Model weights
Parameters × bytes per parameter: 2 bytes at FP16, 1 byte at FP8 and 0.5 byte at INT4.
Platform · Capacity planning
GPU sizing estimator
Estimate the GPU memory and node count your model needs before you ask for a quote.
A transparent planning tool for H100 and L40S capacity. It does not invent a price; it produces a configuration you can take into a scoping conversation.
REF · SC-2026-SIZE
GPU · H100 80GB / L40S 48GB Output · Memory + node plan Pricing · Quote-based Residency · Malaysia
Your workload
REF · SC-2026-SIZEPlanning estimate
Estimated GPU memory
— GB
- Model weights
- —
- KV cache
- —
- Runtime / activation allowance
- —
Recommended GPUs
—
Node plan
—
Planning estimate only. Actual memory depends on architecture, quantisation, batch size, sequence length and serving framework.
Send this configuration
Opens your email app with the configuration pre-filled. A solution architect will confirm feasibility and pricing.
Method
How the estimate is calculated
Three transparent terms. No benchmark multiplier, no synthetic comparison against another provider.
KV cache
A per-model allowance × context window × concurrent requests. Inference only; training uses activation memory instead.
Overhead and node plan
A workload-specific allowance for runtime, activations, optimiser state and fragmentation, then GPU count at 85% usable memory.
Limits
What the estimate does not include
These are the variables that move a real deployment, and the reason the output is a starting point rather than a quote.
- Model architecture effects such as grouped-query or multi-head latent attention, and mixture-of-experts routing.
- Serving framework behaviour, batching policy, tensor-parallel layout and CUDA graph memory.
- Storage throughput, dataset size, checkpoint frequency and network fabric requirements.
- Redundancy beyond the high-availability toggle, disaster recovery and cross-site replication.
Pricing
Why we quote instead of publishing a price
Two workloads with the same GPU count can have very different commercial shapes: one runs for a week, another for three years; one needs a dedicated node and private connectivity, another is an on-demand burst. A published list price would either overcharge the first or under-deliver to the second. We itemise the quote against the actual configuration, term and support model.
FAQ
Estimator questions
Is this estimate a quote?
No. It is a planning estimate for GPU memory and node count. Pricing, availability and the final configuration are confirmed in an itemised quote after a scoping conversation.
Why does the estimator not show a monthly price?
Because the rate depends on the GPU shape, commitment length, storage profile, connectivity and support model. A single list price would not survive those variables, so we quote against the configuration you actually need.
Does the estimate account for quantisation quality?
It accounts for the memory reduction from FP16, FP8 and INT4, but not for the quality trade-off of a particular quantisation method. That trade-off should be evaluated against your own corpus before production.
Can the estimate cover multi-node training?
The node plan rounds up to whole 8-GPU nodes when the memory requirement exceeds a single node. Multi-node interconnect and scheduling requirements are confirmed during the scoping review.
Platform · Capacity planning
Take the configuration to a scoping call
Send the estimate with your model, data class and target timeline. We will confirm the GPU shape, availability window and an itemised quote.