AMD Instinct MI325X
Memory-flagship CDNA 3 accelerator with 256 GB of HBM3e — the largest single-chip memory in the inference market — and 6 TB/s of memory bandwidth. Pin-compatible with MI300X OAM platforms, positioning AMD against H200 / B100 on capacity-bound LLM workloads.
VRAM
256 GB
Memory
HBM3e
Bandwidth
6,000 GB/s
FP16
1,307 TF
TDP
1000 W
Memory and compute
Architecture and power
Notes
FP16/BF16 figures are dense matrix throughput (no sparsity), matching the convention used for MI300X. With 2:4 sparsity, peak FP16 is 2614 TFLOPS.
Models that may fit on AMD Instinct MI325X
Estimates based on INT8 quantization. Actual fit depends on framework and batch size.
More GPUs by AMD
The manufacturer's other cards in the catalog, largest memory first, with bandwidth and power where published.
Frequently asked questions about AMD Instinct MI325X
Answered from the entry's own figures: memory, which models fit, bandwidth, power and the card's class.
AMD Instinct MI325X has 256 GB of HBM3e at 6,000 GB/s. At INT8 a model's weights take about one byte per parameter, so the card holds roughly a 204B-parameter model with room for context; larger models need several cards or a lower precision.
From the catalog's estimates at INT8, Gemma 4 31B IT, Qwen3 235B A22B Thinking 2507, Llama Nemotron Embed VL 1B V2 and GLM Z1 9B 0414 fit on a single AMD Instinct MI325X with their default context. The list on this page shows the estimated memory each takes; the capacity planner sizes any model against this card for your own context length, batch size and request rate.
6,000 GB/s. Bandwidth bounds how fast a model generates tokens, because every token reads the whole set of weights and the KV cache from memory; at this rate a 204B model at INT8 could be read on the order of 29 times a second on one card, before batching and framework efficiency.
AMD Instinct MI325X has a TDP of 1000 W, with a maximum of 1000 W, in the OAM form factor. Budget the TDP per card plus host overhead when sizing a server's power and cooling.
AMD Instinct MI325X is a datacenter accelerator, meant for servers and multi-GPU serving, built on the CDNA 3 architecture and released in 2024. It supports a multi-GPU interconnect at 896 GB/s, so a model larger than one card can be split across several.
Specifications from the manufacturer's published material; a figure that is not published is shown as such, and a derived one is explained in the notes.
Added Apr 30, 2026 · Last updated Apr 30, 2026
Explore more
Companion tools that draw on the same catalog and routing engine.
Put intelligence in the middle.
One gateway in front of every model, with your policies applied and every decision on record. Start with $5 of credit and 5,000 routing decisions a month, no card required.