AMD Instinct MI300X, MI325X and MI355X are AI accelerators whose headline difference is memory: AMD specifies 192 GB, 256 GB and 288 GB respectively. The live comparison below shows how their rental prices stack up against NVIDIA H200 and B200. Choose AMD for memory-bound large-model inference when its price per GB and measured serving cost win; stay with NVIDIA when your required software or deadline makes migration a poor trade.
Start with the memory boundary
Choose the smallest accelerator that fits your workload with room for the serving configuration you intend to use. Do not budget for model weights alone. Include the memory your actual run needs, then test at the target concurrency and context length. That gives the memory comparison a job: deciding which configurations deserve a rental trial.
| Spec | MI300X | MI325X | MI355X | H100 | H200 | B200 |
|---|---|---|---|---|---|---|
| VRAM | 192 GB | 256 GB | 288 GB | 80 to 94 GB | 141 GB | 180 to 192 GB |
| Memory type | HBM3 | HBM3e | HBM3e | HBM3 | HBM3e | HBM3e |
| Memory bandwidth | 5,300 GB/s | 6,000 GB/s | 8,000 GB/s | 3,350 GB/s | 4,800 GB/s | 8,000 GB/s |
| FP16 (dense) | 1,307.4 TFLOPS | 1,307.4 TFLOPS | 2,500 TFLOPS | 989 TFLOPS | 989 TFLOPS | 2,250 TFLOPS |
| FP16 (with sparsity) | 2,614.9 TFLOPS | 2,614.9 TFLOPS | 5,000 TFLOPS | 1,979 TFLOPS | 1,979 TFLOPS | 4,500 TFLOPS |
| FP8 (dense) | 2,614.9 TFLOPS | 2,614.9 TFLOPS | 5,000 TFLOPS | 1,979 TFLOPS | 1,979 TFLOPS | 4,500 TFLOPS |
| FP8 (with sparsity) | 5,229.8 TFLOPS | 5,229.8 TFLOPS | 10,100 TFLOPS | 3,958 TFLOPS | 3,958 TFLOPS | 9,000 TFLOPS |
| FP4 (dense) | Not published | Not published | 10,100 TFLOPS | Not published | Not published | 9,000 TFLOPS |
| FP4 (with sparsity) | Not published | Not published | Not published | Not published | Not published | 18,000 TFLOPS |
| TDP | 750 W | 1,000 W | 1,400 W | 700 W | 700 W | 1,000 W |
| Launch year | 2023 | 2024 | 2025 | 2022 | 2024 | 2024 |
These are family specifications, not a promise about every rented server. AMD's theoretical compute figures are VENDOR CLAIM, not independently measured model throughput. Compare dense with dense and sparse with sparse. Do not compare AMD's sparse peak against NVIDIA's dense row and call the difference a speedup. A blank FP4 entry also gives you no basis for assigning a native FP4 throughput figure.
The AMD MI325X is primarily a memory upgrade over MI300X in this comparison: their dense FP16 and FP8 entries match. MI355X raises both memory capacity and the listed dense compute peaks. Start with MI300X versus H100 for an existing H100 workload, then H200 versus MI300X if memory is the constraint. For a larger AMD candidate, use H200 versus MI355X.
Price per GB is a screening tool. Divide the live per-GPU price by that GPU's memory capacity. Then compare the cost of the full configuration that actually serves your model. Reject a nominal memory bargain if it misses your latency target, requires unused GPUs in the billing unit, or takes too much engineering time to make reliable.
The lineup, with announcement dates kept separate
AMD's dated announcements establish the following chronology. An introduction date, a launch date and a production statement answer different questions; none is a substitute for checking a rental offer.
| Product | Dated AMD record | How to use it |
|---|---|---|
| MI250X | Announced 8 November 2021 | Earlier generation; keep it separate from the three rental candidates here. |
| MI300X | Introduced 13 June 2023; availability announcement 6 December 2023 | Starting point for the memory and cost comparison. |
| MI325X | Final launch announcement 10 October 2024 | Compare the larger memory allocation against MI300X. |
| MI350X | Unveiled 12 June 2025 | Distinct product from MI355X, despite the shared launch. |
| MI355X | Unveiled 12 June 2025 | Test against the larger NVIDIA candidates. |
| MI400 series | Announced 2 June 2024 for 2026 | Roadmap announcement; do not substitute it for a specific model. |
| Helios | Announced as an MI400-based rack on 12 June 2025; AMD identified MI455X on 5 January 2026 | AMD's 23 July 2026 update described Helios as in production. |
AMD's CDNA 4 whitepaper, accessed 28 September 2026, describes MI350X as air-cooled and MI355X platforms as direct liquid-cooled. That distinction matters when requesting a purchase quote: specify the exact GPU and cooling design, rather than asking for a generic MI350 server.
For the MI450 name, AMD's 24 February 2026 Meta agreement specifies a custom MI450-based GPU. Treat that agreement as its own product context. Do not use the name as a shortcut for the MI355X configuration you can evaluate here.
AMD MI300X price: historical estimates, not a retail quote
For an AMD Instinct MI300X price reference, Tom's Hardware reported Citi's ANALYST ESTIMATE on 2 February 2024: roughly $10,000 per accelerator sold to Microsoft, and around $15,000 for other customers. Those were estimated volume unit prices. They were neither complete-server prices nor individual purchase offers. In the same report, AMD said it did not publicly share MI300 pricing.
Do not put either estimate into a September 2026 procurement budget as a current quote. Ask for a dated system proposal with the accelerator model, quantity, support term and installation requirements spelled out. Compare it with the same scope on the NVIDIA side; the NVIDIA purchase-price guide gives that comparison its own space.
Supermicro's US eStore listed a starting price of $254,032.19 on 28 September 2026 for the AS-8126GS-TNMR system, advertised with eight MI325X or MI350X accelerators; the listing did not say which GPU the starting price included. It is useful as a quote-request reference, not an AMD Instinct MI325X unit price. Do not divide it into a chip price.
For MI325X and MI355X, ask for a complete system quote rather than a card price. Rent first unless you already have a validated workload and a complete system quote. For buying, budget the installed system, power, cooling, maintenance and the people responsible for it. Compare ownership against the hours you expect to use, not continuous utilization assumed for convenience.
Rent the right billing unit
The live table shows provider-level comparisons for the three AMD families. Use it to shortlist an offer, then confirm the system configuration and billing terms before starting a job.
No provider has an in-stock on-demand offer for MI300X, MI325X, MI355X right now.
For hyperscaler context, AMD's 21 May 2024 announcement introduced Azure ND MI300X v5 general availability. Oracle announced OCI BM.GPU.MI300X.8 on 26 September 2024 and MI355X general availability on 14 October 2025. Microsoft's documentation, accessed 28 September 2026, describes its ND MI300X v5 VM as an eight-accelerator system; Oracle's announcements describe eight-GPU shapes too. These are dated product records, not current capacity confirmations.
Among neoclouds, Vultr's 9 September 2025 announcement described MI355X bare-metal servers and eight-GPU VM plans. Hot Aisle's MI300X documentation, accessed 28 September 2026, describes single-GPU KVM VMs and complete eight-GPU Dell PowerEdge XE9680 servers. Compare the live listings on Vultr and Hot Aisle without assuming their billing units are interchangeable.
DigitalOcean announced MI325X GPU Droplets on 17 July 2025. RunPod's 27 August 2026 MI300X page describes Secure Cloud configurations with up to eight accelerators. TensorWave's MI300X documentation, accessed 28 September 2026, describes an eight-GPU bare-metal node with optional managed Kubernetes or Slurm. These examples give you deployment formats to ask about; use current offers to decide where to spend.
Open the MI300X rental comparison, MI325X rental comparison and MI355X rental comparison when selecting the actual offer. Check whether your budget covers a single accelerator or a whole node. Ask for the software image and network configuration with the quote, especially if your planned test crosses server boundaries.
Published tests narrow the choice
SemiAnalysis' 22 December 2024 training investigation followed a five-month evaluation of MI300X against H100 and H200. It found public AMD software materially less usable than custom development images. Its 21 December work-in-progress build beat H100/H200 on BF16 Llama 3 8B and a four-layer Llama 3 70B proxy, but used an unmerged AMD branch. That is evidence about those builds and tests, not a current stock-install guarantee.
SemiAnalysis' 16 February 2026 InferenceX v2 assessment found competitive MI355X versus B200 results on DeepSeek-R1 FP8 using SGLang, disaggregated prefill and MoRI/Dynamo respectively. The outcome depended on the throughput-versus-interactivity target. The same assessment found a weakness for MI355X when combining FP4, disaggregated serving and wide expert parallelism. Choose a comparable test setup before applying either conclusion.
Artificial Analysis wrote on 8 June 2025: "AMD MI300X systems achieved higher peak system throughput at high concurrency." AMD commissioned that work and supplied access to DigitalOcean hardware. The tests used eight-GPU systems: FP8 DeepSeek R1 on SGLang against H200, and FP8 Llama 4 Maverick on vLLM against H100, with 1,000 input and 1,000 output tokens and three-minute load phases. Software builds differed across vendors. The high-concurrency qualification belongs with the result.
MLPerf supplies benchmark context, but participation alone does not establish a winner. AMD's 2 April 2025 report identified MI325X submissions for Llama 2 70B and Stable Diffusion XL. In MLCommons' 16 June 2026 Training v6.0 supplemental discussion, AMD identified Flux.1 FP8 data-parallel training on 64 MI325X GPUs. MLCommons published Inference v6.1 on 16 September 2026; AMD's accompanying report identified a Dell/MangoBoost GPT-OSS-120B submission combining 16 MI300X GPUs in Korea with 16 MI355X GPUs in the United States. Do not turn these different setups into a single-GPU ranking.
Customer agreements are context, not your benchmark
AMD and OpenAI announced a six-gigawatt, multigeneration agreement on 6 October 2025. AMD's 23 July 2026 update said OpenAI expected Helios online beginning in Q4 2026. AMD and Meta announced an agreement for up to six gigawatts on 24 February 2026, with first-gigawatt shipments scheduled for the second half of 2026 using a custom MI450-based GPU and Helios.
Those agreements justify evaluating AMD seriously. They do not establish the cost or performance of your workload, and scheduled deployments are not completed deliveries. Keep the procurement decision tied to the system you can test and the offer you can sign.
Keep the software trial small
PyTorch's HIP documentation, dated 11 August 2026, says its ROCm implementation reuses torch.cuda interfaces. AMD's HIPIFY documentation, accessed 28 September 2026, says it cannot translate a library without a HIP equivalent and recommends correctness testing and optimization after migration. Read the ROCm versus CUDA guide for the compatibility work.
For this rental decision, freeze the model, precision, prompts, concurrency and latency requirement. Test the full serving path, including your extensions. Record the software image with the results. Compare whole-system cost for the same useful output and quality target, then add the engineering time needed to get there.
Start with MI300X when it fits. Pay for MI325X memory or AMD MI355X performance only when your trial demonstrates the benefit. Choose AMD when memory-bound inference delivers a lower validated serving cost; keep NVIDIA when a required dependency, latency target or migration deadline makes it the better working system.
Sources
-
AMD MI200 announcement, 2021-11-08.
-
AMD MI300X introduction, 2023-06-13.
-
AMD MI300X launch, 2023-12-06.
-
AMD MI325X launch, 2024-10-10.
-
AMD Advancing AI 2025, 2025-06-12.
-
AMD 2024 roadmap, 2024-06-02.
-
AMD CES 2026, 2026-01-05.
-
AMD Advancing AI 2026, 2026-07-23.
-
AMD CDNA 4 whitepaper, retrieved 2026-09-28.
-
AMD and Meta agreement, 2026-02-24.
-
Tom's Hardware: Citi MI300X price estimates, 2024-02-02.
-
Supermicro AS-8126GS-TNMR store, retrieved 2026-09-28.
-
AMD: Azure MI300X announcement, 2024-05-21.
-
Microsoft ND MI300X v5 documentation, retrieved 2026-09-28.
-
Oracle MI300X announcement, 2024-09-26.
-
Oracle MI355X announcement, 2025-10-14.
-
Vultr MI355X announcement, 2025-09-09.
-
Hot Aisle MI300X documentation, retrieved 2026-09-28.
-
DigitalOcean MI325X announcement, 2025-07-17.
-
RunPod MI300X documentation, 2026-08-27.
-
TensorWave MI300X documentation, retrieved 2026-09-28.
-
SemiAnalysis training investigation, 2024-12-22.
-
SemiAnalysis InferenceX v2, 2026-02-16.
-
Artificial Analysis MI300X comparison, 2025-06-08.
-
AMD MLPerf Inference v5.0 report, 2025-04-02.
-
MLCommons Training v6.0 supplemental discussion, 2026-06-16.
-
MLCommons Inference v6.1 results, 2026-09-16.
-
AMD Inference v6.1 report, 2026-09-16.
-
AMD and OpenAI agreement, 2025-10-06.
-
PyTorch HIP semantics, 2026-08-11.
-
AMD HIPIFY documentation, retrieved 2026-09-28.
-
AMD MI300X datasheet, retrieved 2026-09-28.
-
AMD MI325X datasheet, retrieved 2026-09-28.
-
AMD MI355X brochure, retrieved 2026-09-28.