AMD Instinct MI325X
AMD: specifications, comparisons and alternatives.
The AMD Instinct MI325X is a CDNA 3 accelerator for AI training, inference and HPC, with an emphasis on memory. According to AMD, it has 256 GB of HBM3E, 6 TB/s of memory bandwidth and a board power of 1,000 W. According to the page, the OAM module launched on 10 October 2024. It is a newer variant within the MI300 series.
AMD
- 256 GB
- Memory
- HBM3E
- Memory type
- 6 TB/s
- Memory bandwidth
- 1.000 W
- Power (board power or TDP)
You can see the current price and availability with the partner as soon as a link has been set up.
How are the points calculated?
Clouwino points are an arithmetic summary of the specifications the manufacturer states, on a scale of 0 to 100. It is not a benchmark and not a judgement on quality or suitability: a product with fewer points may suit your situation better. Products are compared with other products of the same type. For each property we count how far a product sits between the lowest and highest value in the group:
- FP32 compute (for: consumer), carries heavy weight
- Memory (for: consumer), carries heavy weight
- Memory bandwidth (for: consumer), carries heavy weight
- Compute units (for: consumer), carries light weight
- Boost clock (for: consumer), carries light weight
- Memory (for: professional and data centre), carries medium weight
- Memory bandwidth (for: professional and data centre), carries medium weight
- FP16/BF16 (for: professional and data centre), carries heavy weight
- FP8 (for: professional and data centre), carries medium weight
- FP64 (for: professional and data centre), carries medium weight
If too much data is missing, we do not show points. Points are relative: when a new product is added, they can shift. Read more at how we compare.
Specifications of AMD Instinct MI325X
All specifications the manufacturer lists, per item. Want to see another product next to it? Pick one below.
| Specificatie | AMD Instinct MI325X |
|---|---|
| Overview | |
| Clouwino points | 56 punten |
| General | |
| Architecture | CDNA 3 |
| Manufacturing process | TSMC 5nm | 6nm FinFET |
| Form | OAM |
| Launch | 2024-10 |
| Compute | |
| FP32 compute | 163,4 TFLOPS |
| FP64 | 81,7 TFLOPS |
| FP16/BF16 | 1.300 TFLOPS |
| FP8 | 2.610 TFLOPS |
| FP4 | Not stated by the manufacturer |
| Performance notes | Same compute figures as MI300X. FP16/BF16: 1.3 PFLOPS dense (2.61 PFLOPS with structured sparsity). FP8: 2.61 PFLOPS dense (5.22 PFLOPS sparse). FP64 81.7 vector (matrix 163.4), FP32 163.4 TFLOPS. PFLOPS converted to TFLOPS (unit only). FP4 not stated. |
| Memory | |
| Memory | 256 GB |
| Memory type | HBM3E |
| Memory bandwidth | 6 TB/s |
| Power and connections | |
| Power (board power or TDP) | 1.000 W |
| GPU interconnect | 8x Infinity Fabric, 128 GB/s per link; PCIe 5.0 x16 |
| Multiple instances | Not stated by the manufacturer |
Note: Figures are per GPU. TBP 1000 W peak. AMD footnote: memory figures 'based on current specifications and/or estimation'; actual results may differ. Partitioning not described (SR-IOV only). Launch date 10-10-2024 according to the page.
Alternatives to AMD Instinct MI325X
Products with similar specifications from the same overview. This is not a ranking and not buying advice.
- AMD Instinct MI300X192 GB · HBM3 · 5,3 TB/sCompare with AMD Instinct MI325X →
- AMD Instinct MI355X288 GB · HBM3E · 8 TB/sCompare with AMD Instinct MI325X →
- NVIDIA H200 SXM141 GB · HBM3e · 4,8 TB/sCompare with AMD Instinct MI325X →
- NVIDIA H100 SXM80 GB · HBM3 · 3,35 TB/sCompare with AMD Instinct MI325X →
- AMD Instinct MI250X128 GB · HBM2e · 3,2 TB/sCompare with AMD Instinct MI325X →
- NVIDIA Rubin (per GPU)288 GB · HBM4 · 22 TB/sCompare with AMD Instinct MI325X →
What the MI325X means for the MI300 series
The MI325X shares the CDNA 3 architecture and the TSMC 5nm and 6nm FinFET technology with the MI300X, but has more and newer memory: 256 GB of HBM3E instead of 192 GB of HBM3. According to the data, the compute figures have stayed the same.
Like the MI300X, it is an OAM module with 8 Infinity Fabric links of 128 GB/s per link and a PCIe 5.0 x16 connection. The page describes partitioning only as SR-IOV. AMD notes that the memory figures are based on current specifications and/or estimates, and that actual results may differ.
If you know the MI300X, you will mainly see a memory upgrade within the same architecture.
Bear in mind that the MI325X is mainly interesting when memory capacity is the limiting factor in your setup. Because the compute figures are the same as those of the MI300X, a choice between the two mainly comes down to memory size, bandwidth and the power available in your server room.
MI325X numbers in plain English
The large memory allows bigger models or longer contexts on a single accelerator, and 6 TB/s of bandwidth feeds the data in. See what is VRAM for the basics. The board power is 1,000 W as a peak value.
The compute figures are per accelerator. FP16 and BF16 are 1.3 PFLOPS dense and 2.61 PFLOPS with structured sparsity. FP8 is 2.61 PFLOPS dense and 5.22 PFLOPS with sparsity. FP64 is 81.7 TFLOPS for vector (163.4 for matrix) and FP32 is 163.4 TFLOPS. AMD does not list FP4.
That difference in memory size can determine whether a model fits on one accelerator or has to be split across several.
Choosing an MI325X system
At 1,000 W per module, cooling plays the leading role. A server with multiple OAM modules needs sufficient power and airflow or liquid cooling, depending on what the supplier offers.
- See whether the server platform supports OAM modules of this power.
- Read the checklist for GPU instances if you use the cloud.
- Only compare figures with the same units and the same label.
When comparing with other accelerators, it is wise to use only the same units and the same labels.
You can see the current price and availability at the partner via the button on the page.
Frequently asked questions
What is the memory of the AMD Instinct MI325X?
The MI325X has 256 GB of HBM3E with a bandwidth of 6 TB/s. AMD notes that these memory figures are based on current specifications and/or estimates and may differ.
Are the MI325X compute figures different from those of the MI300X?
According to the data, the compute figures are the same: FP16 and BF16 1.3 PFLOPS dense, FP8 2.61 PFLOPS dense. The difference lies mainly in the memory, its type and the board power of 1,000 W.
When was the MI325X launched?
According to the AMD page, the launch date is 10 October 2024. The MI325X is an OAM accelerator with a board power of 1,000 W as a peak value. The accelerator is intended for AI training, inference and HPC in servers.
Last updated on
