AMD Instinct MI100 CDNA Next-Gen GPU Based Accelerator For HPC Launching This Month, Indicate Leaked Documents

AMD’s powerful Instinct MI100 accelerator is rumored to launch this month. Documents confirming the official arrival of the next-gen CDNA-based GPU Accelerator meant for High-Performance Computing (HPC) and Data Server segments, have allegedly leaked online. The AMD Instinct MI100 GPU confirms AMD is following NVIDIA’s footsteps in terms of naming their GPUs that will serve specific industries and not gaming.

AMD Instinct MI100 Graphics Accelerator, not a Graphics Card in the traditional sense, could arrive on November 16, indicates a leaked set of documents. The AMD GPU, based on the CDNA Architecture, will compete directly against NVIDIA’s NVIDIA A100 GPU, which is based on the Ampere Architecture and is tagged as GA100 GPU.

AMD appears to be dropping its Radeon branding for the Instinct series of accelerators. NVIDIA has already dropped the Tesla and Quadro series in favor of simpler “Axxx” branding.

AMD Instinct MI100 With First-Generation CDNA Architecture Specifications, Features:

AMD’s CTO Mark Papermaster had confirmed the existence of the AMD Instinct MI100 almost 5 months ago. Mark had indicated that AMD will be introducing the cDNA based Instinct GPU by the second half of 2020. After a little delay, AMD appears ready to launch the most powerful data center GPU the company has ever built under the leadership of RTG’s new chief, David Wang.

The MI100 is an accelerator-based on CDNA architecture. This card will feature a new Arcturus GPU that does not have the standard rendering technologies as they are unnecessary to datacenter and HPC accelerators. AMD had earlier decided to split its GPU architectures for computing (CDNA) and gaming-oriented (RDNA) series. The company clearly intends to capture a bigger part of the data center as well as gaming markets at the same time.

The CDNA architecture has been designed specifically for the HPC segment and will compete against NVIDIA’s Ampere A100 and similar accelerator cards. The Radeon Instinct MI100 ‘Arcturus’ GPU will reportedly feature several variants. The flagship variant is D34303 SKU which makes use of the XL variant. Apart from the MI100, AMD is also reportedly offering Instinct V640 and V620 GPU accelerators.

The AMD Instinct MI100 will feature 34 TFLOPs of FP32 compute per GPU. Each Radeon Instinct MI100 GPU will have a TDP of 300W. Each GPU will feature 32 GB of HBM2e memory which should have a total of 1.225 TB/s bandwidth.

AMD Instinct MI100 Vs. NVIDIA A100 GPU:

Leaked data suggests the AMD Radeon Instinct MI100 is being compared to the NVIDIA Volta V100 and the NVIDIA Ampere A100 GPU accelerators. NVIDIA A100 configuration comes in two TDP profiles: the 400W configuration in the SXM form factor and the 250W configuration which comes in the PCIe form factor.

[Image Credit: WCCFTech]
Data suggest the Radeon Instinct MI100 has about 13 percent better FP32 performance when compared to the NVIDIA Ampere A100 and over 2x performance increase against the earlier Volta V100 GPUs. The performance to value ratio of AMD MI100 is claimed to be around 2.4x better value compared to the V100S and 50 percent better value than the Ampere A100. It is also shown that the performance scaling is near-linear even with up to 32 GPU configurations in Resenet.

Slides also indicate AMD has built the MI100 Accelerator primarily for the Oil & Gas, Academia, and HPC and Machine Learning. NVIDIA appears to have cornered the HPC workloads such as AI & data analytics, for now. This is because NVIDIA offers the benefit of Multi-Instance GPU architecture over AMD. In this particular area, NVIDIA GPUs have 2.5x better FP64 performance, 2x better FP16 performance, and twice the tensor performance due to the latest Tensor cores within the NVIDIA Ampere A100 GPU.

Alap Naik Desai
A B.Tech Plastics (UDCT) and a Windows enthusiast. Optimizing the OS, exploring software, searching and deploying solutions to strange and weird issues is Alap's main interest.