RunPod Community CloudIn stock
On-Demand from $0.50
From
$0.50 / GPU / hr
On-DemandVisit website →
Rental price index
AMD MI300X
🏢192 GB per GPU, more than any H100 or H200. The capacity play.
Prices collected 2026-09-12 · 2 providers, 2 configurations
Compare vs other GPUs →On-demand
Index median price per GPU per hour, by billing type. Nothing is interpolated.
At a glance
Median on-demand
$1.45/GPU/hr
Cheapest on-demand
$0.50RunPod Community Cloud
RunPod Community CloudFloor, any billing
$0.50incl. spot
Per month at 720h
$1,040
AMD MI300X pricing by provider
One card per provider, cheapest first. A configuration is one priced unit — an instance type, a SKU, a marketplace offer — and the per-GPU price is that unit's price divided by its GPU count. Open a card to see every configuration with its vCPU, RAM and region where the provider publishes them.
2 providers · 2 configurations
RunPod Secure CloudIn stock
USUSA1 config1xcollected 2026-09-12
On-Demand from $2.39
From
$2.39 / GPU / hr
On-DemandVisit website →What fits on a AMD MI300X
Weights + KV cache at 8K context + runtime overhead, against 173 GB usable of the 192 GB on the card. Capacity follows total parameters, even for mixture-of-experts models — any expert may be needed next. Where one card is not enough, the table says how many are, and prices the set at today's median.
| Model | Params | 4-bit | 8-bit | 16-bit |
|---|---|---|---|---|
| 2.8T104B active | 10× GPU$14.45/hr | 18× GPU$26.01/hr | 33× GPU$47.69/hr | |
| 552B16B active | 2× GPU$2.89/hr | 4× GPU$5.78/hr | 7× GPU$10.12/hr | |
 Llama 4 Maverick | 400B17B active | 2× GPU$2.89/hr | 3× GPU$4.34/hr | 5× GPU$7.23/hr |
 gpt-oss-120b | 117B5.1B active | ✓ 67.9 GB~1478 tok/s | ✓ 126.4 GB~782 tok/s | 2× GPU$2.89/hr |
| Llama 3.1 70B | 70B | ✓ 44.5 GB~108 tok/s | ✓ 79.5 GB~57 tok/s | ✓ 145.1 GB~30 tok/s |
 Gemma 4 31B | 31B | ✓ 21.2 GB~243 tok/s | ✓ 36.7 GB~129 tok/s | ✓ 65.7 GB~68 tok/s |
 Qwen3 Coder 30B-A3B | 30.5B3.3B active | ✓ 19 GB~2284 tok/s | ✓ 34.3 GB~1209 tok/s | ✓ 62.9 GB~642 tok/s |
| Qwen3.8 27B | 27B | ✓ 18.7 GB~279 tok/s | ✓ 32.2 GB~148 tok/s | ✓ 57.6 GB~79 tok/s |
| gpt-oss-20b | 21B3.6B active | ✓ 13.7 GB~2094 tok/s | ✓ 24.2 GB~1108 tok/s | ✓ 43.9 GB~589 tok/s |
| Llama 3.1 8B | 8B | ✓ 6.9 GB~942 tok/s | ✓ 10.9 GB~499 tok/s | ✓ 18.4 GB~265 tok/s |
tok/s is decode throughput from memory bandwidth, derated 20%, single stream — an estimate, not a benchmark. Set cost uses today's median × GPUs required.
Specifications
Memory192 GB HBM3
Memory bandwidth5,300 GB/s
ArchitectureCDNA 3
VendorAMD
ClassDatacenter
Released2023-12
Bandwidth is listed because it governs generation speed: each token streams the active weights once, so tokens per second is bandwidth divided by bytes read. TFLOPs decide prompt processing and training, not how fast text appears.
Alternatives
NVIDIA B200180 GB · $6.39/hr median · 15 providersNVIDIA H200141 GB · $4.30/hr median · 16 providersNVIDIA H200 NVL141 GB · $3.72/hr median · 5 providers
Questions
How much does it cost to rent a AMD MI300X?
As of 2026-09-12, the median on-demand price across 2 providers is $1.45 per GPU-hour, about $1,040 a month at 720 hours. The cheapest we saw was $0.50 at RunPod Community Cloud.
Why do prices for the same GPU vary so much?
Hyperscalers bundle CPU, RAM and network into an instance price and charge for the ecosystem around it. Neoclouds sell the GPU more directly. Marketplaces are independent hosts competing on price, with the trade-offs of shared, third-party hardware. The per-GPU figure is the instance price divided by GPU count, which is the only way to compare an 8-GPU node with a single-card listing.
Is the median the price I will pay?
No — it is the middle of the market. Each provider's own median is taken first so a provider with many instance sizes counts once, then the median across providers. Spot and reserved rates are excluded from it; they set the floor shown separately. Use it to judge whether a quote is high or low, then verify with the provider.
Which models can a AMD MI300X run?
With 192 GB per GPU, the table above shows which reference models fit on one card at 4, 8 and 16-bit, using the same memory arithmetic as the rest of Runyard: weights plus KV cache plus runtime overhead, with capacity following total parameters even for mixture-of-experts models. Where a model needs more than one GPU it says how many and what that set costs at today's median.
Own the hardware instead? Work out what your card holds with the same arithmetic.
Check what fits →