Nearly two years after AMD confirmed the next-gen Arcturus GPU, the Radeon Instinct MI100, an HPC AI Accelerator, has appeared online. Interestingly, specifications and features of the high-performance computer have leaked online as well. Although they might change as the test rig is apparently still under development, the platform appears to give competition to NVIDIA’s next-generation Ampere GPU based HPC.
AMD’s upcoming Radeon Instinct MI100 HPC accelerator, featuring the Arcturus GPU, has leaked online. Although AMD had long back confirmed the existence of Arcturus GPU, the details about the AMD’s next HPC/AI accelerator are just surfacing. Interestingly, the AMD Arcturus XL GPU could be a single huge monolithic die. In other words, it will not bear the new-generation chiplet based design like AMD’s Ryzen CPU lineup that runs on the ZEN 2 Architecture.
AMD Arcturus GPU Powered Radeon Instinct MI100 HPC / AI Accelerator Testbench Prototype Specifications And Features:
The AMD Arcturus GPU Powered Radeon Instinct MI100 HPC / AI Accelerator testbench is powered on ‘D34303’, which is a single, monolithic XL die. The Radeon Instinct MI100 test board reportedly has a TDP of 200W. The card also features 32 GB of HBM2 memory with pin speeds of 1.0 – 1.2 GHz. It is quite possible that the final design of the Arcturus GPU could pack Samsung’s latest HBM2E ‘Flashbolt’ memory which offers 3.2 Gbps speeds for up to 1.5 Tb/s of bandwidth.
According to the naming scheme, the Arcturus XL GPU-based Radeon Instinct MI100 would have a performance of around 100 TFLOPs of INT8. That’s a shocking 66 percent boost over the INT8 (AI/DNN) compute horsepower. While the FP16 compute might be rated at around 50 TFLOPs, 25 TFLOPs of FP32, the FP64 compute could be 12.5 TFLOPs. Needless to add, these are substantial jumps in the performance benchmarks, and could be attributed to the significantly updated graphics architecture, boosted clock speeds or simply due to the higher number of CUs.
The cache size of the AMD Arcturus GPU has been substantially increased, and also the CU count has doubled from 64 to 128. This is twice as many CUs as Vega 10. If AMD is using 64 Stream Processors per CU like their current and confirmed upcoming GPU designs, then that gives the HPC about 8192 Stream Processors.
AMD Arcturus GPU A Parallel Development Or Fork Of Vega Architecture?
It is quite likely that the AMD Arcturus is a Vega derivative. However, AMD will only release or develop the same for the High-Performance Computing segment. Although similar in design, the AMD Arcturus and Vega will remain independent of each other, and while the Arcturus will be developed and sold in the HPC market which consists of AI/DNN and datacenter customers, the Vega will be developed for the gaming/consumer/prosumer segment.
— Anthony Garreffa (@anthony256) November 6, 2018
It is apparent that AMD isn’t trying to go directly against the products that NVIDIA has for the HPC Segment. The NVIDIA’s next-generation Ampere GPU for this segment is rumored to offer up to 18 TFLOPs of FP64 compute. However, AMD could be altering the playing field by offering more FLOPs at a competitive price. Although AMD hasn’t officially acknowledged or confirmed anything, experts indicate the HPC market may get an AMD Instinct branded product in the second half of this year.