
AMD
AMD Instinct MI300X 192GB
CDNA 3 OAM accelerator with 192 GB HBM3, deployed eight to a universal baseboard for large-model training and inference.
| Dostupnost | Předobjednávka |
|---|---|
| SKU | AXL-GPU-MI300X |
| Minimální objednávka | 8 units |
| Stav | Nové |
| Ceny | Dle cenové nabídky |
Přehled
The Instinct MI300X pairs 192 GB of HBM3 with 5.3 TB/s of memory bandwidth, so a single accelerator can hold model weights that would otherwise be split across several devices. That capacity suits large language model inference and training on the AMD ROCm software stack, alongside memory-bound HPC work. The MI300X is an OCP Accelerator Module rather than a card: it mounts eight to a universal baseboard inside a qualified chassis, so platform, power and cooling have to be specified together with the modules.
AXISLINK therefore quotes MI300X as complete baseboard or system configurations rather than loose parts, because that is how the platform is deployed and serviced. Sourcing is through verified channels, serial numbers are documented on the quotation, and we confirm the exact server model and firmware level before the order is placed. Export paperwork, insurance and freight are handled from Hong Kong, and end-user information may be required for this class of accelerator.
Specifikace
| Architecture | AMD CDNA 3 |
|---|---|
| Memory | 192 GB HBM3 |
| Memory bandwidth | 5.3 TB/s |
| Compute units | 304 |
| Form factor | OAM module, mounted 8 per universal baseboard (UBB) |
|---|---|
| Interconnect | AMD Infinity Fabric peer-to-peer links between modules |
| Host interface | PCIe Gen5 x16 |
| Software | AMD ROCm |
| Peak board power | 750 W |
|---|---|
| Cooling | Chassis level, air or liquid cooled depending on platform |
Vyžádat nabídku
Cenová nabídka na AMD Instinct MI300X 192GB. Odpovíme formální proforma fakturou, obvykle do 24 hodin.
Další v kategorii GPU a akcelerátory
Zobrazit kategorii
NVIDIA
NVIDIA H200 141GB SXM
Hopper SXM module with 141 GB HBM3e for large-model training and memory-bound inference on HGX baseboards.

NVIDIA
NVIDIA L40S 48GB
Ada Lovelace data-center card with 48 GB GDDR6 ECC for inference, fine-tuning, rendering and virtual workstations.

NVIDIA
NVIDIA H100 80GB
Hopper-architecture data-center GPU for large-scale AI training and inference. PCIe and SXM variants.