How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
MI325X is AMD’s answer to NVIDIA’s H200 — the same compute silicon as MI300X, with a third more memory. Announced 10 October 2024, it carries 256 GB of HBM3E memory 1.
Deep diveThe technical detail
AMD’s own comparison at launch, against NVIDIA’s H200: “1.8X more capacity and 1.3x more bandwidth”, plus “1.3X greater peak theoretical FP16 and FP8 compute performance” 1 — AMD’s own benchmark, aimed at a specific named competitor rather than a general claim.
2.Launch and history
SimpleStart here
Large AI models are often limited less by how fast a chip can calculate and more by how much of the model can fit in memory at once. MI325X answers that specific problem: same compute units as MI300X, a lot more memory to hold bigger models on fewer chips.
Deep diveThe technical detail
One thing to flag plainly: AMD’s own roadmap release in June 2024 previewed MI325X at 288 GB of HBM3E memory and 6 TB/s of bandwidth 2, but the chip that actually shipped in October 2024 carries 256 GB 13 — a real AMD-to-AMD change between preview and shipping product. This page reports the shipping figure, 256 GB, as current, and notes the earlier 288 GB figure as a since-revised preview rather than silently picking one number.
3.What’s inside it
SimpleStart here
The same CDNA 3 chiplet design as MI300X — 304 compute units, 153 billion transistors — with upgraded, higher-capacity memory stacks 34.
Deep diveThe technical detail
AMD’s own MI325X product page lists 304 compute units and 19,456 stream processors, identical to MI300X 35. The 153-billion-transistor figure that AMD states on MI300X’s Hot Chips 2024 slide for the underlying die is repeated on MI325X’s own product page, consistent with both chips sharing the same compute silicon and differing mainly in the HBM3E stacks attached to it 43.
4.Full spec table
SimpleStart here
304 compute units, 256 GB of HBM3E memory, up to 6.0 TB/s of memory bandwidth, a TDP of 1,000W peak 3.
Deep diveThe technical detail
| Spec | AMD Instinct MI325X |
|---|
| Compute units | 304 |
|---|
| Stream processors | 19,456 |
|---|
| Transistors | 153 billion |
|---|
| Memory | 256 GB HBM3E, 8,192-bit |
|---|
| Memory bandwidth | 6.0 TB/s peak |
|---|
| Infinity Cache | 256 MB |
|---|
| Process node | TSMC 5nm (compute) / 6nm (I/O) |
|---|
| Peak engine clock | 2,100 MHz |
|---|
| FP64 | 81.7 TFLOPS (vector) / 163.4 TFLOPS (matrix) |
|---|
| FP32 | 163.4 TFLOPS |
|---|
| FP16 / BF16 | 1.3 PFLOPS |
|---|
| FP8 | 2.61 PFLOPS (2.62 with sparsity) |
|---|
| INT8 | 2.6 POPS |
|---|
| TDP | 1,000W peak |
|---|
Figures per AMD’s own MI325X product page 3. AMD’s launch release states this as a test-configuration power figure rather than a formally rated TDP specification, so this page describes it as AMD’s own stated peak power draw 1.
5.Where it’s made
SimpleStart here
Made by TSMC, the same process split as MI300X: compute chiplets at 5 nanometres, input/output chiplets at 6 nanometres 3.
Deep diveThe technical detail
AMD’s own product page states the same node pairing used for MI300X, consistent with MI325X being a memory-upgraded version of the same underlying chiplet platform rather than a new design 35.
6.Which systems use it
SimpleStart here
AMD named Dell Technologies, Eviden, Gigabyte, HPE, Lenovo and Supermicro as launch partners, with production shipments from the fourth quarter of 2024 and full system availability from early 2025 1.
Deep diveThe technical detail
AMD’s own October 2024 release does not name a hyperscale cloud customer for MI325X the way it later would for MI300X on Azure and Oracle — its named partners at launch are all server-hardware makers, not cloud operators 1. Specialist cloud providers picked it up soon after: DigitalOcean and Vultr both offer MI325X rentals (see pricing below), and it appears on India’s national AI compute portal.
7.Official pricing
SimpleStart here
DigitalOcean lists MI325X from $3.80 per GPU-hour on-demand ($2.88 with a 12-month contract); Vultr lists it from $2.00 per GPU-hour 67.
Deep diveThe technical detail
| Where | Price | Terms |
|---|
| DigitalOcean | $3.80/GPU-hr | On-demand, official pricing page 6 |
| DigitalOcean | $2.88/GPU-hr | 12-month reserved contract, official pricing page; DigitalOcean’s original 2024 launch post quoted a lower introductory contract rate of $1.69/GPU-hr, since superseded 68 |
| Vultr | $2.00/GPU-hr | “Starting at”, official pricing page 7 |
| IndiaAI Mission compute portal, 1 GPU | ₹169.20/hour | On-demand list price 9 |
For comparison, the IndiaAI portal lists MI300X at ₹168.20 an hour on the same page — almost identical to MI325X despite the extra memory 9.
9.What came before, what came next
SimpleStart here
Came before: MI300X, announced December 2023. Came after: the MI350 series (MI350X and MI355X), announced June 2025.
Deep diveThe technical detail
AMD’s own June 2025 release describes the MI350 series as up to a “4x generation-on-generation AI compute increase” over MI300X, and a “35x generational leap in inferencing” — AMD’s own marketing framing for the jump that follows MI325X 10.
10.Hidden in plain sight
SimpleStart here
MI325X’s entire pitch is memory, not raw speed — its compute figures are close to identical to MI300X’s, and the story AMD tells about it is capacity and bandwidth, aimed at fitting bigger AI models on fewer chips.
Deep diveThe technical detail
The clearest evidence of that framing is AMD’s own comparison language: every headline number in its October 2024 release compares MI325X to NVIDIA’s H200 on memory capacity and bandwidth first (“1.8X more capacity and 1.3x more bandwidth”), with compute performance mentioned second 1 — a deliberate sequencing that tells you which specification AMD wanted a reader to notice first.