Automatic GPU sizing on AMD Developer Cloud

Stop paying full MI300X
prices for small jobs.

NeuralGrid profiles every job you submit and sizes it to the smallest AMD Instinct partition that can actually run it — automatically, on AMD Developer Cloud. Same output. Lower bill.

40% average cost reduction vs always requesting a full MI300X node

Get started
Compute Estimatorlive
Job in: llama-3-8b · inference
T1
T2
T3
VRAM
8.5 GB
Confidence
HIGH
Routed toMI210 partition · $0.0023

The problem

You're overpaying for GPU compute

Most teams default to the biggest GPU available. Here's what that actually costs.

TaskTier neededTier typically usedWaste factor
LLM inference, 7B modelT1 · MI210 partitionT3 · Full MI300X5–10×
Image generation (SDXL-class)T2 · MI300X partitionT3 · Full MI300X3–5×
Fine-tune, small LLMT2 · MI300X partitionT3 · Full MI300X3–4×
Audio generationT1 · MI210 partitionT3 · Full MI300X5–8×

How it works

One API call, cheapest matching GPU

  1. 1

    Submit your job

    Send a prompt, image, or audio job through one API call.

  2. 2

    Compute Estimator classifies it

    VRAM and tier requirements are calculated automatically.

  3. 3

    Tier is selected

    The cheapest GPU tier that can handle the job is chosen.

  4. 4

    Routed to a GPU provider

    Your job runs on the matched node and returns a result.

Platform spotlight

AMD MI300X

192GB HBM3

VRAM per accelerator

Built on AMD

NeuralGrid runs exclusively on AMD Developer Cloud.

A single MI300X carries enough VRAM to run 70B+ models on one node — and enough headroom to partition into smaller instances for lighter jobs. NeuralGrid's scheduler packs jobs onto the smallest AMD Instinct partition that fits, across MI210 and MI300X capacity, entirely within AMD Developer Cloud.

How NeuralGrid compares

None of the existing GPU marketplaces automatically route jobs to the cheapest sufficient tier.

PlatformAutomatic tier routing
NeuralGrid✓

GPU tiers and pricing

Every job is automatically matched to one of three tiers based on what it actually needs.

T1 — Lite
VRAM range
0–16GB
Hardware
AMD Instinct MI210 (partitioned)
Price range
$0.04–0.08/hr
Example workload
Small LLM inference (Llama-3-8B)
T2 — Standard
VRAM range
16–64GB
Hardware
AMD Instinct MI300X (partitioned)
Price range
$0.18–0.35/hr
Example workload
SDXL-class image generation
T3 — Power
VRAM range
64GB+
Hardware
AMD Instinct MI300X (full node, 192GB)
Price range
$0.60–1.10/hr
Example workload
Llama-3-70B inference