Updated
Spheron AI GPU Catalog

AMD MI300X GPU: 192GB HBM3 Specs, Pricing & Rental. Rent MI300X GPU from $3.59/hr

192GB HBM3 per GPU, 5.3 TB/s bandwidth, ROCm on Ubuntu. Per-minute billing on MI300X GPU rentals after a 20-minute minimum (60 on 2x), in 1x and 2x instances, with larger configurations on request.

No long contracts · Per-minute billing after the minimum runtime · 1x and 2x self-serve, larger on request

At a glance

Renting an AMD MI300X on Spheron AI costs $3.59/hr per GPU on-demand, in 1x and 2x instances you launch yourself from the same console as the NVIDIA GPUs, with no quota request or sales call. Larger MI300X configurations, including 8-GPU nodes, are available as reserved capacity on request. Each GPU has 192 GB of HBM3 at 5.3 TB/s, enough to serve a 70B model in BF16 on one card. The stack is ROCm, not CUDA. Billing is per minute after a minimum runtime of 20 minutes on 1x and 60 minutes on 2x, and instances bill until you destroy them. No MI300X spot tier is listed on Spheron AI right now. Runpod lists MI300X at $2.39/hr, but Runpod's own console showed it Unavailable on 17 Sep 2026, and a listed price is not a GPU you can launch. In an August 2026 survey of published on-demand rates, DigitalOcean listed MI300X at $2.59/hr and Azure ND MI300X v5 started at $6.00 per GPU-hour, sold only as 8-GPU VMs.

Dedicated$3.59/hr
GPU ArchitectureAMD CDNA 3
VRAM192 GB HBM3
Memory Bandwidth5.3 TB/s

AMD MI300X pricing

ProviderPrice/hrvs Spheron AIAvailability
Spheron AIYour price
$3.59/hr-Bookable now
Runpod
$2.39/hr33% cheaperUnavailable
DigitalOcean
$2.59/hr28% cheaperListed rate
Crusoe
$3.45/hr4% cheaper8x nodes only
Azure ND MI300X v5
$6.00/hr1.7x more expensive8x nodes only
Oracle BM.GPU.MI300X.8
$6.00/hr1.7x more expensive8x nodes only

Spheron AI rates are live marketplace prices and move with partner supply, so what is shown is what you can book right now. Competitor figures are published on-demand rates taken from each provider's own pricing page. Each is measured against whichever of our own rates it actually sits against: our on-demand rate when it lands at or above it, and our spot rate when it falls below, because spot is preemptible and not the same product. A listed rate is not the same thing as available capacity: check before you plan a run. Where a competitor row carries a capacity note, that is what that provider's own console reported on 17 Sep 2026. It is a dated observation, not a live feed, and their supply may have changed since.

AMD MI300X specifications

GPU Architecture
AMD CDNA 3
VRAM
192 GB HBM3
Memory Bandwidth
5.3 TB/s
FP16 Tensor (dense)
1,307 TFLOPS
FP8 Tensor (dense)
2,615 TFLOPS
FP32 Performance
163.4 TFLOPS
Stream Processors
19,456
Process Node
TSMC 5nm / 6nm
Form Factor
OAM
TDP
750W
Software Stack
ROCm (Ubuntu image)
System RAM
241 GB (1x), 481 GB (2x)
vCPUs
8 or 13 (1x), 26 (2x)
Storage
13,194 GB NVMe
Region
Michigan, United States
Custom & Reserved

Need More MI300X Than What's Listed?

Reserved Capacity

Commit to a duration, lock in availability and better rates

Custom Clusters

8 to 512+ GPUs, specific hardware, InfiniBand configs on request

Supplier Matchmaking

Spheron AI sources from its certified data center network, negotiates pricing, handles setup

Need more MI300X capacity? Tell us your requirements and we'll source it from our certified data center network.

Typical turnaround: 24–48 hours

When to pick the MI300X

Scenario 01

Pick the MI300X if

Your model does not fit where you want it. A 70B model in BF16 needs about 141GB for weights, which means two H100s, or FP8 on an H200. One MI300X holds it with roughly 30GB to spare, and a 2x instance holds Mixtral 8x22B in BF16. Your stack also has to be PyTorch, vLLM or SGLang, because those are what run cleanly on ROCm.

Recommended fit
Scenario 02

Pick the H200 or H100 instead if

The model fits in 141GB or 80GB with the KV cache you need. Hopper on CUDA is faster per GPU than MI300X on ROCm today: SemiAnalysis measured MI300X about 14% behind H100 on BF16 matrix math, and our own 8B LoRA job trained in about 15 minutes on H100 against 20 on MI300X. If memory is not the bottleneck, the extra capacity buys you nothing.

Recommended fit
Scenario 03

Pick the B200 instead if

You want 192GB per GPU and the most throughput CUDA can give you. B200 carries the same capacity on HBM3e with Blackwell compute and native FP4, and it runs TensorRT-LLM. MI300X is the pick when your stack already runs on ROCm and one 192GB card is the requirement, not peak tokens per second.

Recommended fit
Scenario 04

Stay on CUDA if

Your serving or training path depends on TensorRT-LLM, FlashAttention 3's CUDA kernels, NVIDIA NIM, or custom CUDA kernels you have not ported with HIPIFY. For multi-node training, 1x and 2x are what you launch yourself on Spheron AI, and larger MI300X configurations are arranged as reserved capacity on request. Test the port on a 1x instance before you move production traffic.

Recommended fit

AMD MI300X use cases

Use case / 01
🧠

70B-class inference on one GPU

At BF16, Llama 3.1 70B needs about 141GB for weights. One MI300X holds that with roughly 30GB left for KV cache, so you serve the full-precision model with no tensor parallelism and no quantization step.

Llama 3.1 70B Instruct in BF16 on a 1x instanceQwen2.5 72B served without FP8 or INT4Mixtral 8x7B at FP16 with room for large batchesMixtral 8x22B in BF16 across a 2x instance
Use case / 02
📜

Long-context and high-concurrency serving

When the model already fits, the spare memory goes to KV cache. A 13B model in BF16 leaves over 140GB for cache on one card, which means longer contexts and more concurrent sequences before the scheduler starts preempting.

Long-document chat on 13B to 32B modelsRAG backends with large retrieved contextBatch summarization at high sequence countsAgent backends holding many open sessions
Use case / 03
🎯

Fine-tuning with memory headroom

LoRA on a 70B base in BF16 fits on one card at small batch sizes with activation checkpointing, and full fine-tunes of 7B-class models run without sharding optimizer state. Expect to tune batch size: our Llama 3.1 8B LoRA run needed batch 2 on ROCm where H100 ran batch 4.

Llama 3.1 8B LoRA with torchtune, one config change from CUDALoRA on 70B BF16 bases with activation checkpointingFull fine-tunes of 7B-class models on one GPUDPO with LoRA adapters on 7B to 13B models
Use case / 04
🧪

ROCm evaluation before a bigger AMD commitment

A 1x instance with a 20-minute minimum is a cheap way to find out whether your stack runs on ROCm before you reserve a larger AMD footprint or port a production service.

Run your vLLM or SGLang serving config on ROCmPort and test custom kernels with HIPIFYCompare tokens per dollar against your H100 or H200 fleetCheck numerical parity between CUDA and ROCm outputs

AMD MI300X benchmarks

Llama 3.1 70B BF16 weights
~141 GB on 1 GPU
needs 2x H100 80GB, or FP8 on one H200
VRAM capacity
192 GB HBM3
2.4x H100 (80 GB), 1.36x H200 (141 GB)
Memory bandwidth
5.3 TB/s
vs 4.8 TB/s on H200 and 3.35 TB/s on H100
Llama 3.1 8B LoRA, 375 steps (training)
~20 min
vs ~15 min on H100 SXM5 in our 13 Sep 2026 run
Realized BF16 GEMM
~620 TFLOP/s
about 14% behind H100's ~720 in SemiAnalysis tests
Realized FP8 GEMM
~990 TFLOP/s
about 22% behind H100's ~1,280 in SemiAnalysis tests

Serve Llama 3.1 70B in BF16 on one MI300X with vLLM

One MI300X holds the roughly 141GB of BF16 weights for Llama 3.1 70B with room left for KV cache, so there is no tensor parallelism and no quantization step. The commands use AMD's prebuilt vLLM image and the device flags from AMD's ROCm documentation.

bash
Spheron AI
# 1. On the instance, check which ROCm release the host image shipscat /opt/rocm/.info/version # 2. Pull AMD's vLLM image for Instinct GPUs. Pick the tag from AMD's ROCm#    vLLM docs that matches the host release; this is one AMD listed on 17 Sep 2026.IMAGE=rocm/vllm:rocm7.14.1_cdna_ubuntu24.04_py3.14_pytorch_2.11_vllm_0.23.0docker pull $IMAGE # 3. Serve Llama 3.1 70B Instruct in BF16 on one MI300X#    (HF_TOKEN needs approved access to the Llama 3.1 weights)docker run -it --rm \  --device /dev/kfd --device /dev/dri \  --group-add=video --ipc=host --network=host \  --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \  -v ~/.cache/huggingface:/root/.cache/huggingface \  -e HF_TOKEN=$HF_TOKEN \  $IMAGE \  vllm serve meta-llama/Llama-3.1-70B-Instruct \    --dtype bfloat16 \    --max-model-len 32768 \    --gpu-memory-utilization 0.90 \    --port 8000 # 4. From a second shell, hit the OpenAI-compatible endpointcurl http://localhost:8000/v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "meta-llama/Llama-3.1-70B-Instruct",    "messages": [{"role": "user", "content": "Why does a 70B model fit on one MI300X?"}]  }'

For models too large for one card, such as Mixtral 8x22B in BF16, rent a 2x instance and add --tensor-parallel-size 2. If the instance has Podman rather than Docker, the same flags work with podman run.

MI300X vs alternatives

MI300X Guides & Benchmarks

More GPU Selection guides →
01Technical Brief

AMD MI300X Release Date and Cloud Availability

The AMD MI300X release date is 6 December 2023, when AMD launched it alongside ROCm 6, with Microsoft already announcing the Azure ND MI300X v5 series and Oracle planning bare metal MI300X instances. Azure's ND MI300X v5 VMs reached general availability on 21 May 2024, with Hugging Face as the first customer. The MI300X is now listed across several clouds, from 8-GPU nodes on Azure, Oracle and Crusoe to on-demand listings at DigitalOcean and Runpod, and our MI300X pricing survey tracks the published rates.

A listing is not the same thing as capacity: Runpod's own console showed MI300X as Unavailable on 17 Sep 2026. Spheron AI added the MI300X on 10 Sep 2026 as its first AMD GPU. It rents as a virtual machine in 1x and 2x instances in Michigan, on a fixed Ubuntu image with ROCm, and larger MI300X configurations including 8-GPU nodes are available as reserved capacity on request. No MI300X spot tier is listed on Spheron AI right now. The live rate sits next to every NVIDIA GPU on the pricing page.

02Technical Brief

MI300X VRAM and Memory Bandwidth: 192GB HBM3

MI300X VRAM is 192 GB of HBM3 per GPU, with a peak memory bandwidth of 5.3 TB/s through an 8,192-bit memory interface. That is 2.4x the 80 GB on an H100 and 51 GB more than the H200's 141 GB. The memory generation is HBM3, not HBM3e: the H200 and B200 use HBM3e, yet the MI300X still posts higher bandwidth than the H200's 4.8 TB/s. For batch-1 decode on a 70B BF16 model, which reads about 141 GB of weights per token, that bandwidth sets a theoretical ceiling near 38 tokens per second, against about 34 on H200. Real throughput lands below the ceiling on either card.

Capacity is what 192 GB buys. Llama 3.1 70B in BF16 fits on one card with about 30 GB left for KV cache, and a 2x instance's 384 GB holds Mixtral 8x22B in BF16 with room for batching. It does not buy compute: in SemiAnalysis's independent GEMM benchmarks the MI300X realized about 620 TFLOP/s in BF16 against about 720 on H100. Rent the MI300X when memory is the bottleneck, and read the MI300X vs H200 comparison if you are choosing between the two.

FAQ / 15

AMD MI300X FAQ

On Spheron AI the AMD MI300X costs $3.59/hr per GPU on-demand, and a 2x instance bills both GPUs at that rate. Billing is per minute after a minimum runtime of 20 minutes on 1x and 60 minutes on 2x, with no contract. No MI300X spot tier is listed on Spheron AI right now. The rate is a live marketplace price, so check it before a long run.

No MI300X spot tier is listed on Spheron AI right now, so $3.59/hr on-demand is the Spheron AI rate to plan around. For reference, Runpod lists MI300X at $2.39/hr, but Runpod's own Pod console showed it as Unavailable on 17 Sep 2026, so that rate could not be deployed that day. DigitalOcean listed MI300X at $2.59/hr on-demand in an August 2026 survey of published rates; its capacity was not checked. On Spheron AI, keep the bill down by using a 1x instance (a 20-minute minimum instead of 60), starting the job as soon as the instance boots, and destroying it the moment the job ends, since there is no stopped state.

Yes. A 1x MI300X instance has a 20-minute minimum runtime and a 2x instance has a 60-minute minimum. The minimum is charged in full even if you destroy the instance earlier, and billing is per minute after it. The deploy wizard shows the minimum before you launch.

No. MI300X instances can be restarted but not stopped or paused, and an instance bills at the full hourly rate until you destroy it. Attached volumes are not available on these instances either, so copy checkpoints and results to your own storage before you destroy one. Destroying the instance is the only way to stop billing.

Every MI300X instance boots one fixed image, Ubuntu with ROCm. There is no image picker, and we do not pin a ROCm version on this page, so check it on the machine with cat /opt/rocm/.info/version. The practical approach is to run your framework in a container that brings its own ROCm userspace, such as AMD's rocm/vllm or rocm/pytorch images, and pick the tag that matches the host's ROCm release.

Most PyTorch code will, without changes. ROCm builds of PyTorch expose AMD GPUs through the torch.cuda API, so device="cuda" works. What will not run: TensorRT-LLM, FlashAttention 3's CUDA kernels, NVIDIA NIM containers, and any hand-written CUDA kernel until you port it with AMD's HIPIFY tools. Test your exact stack on a 1x instance before you commit a workload to it.

Yes. vLLM lists the MI300X among its supported AMD GPUs, and AMD publishes prebuilt rocm/vllm images for Instinct cards. PyTorch ships official ROCm builds, and SGLang supports AMD GPUs on ROCm. We ran torchtune's Llama 3.1 8B LoRA recipe on a rented MI300X with the same recipe and framework code we used on H100. Two things changed: the container image, and one line of the config. Batch size 4 hit an out-of-memory error on ROCm, so that run used batch size 2 with gradient accumulation 8, the same effective batch of 16 as the H100 run.

At BF16, a 70B model such as Llama 3.1 70B needs about 141GB for weights, so it fits on one MI300X with roughly 30GB left for KV cache at 90% memory utilization. Mixtral 8x7B at FP16, about 93GB, fits with room for large batches. A 2x instance gives you 384GB, enough for Mixtral 8x22B in BF16 with tensor parallelism across both GPUs. A 405B model does not fit on 2x at 8-bit; it needs 4-bit weights, so check that your quantization kernels have a ROCm build first.

Rent the MI300X when a model needs more than 141GB on one GPU: it has 192GB of HBM3 at 5.3 TB/s against the H200's 141GB of HBM3e at 4.8 TB/s. Rent the H200 when the model fits in 141GB and you want CUDA, TensorRT-LLM and FP8 through Transformer Engine, which usually means lower latency at small batch sizes. Spheron AI rents both, so you can run the same model on each and compare.

For jobs that fit in 80GB, the H100. It is faster per GPU on today's software: in our same-recipe Llama 3.1 8B LoRA run, training took about 15 minutes on H100 and about 20 on MI300X, and SemiAnalysis measured MI300X about 14% behind H100 on BF16 matrix math. The MI300X wins when a model needs more than 80GB, because one 192GB card replaces a tensor-parallel pair of H100s.

Read the H100 vs MI300X benchmark

MI300X instances on Spheron AI run in one region: Michigan, United States. There is no region choice for this GPU, so plan for US Midwest latency if you serve users on other continents.

Yes, as reserved capacity. What you launch yourself on Spheron AI is a 1x or 2x MI300X virtual machine. Larger MI300X configurations, including 8-GPU nodes, are available on request: use the Request a Quote form on this page and the Spheron AI team arranges it, so it is quoted rather than instantly launchable. Azure ND MI300X v5, Oracle BM.GPU.MI300X.8 and Crusoe also sell 8-GPU MI300X nodes self-serve.

$48 an hour at the lowest published rates. Azure ND MI300X v5 and Oracle BM.GPU.MI300X.8 both published $6.00 per GPU-hour on-demand in August 2026, Azure in its cheapest regions, and both sell only full 8-GPU nodes, so there is no single-GPU option. On Spheron AI a single MI300X is $3.59/hr. Azure also publishes a spot rate near $1.11 per GPU-hour on the same VM, but spot capacity can be reclaimed without notice.

The AMD MI300X price on Spheron AI is $3.59/hr per GPU per hour on-demand, charged per minute once the minimum runtime (20 minutes on 1x, 60 on 2x) has passed. Published on-demand MI300X rates elsewhere include Runpod's $2.39/hr, which Runpod's console showed as unavailable on 17 Sep 2026. An August 2026 survey of published rates found DigitalOcean at $2.59/hr, Crusoe at $3.45/hr, and $6.00 per GPU-hour in Azure's cheapest regions and on Oracle. The rate moves with marketplace supply, so the live figure on this page is the one that applies.

On Spheron AI the AMD MI300X is $3.59/hr per GPU, billed per minute after a 20-minute minimum runtime (60 minutes on 2x), with no contract. Compared against published rates elsewhere, all per GPU: Runpod lists $2.39/hr (unavailable, 33% cheaper); DigitalOcean lists $2.59/hr (28% cheaper); Crusoe lists $3.45/hr (sold only as 8-GPU nodes, with no single-GPU option, 4% cheaper); Azure ND MI300X v5 lists $6.00/hr (sold only as 8-GPU nodes, with no single-GPU option, 1.7x more expensive); and Oracle BM.GPU.MI300X.8 lists $6.00/hr (sold only as 8-GPU nodes, with no single-GPU option, 1.7x more expensive). Each comparison above is against our on-demand rate. Competitor rates are published on-demand figures from each provider's own pricing page, and the capacity notes are what their own consoles reported on 17 Sep 2026. A listed rate is not the same thing as available capacity.

AMD MI300X alternatives and related GPUs