| AMD Instinct | |
| Designfirm: | AMD |
| Marketed By: | AMD |
| Model: | MI Series |
| Transistors1: | 5.7B (Polaris10) 14 nm |
| Transistors2: | 8.9B (Fiji) 28 nm |
| Transistors3: | 12.5B (Vega10) 14 nm |
| Transistors4: | 13.2B (Vega20) 7 nm |
| Transistors5: | 25.6B (Arcturus) 7 nm |
| Transistors6: | 58.2B (Aldebaran) 6 nm |
| Transistors7: | 146B (Antares) 5 nm |
| Transistors8: | 153B (Aqua Vanjaram) 5 nm |
| Numcores: | 36-304 Compute Units (CUs) |
AMD Instinct is AMD's brand of data center GPUs.[1] [2] It replaced AMD's FirePro S brand in 2016. Compared to the Radeon brand of mainstream consumer/gamer products, the Instinct product line is intended to accelerate deep learning, artificial neural network, and high-performance computing/GPGPU applications.
The AMD Instinct product line directly competes with Nvidia's Tesla and Intel's Xeon Phi and Data Center GPU lines of machine learning and GPGPU cards.
The brand was originally known as AMD Radeon Instinct, but AMD dropped the Radeon brand from the name before AMD Instinct MI100 was introduced in November 2020.
In June 2022, supercomputers based on AMD's Epyc CPUs and Instinct GPUs took the lead on the Green500 list of the most power-efficient supercomputers with over 50% lead over any other, and held the top first 4 spots.[3] One of them, the AMD-based Frontier is since June 2022 and as of 2023 the fastest supercomputer in the world on the TOP500 list.[4] [5]
| Launch date | Architecture | Lithography | Compute Units | Memory | PCIe support | Form factor | Processing power | TBP | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Size | Type | Bandwidth (GB/s) | FP16 | BF16 | FP32 | FP32 matrix | FP64 performance | FP64 matrix | INT8 | INT4 | ||||||||
| MI6 | 2016-12-12[6] | GCN 4 | 14 nm | 36 | 16 GB | GDDR5 | 224 | 3.0 | PCIe | 5.7 TFLOPS | N/A | 5.7 TFLOPS | N/A | 358 GFLOPS | N/A | N/A | N/A | 150 W |
| MI8 | GCN 3 | 28 nm | 64 | 4 GB | HBM | 512 | 8.2 TFLOPS | 8.2 TFLOPS | 512 GFLOPS | 175 W | ||||||||
| MI25 | GCN 5 | 14 nm | 16 GB | HBM2 | 484 | 26.4 TFLOPS | 12.3 TFLOPS | 768 GFLOPS | 300 W | |||||||||
| MI50 | 2018-11-06[7] | 7 nm | 60 | 1024 | 4.0 | 26.5 TFLOPS | 13.3 TFLOPS | 6.6 TFLOPS | 53 TOPS | 300 W | ||||||||
| MI60 | 64 | 32 GB | 29.5 TFLOPS | 14.7 TFLOPS | 7.4 TFLOPS | 59 TOPS | 300 W | |||||||||||
| MI100 | 2020-11-16 | CDNA | 120 | 1200 | 184.6 TFLOPS | 92.3 TFLOPS | 23.1 TFLOPS | 46.1 TFLOPS | 11.5 TFLOPS | 184.6 TOPS | 300 W | |||||||
| MI210 | 2022-03-22[8] | CDNA 2 | 6 nm | 104 | 64 GB | HBM2e | 1600 | 181 TFLOPS | 22.6 TFLOPS | 45.3 TFLOPS | 22.6 TFLOPS | 45.3 TFLOPS | 181 TOPS | 300 W | ||||
| MI250 | 2021-11-08[9] | 208 | 128 GB | 3200 | OAM | 362.1 TFLOPS | 45.3 TFLOPS | 90.5 TFLOPS | 45.3 TFLOPS | 90.5 TFLOPS | 362.1 TOPS | 560 W | ||||||
| MI250X | 220 | 383 TFLOPS | 47.92 TFLOPS | 95.7 TFLOPS | 47.9 TFLOPS | 95.7 TFLOPS | 383 TOPS | 560 W | ||||||||||
| MI300A | 2023-12-06[10] | CDNA 3 | 6 & 5 nm | 228 | 128 GB | HBM3 | 5300 | 5.0 | APU SH5 socket | 980.6 TFLOPS 1961.2 TFLOPS (with Sparsity) | 122.6 TFLOPS | 61.3 TFLOPS | 122.6 TFLOPS | 1961.2 TOPS 3922.3 TOPS (with Sparsity) | N/A | 550 W 760 W (with liquid cooling) | ||
| MI300X | 304 | 192 GB | OAM | 1307.4 TFLOPS 2614.9 TFLOPS (with Sparsity) | 163.4 TFLOPS | 81.7 TFLOPS | 163.4 TFLOPS | 2614.9 TOPS 5229.8 TOPS (with Sparsity) | N/A | 750 W | ||||||||
| MI325X | 2024-06-02[11] | 288 GB | HBM3e | 6000 | ||||||||||||||
The MI6 is a passively cooled, Polaris 10 based card with 16 GB of GDDR5 memory and with a <150 W TDP. At 5.7 TFLOPS (FP16 and FP32), the MI6 is expected to be used primarily for inference, rather than neural network training. The MI6 has a peak double precision (FP64) compute performance of 358 GFLOPS.[14]
The MI8 is a Fiji based card, analogous to the R9 Nano, has a <175W TDP. The MI8 has 4 GB of High Bandwidth Memory. At 8.2 TFLOPS (FP16 and FP32), the MI8 is marked toward inference. The MI8 has a peak (FP64) double precision compute performance 512 GFLOPS.[15]
The MI25 is a Vega based card, utilizing HBM2 memory. The MI25 performance is expected to be 12.3 TFLOPS using FP32 numbers. In contrast to the MI6 and MI8, the MI25 is able to increase performance when using lower precision numbers, and accordingly is expected to reach 24.6 TFLOPS when using FP16 numbers. The MI25 is rated at <300W TDP with passive cooling. The MI25 also provides 768 GFLOPS peak double precision (FP64) at 1/16th rate.[16]
The MI300A and MI300X are data center accelerators that use the CDNA 3 architecture, which is optimized for high-performance computing (HPC) and generative artificial intelligence (AI) workloads. The CDNA 3 architecture features a scalable chiplet design that leverages TSMC’s advanced packaging technologies, such as CoWoS (chip-on-wafer-on-substrate) and InFO (integrated fan-out), to combine multiple chiplets on a single interposer. The chiplets are interconnected by AMD’s Infinity Fabric, which enables high-speed and low-latency data transfer between the chiplets and the host system.
The MI300A is an accelerated processing unit (APU) that integrates 24 Zen 4 CPU cores with four CDNA 3 GPU cores, resulting in a total of 228 CUs in the GPU section, and 128 GB of HBM3 memory. The Zen 4 CPU cores are based on the 5 nm process node and support the x86-64 instruction set, as well as AVX-512 and BFloat16 extensions. The Zen 4 CPU cores can run general-purpose applications and provide host-side computation for the GPU cores. The MI300A has a peak performance of 61.3 TFLOPS of FP64 (122.6 TFLOPS FP64 matrix) and 980.6 TFLOPS of FP16 (1961.2 TFLOPS with sparsity), as well as 5.3 TB/s of memory bandwidth. The MI300A supports PCIe 5.0 and CXL 2.0 interfaces, which allow it to communicate with other devices and accelerators in a heterogeneous system.
The MI300X is a dedicated generative AI accelerator that replaces the CPU cores with additional GPU cores and HBM memory, resulting in a total of 304 CUs (64 cores per CU) and 192 GB of HBM3 memory. The MI300X is designed to accelerate generative AI applications, such as natural language processing, computer vision, and deep learning. The MI300X has a peak performance of 653.7 TFLOPS of TP32 (1307.4 TFLOPS with sparsity) and 1307.4 TFLOPS of FP16 (2614.9 TFLOPS with sparsity), as well as 5.3 TB/s of memory bandwidth. The MI300X also supports PCIe 5.0 and CXL 2.0 interfaces, as well as AMD’s ROCm software stack, which provides a unified programming model and tools for developing and deploying generative AI applications on AMD hardware.[17] [18] [19]
See main article: ROCm.
Following software is, as of 2022, regrouped under the Radeon Open Compute meta-project.
The MI6, MI8, and MI25 products all support AMD's MxGPU virtualization technology, enabling sharing of GPU resources across multiple users.[20]
MIOpen is AMD's deep learning library to enable GPU acceleration of deep learning. Much of this extends the GPUOpen's Boltzmann Initiative software.[20] This is intended to compete with the deep learning portions of Nvidia's CUDA library. It supports the deep learning frameworks: Theano, Caffe, TensorFlow, MXNet, Microsoft Cognitive Toolkit, Torch, and Chainer. Programming is supported in OpenCL and Python, in addition to supporting the compilation of CUDA through AMD's Heterogeneous-compute Interface for Portability and Heterogeneous Compute Compiler.