AMD announces the Instinct MI200 series of data center and supercomputer accelerators. These chips are based on the CDNA 2 architecture and contain 128GB of HBM2e memory. They are also the first GPUs with an extension multi-slide unit-the design.
AMD has initially released two of its accelerators MI200 . seriesThe company announced this Monday evening during its activities datacenterlivestream. The company comes with the AMD Instinct MI250x and MI250, which differ in the number of computing units. The chips use the CDNA 2 architecture, intended for use in data centers and supercomputers. AMD will later also release the MI210 GPU with PCIe form factor, although the specifications are not yet known.
Instinct MI250 accelerators are the first GPUs with a so-called mcm design, or multi-chip unit. In doing so, several die account combined on one slide. The MI250x and MI250 both contain two of these CDNA2Dies, which is produced on a TSMC 6nm node.
MI200: 128 GB HBM2e and up to 220 compute units
The MI250x features two compute units, each containing 110 compute units, for a total of 220 cu. The MI250, in turn, gets two units with 104 cubic meters, for a total of 208 compute units. This is complemented on both chips by the 128GB HBM2e-ecc memory. The maximum clock speed for both options is 1.7GHz.
Accelerators will also receive four HBM2e controllers with a total memory bandwidth of 3.2 TB/s. MI200 GPUs will also receive eight third-generation Infinity Fabric links, which are used to connect the chips together and memory consistency between different CPUs and GPUs.
The MI200 GPUs also feature AMD’s second generation matrix cores, which are intended to perform fp64 and fp32 matrix calculations for HPC and AI applications. According to the manufacturer, these Matrix cores are four times faster than previous AMD Instinct MI100 accelerators. The MI250x gets a total of 880 Matrix cores while the MI250 has that number at 832.
|AMD Instinct MI250x|
|Model||Unit Calculation||stream processors||memory||Bandwidth||FP64/FP32 . Bus||Matrix FP64 / FP32||FP16 / BF16||form factor|
|AMD Instinct MI250x||220||14.080||128 GB HBM2e (etc)||3,2 TB/sec||Up to 47.9Tflops (Peak)||Up to 95.7Tflops (Peak)||Up to 383Tflops (Peak)||OAM|
|AMD Instinct MI250||208||13.312||128 GB HBM2e (etc)||3,2 TB/sec||Up to 45.3Tflops (Peak)||Up to 90.5Tflops (Peak)||Up to 362.1Tflops (peak)||OAM|
Peak performance: “Up to 47.9Tflops at fp64”
According to AMD, this makes the MI200 series much faster than the first generation MI100-gpu’s . Instinct, as well as faster than the competition. MI250x achieves FP64 Vector performance of up to 47.9Tflops. For example, Nvidia’s A100 will achieve a processing power of 9.7Tflops on such FP64 accounts. The company also talks about peak performance of the FP64 and FP32 Matrix of up to 95.7Tflop/s and of the FP16 and BF16 of peak performance of up to 383Tflops.
The first MI200 accelerators are already being delivered to the US Department of Energy’s Oak Ridge National Laboratory. The chips will be used in the Frontier exascale system. The supercomputer is expected to achieve peak performance of “more than 1.5 exaflops”. Thomas Zacharia of Oak Ridge National Laboratory says the MI200 GPU is one more powerful full node than the Summit supercomputer. It also states that Frontier is currently being installed and will be online “early next year”.
AMD announced the Instinct MI200 series on Monday evening, along with new EPYC Milan-X server processors with 3D cache. The company also gave A taste of its own architecture Zen 4, which will be based on a TSMC 5nm node and will be available on EPYC CPUs with different core types.
“Coffee buff. Twitter fanatic. Tv practitioner. Social media advocate. Pop culture ninja.”