Output / Cost)
- B-Tier Dream Scenario
- C-Tier Dream Scenario
- D-Tier Basic Scenario
Phase 3: Blitz
If we’re able to secure GPU’s & additional equipment, we can add exponential scale to both the creation & publication of the content. Here’s our teach leads dream scenario…
GPU Breakdown
We have two tasks: 1) fine-tuning and 2) generation/inference. We'll need both text generation via LLMs and image generation but the bottlenecks are mostly graphical so tailor performance based on image generation. We assume it takes 30 iterations to generate a single image.
Assuming that each additional fine-tuning design+prompt consists of 150 images+text, this is the estimated performance & cost breakdowns.
GPU | Estimated Time to Finetune | Inference Speed | Estimated Cost per Unit |
nVidia H100 80gb pcie | 15 minutes | 27 it/s | $40,000 |
nVidia A100 80gb pcie | 2 hours | 47 it/s | $20,000 |
nVidia RTX 6000 Ada 48gb | >12 hours | 42 it/s | $9,000 |
nVidia RTX 4090 24gb | >16 hours | 33 it/s | $2,000 |
Caveat: RTX 4090's cannot be used in a data center configuration so cannot be scaled up to accommodate actual customers.
A-Tier Dream Scenario (Max Output / Cost)
- 2x nVidia H100 80gb pcie
- 8x nVidia RTX 6000 Ada
- $152k value
- 8-minute fine-tuning times
- 8x concurrent designers
We use dual H100s for redundancy and since they do scale relatively linearly, it would also halve the fine-tuning time. If one dies, we can still do the processing. Any time not used to train/tune can be used for inference.
We use the cost effective RTX 6000 Adas for the generation but not to train.
B-Tier Dream Scenario
- 2x nVidia A100 80gb pcie
- 8x nVidia RTX 6000 Ada 48gb
- $112k value
- 1-hr fine tuning times
- 8x concurrent designers
Same configuration as the S-Tier but with the lower cost A100s.
C-Tier Dream Scenario
- 8x nVidia RTX 6000 Ada 48gb
- $72k value
- 1.5-hr fine tuning times if dedicated
- 8x concurrent designers while not training
We use only the cost-effective inference setup and also train with it. It cannot train while generating and vice versa.
D-Tier Basic Scenario
- 1x nvidia RTX 6000 Ada 48gb
========
- $9k value
- >12-hour fine tuning
- Use personal 3090 homelab to do inference (2-3 concurrent designers)