The hardware behind the work
Clients ask this a lot, so here's the straight answer: most of our GPU work runs on NVIDIA hardware — rented by the hour, on your cloud, or on a box in your server room. But the honest starting point for every project is a cheaper question: does this workload need a GPU at all?
Benchmark first, rent second
GPU hours are the easiest way to burn a project budget, so we made a rule early on: no hardware gets rented until we've profiled the workload on something small. A surprising amount of client work — data pipelines, tabular forecasting, most classical ML — runs perfectly well on an ordinary CPU node, and we'll tell you that even though a GPU invoice would be more profitable for us.
When a workload does earn a GPU, we size it from measurements, not vibes: model memory footprint, batch throughput, and the latency your product actually needs. Fine-tuning a mid-size open model rarely needs the same iron as training one, and inference almost never does.
We don't own a data centre
We're a small studio. Big training runs happen on rented cloud GPUs; day-to-day development happens on workstation RTX cards. You pay for compute you use, not for hardware we're amortising.
No partnership claims
We use NVIDIA hardware and its software stack (CUDA, TensorRT, Triton) because it's what our tooling targets best. We are not an NVIDIA partner or reseller. Applying to or being accepted into NVIDIA Inception does not mean NVIDIA endorses, sponsors, invests in, or certifies Smartgenhub.
The exact card is a detail
GPU generations change fast. What stays constant is the sizing method: measure the workload, pick the smallest hardware that meets the target, and document why.
Which iron for which job
The tiers we actually reach for, and what each one is good at. Exact models depend on what your cloud region or server room offers on the day — this table is about the sizing logic, not a shopping list.
| Tier | Typical NVIDIA hardware | What we use it for | Honest caveat |
|---|---|---|---|
| CPU node | None — plain x86, or Grace-class Arm CPUs where offered | Data pipelines, tabular models, forecasting baselines, anything where a GPU would sit idle | This covers more projects than most vendors admit |
| Workstation | GeForce RTX / RTX-series workstation cards | Prototyping, small fine-tunes, local evaluation runs before anything is rented | Limited VRAM caps model size; fine for development, not production scale |
| Inference GPU | L4, L40S-class cloud instances | Serving vision and NLP models, embeddings, semantic search, batch document extraction | Throughput per dollar is the metric here, not raw speed |
| Training node | A100 / H100-class, newer generations as clouds roll them out | Larger fine-tunes and the rare from-scratch training run, rented by the hour | Most projects never need this tier — we'll show you the benchmark before renting it |
| Superchip node | Grace Hopper / Grace Blackwell-class, where a provider offers them | Workloads bottlenecked on CPU–GPU memory bandwidth rather than compute | Niche; we recommend it only when profiling shows the bottleneck |
| Edge | Jetson modules (Orin-class) | Vision inference next to the camera — factory lines, kiosks, places without reliable connectivity | Models need quantising and pruning to fit; accuracy trade-offs are measured, not assumed |
Product names above belong to NVIDIA. We list them because clients ask what their money rents — not because we resell hardware or hold any certification we haven't mentioned. Everything on this page describes our own working approach; it has not been reviewed or endorsed by NVIDIA.
What our own demos run on
Every demo on our work page has a compute story. Here it is, including the unglamorous parts.
Invoice field extraction
Fine-tuned on a single rented GPU over a weekend. Production inference runs happily on an L4-class card; the human-review queue it feeds is plain CPU.
Search · On-premSemantic docs search
Embedding and reranking models sit on one on-prem GPU so documents never leave the client's network. Indexing is a CPU batch job that runs overnight.
Vision · E-commerceProduct photo tagger
Trained on rented cloud GPUs; serves batches on an inference card. The same model quantises down to a Jetson if tagging ever needs to happen at the warehouse.
ForecastingDemand forecast baseline
Runs on CPU, full stop. The classical methods usually win, and putting gradient boosting on a GPU here would be spending your money to feel modern.
Pipelines · Data qualityData quality watchdog
Pure CPU and cheap by design — it watches every pipeline we ship, so it has to cost almost nothing to keep running around the clock.
MLOps · MonitoringModel serving template
Container-first: the same template deploys to a CPU node, a single GPU, or a multi-GPU box. Triton handles GPU serving where the model earns it.
Where the hardware can live
Three honest options, and the trade-off each one carries.
Your cloud account
We deploy into your AWS, GCP, or Azure project and rent NVIDIA instances there. You see every invoice, keep every artefact, and can revoke our access the day the project ends.
Your server room
For data that can't leave the building, a single GPU server goes surprisingly far. We spec it, set it up, and hand your team the runbook. Semantic search demo runs exactly this way.
At the edge
Jetson modules for inference where the data is born — no round trip to a data centre. The trade-off is model size, and we'll show you the accuracy cost before committing.
Not sure what your workload needs? Send us a description and we'll benchmark it on the smallest hardware that could plausibly work — before you rent anything.
Ask for a Sizing EstimateIf we're approved: how we'd use the cloud credits
We've applied to the NVIDIA Inception program for startups. If Smartgenhub is accepted and cloud credits are included in the benefits available to us, here is exactly how we plan to spend them — the same benchmark-first discipline, just with someone else picking up part of the GPU bill.
The plan, step by step
- Check which benefits and credit offers are actually available to us in the Inception member portal — offers vary and change.
- Create a startup account with the participating cloud provider, or connect an eligible existing account.
- Activate the credits through the provider's redemption instructions or code.
- Verify the fine print before spending anything: credit expiry, eligible regions and services, GPU availability, and usage limits.
- Enable budget alerts, auto-shutdown on idle instances, and usage monitoring from day one — credits burn just like cash.
- Spend the credits on our current Stage 03 prototype work: model fine-tuning, vision and NLP experiments, and inference benchmarking.
- When the credits run out, continue on paid infrastructure only where the benchmarks and the project budget justify it.
Smallest instance first
Free credits don't change the rule: workloads start on the smallest suitable instance, and only benchmark-backed jobs graduate to larger GPUs.
Client compute stays separate
Any credits would fund our own product prototyping. Client workloads keep running in client accounts, on client invoices — that doesn't change.
Not confirmed yet
Acceptance, credit amounts, eligible providers, validity, and redemption conditions are determined by NVIDIA and its participating cloud providers — none of it is guaranteed, and we'll update this page if our status changes.