Home » Tech » AMD Instinct MI250X – we got to know the specification of the top CDNA 2 graphics accelerator with the Aldebaran chip

AMD Instinct MI250X – we got to know the specification of the top CDNA 2 graphics accelerator with the Aldebaran chip


For several months, there have been regular reports about the specifications of the AMD Instinct MI200 graphics accelerator. The graphics chip, based on the CDNA 2 architecture, is to be the first GPU using an MCM structure, like the current AMD Ryzen processors. The manufacturer has previously confirmed the use of the MCM construction for its flagship accelerator, which will have two main computing blocks (two chiplets). Meanwhile, specific information about the specification of the upcoming CDNA 2 flagship has appeared on the web. It turns out that the accelerator will not be called Instinct MI200, but Instinct MI250X. In addition to it, there will also be a slightly weaker variant of the Instinct MI250.

A partial specification of the AMD Instinct M250X graphics accelerator, which is the flagship model based on the CDNA 2 architecture, has appeared on the network. We will get an MCM chip with 128 GB of HBM2e memory and FP32 processing power at 47.9 TFLOPS.

AMD Instinct MI250X - we got to know the specification of the top CDNA 2 graphics accelerator with the Aldebaran chip [1]

AMD Instinct MI300 – First information about the successor of AMD Aldebaran. Accelerators will use as many as four chiplets

Both the AMD Instinct MI250 as well as the Instinct MI250X will be based on the Aldebaran chip. The variant with the annotation X will be equipped with two chiplets and a total of 110 Compute Units. The flagship CDNA 2 chip will be equipped with 128 GB of HBM2e memory, and the core clock will be around 1700 MHz. The computing power in single-precision tasks (FP32) is expected to reach 47.9 TFLOPS. This is more than double that of the AMD Instinct MI100, whose FP32 performance is 23.1 TFLOPS. Interestingly, the Instinct MI250X will thus come closer to the Intel Ponte Vecchio accelerator. In August, it was announced that the Xe-HPC accelerator achieves FP32 performance at the level of 45 TFLOPS.

Intel Ponte Vecchio – characteristics of the architecture of an efficient graphics accelerator designed for the HPC market

According to the disclosed information, AMD Instinct MI250X is to achieve very high performance in double-precision (FP64) calculations. Here, the power expressed in TFLOPS is also to reach the value of 47.9. For comparison, the Instinct MI100 achieves here a performance of 11.5 TFLOPS, while the NVIDIA A100 accelerator only 9.7 TFLOPS. In half-precision calculations (FP16), the AMD Instinct MI250X is to offer performance at the level of 383 TFLOPS (185 TFLOPS for the AMD Instinct MI100 for comparison). The TDP factor of the accelerator is to reach 500 W. It seems that AMD is preparing an accelerator with really huge possibilities for the HPC market.

AMD Instinct MI250X - we got to know the specification of the top CDNA 2 graphics accelerator with the Aldebaran chip [2]

Source: VideoCardz, Twitter @ExecuFix

.

Leave a Comment