On eBay, a third-party seller is, or was, selling an AMD Instinct MI100 GPU for under €500, as found by Twitter user "Madness!" (@Madness727). This particular model is without a cooler and does not work. However, it does allow the viewer to see the internal design of the Arcturus processor-based graphics accelerator.
Cooler-less AMD Instinct MI100 GPU accelerator appears on eBay, selling for 500 EUR, or around 494 USD
In November 2020, the AMD Instinct MI100 launched for $6,400. This card was considered the world's fastest high-performance computing graphics card at the time. The MI100 was created to offer the most efficient double-precision computing capabilities while simultaneously transmitting extreme GPU performance, focusing on AI workloads. AMD/s MI100 utilized the CDNA architecture to rival the NVIDIA Ampere A100 accelerator.
The AMD Instinct MI100 was designed on the 7nm CDNA GPU architecture, offering 120 Compute Units (7,680 stream processors). The CDNA GPU measures around 720 mm² and offer a clock speed of 1500 MHz with a power consumption of 300W. The cooler was considered passive, making sense due to the large amounts of data passed through the accelerator. The peak workload performance of the Arcturus GPU is broken down at:
- 11.5 TFLOPs (FP64)
- 23.1 TFLOPs (FP32)
- 185 TFLOPs (FP16)
The Instinct MI100 accelerator offered 32 GB of HBM2 memory and used the Infinity Fabric X16 interconnect to communicate to the board's four and eight GPU configurations (bandwidth rated at 276 GB/s). AMD board partners for the Instinct MI100 included HPE, Dell, Gigabyte, and Supermicro.
Since this card no longer works, it is considered a collector's item by a niche category of technology enthusiasts.
AMD's MI100 successor, the MI250 series, has already been released. Still, the company is currently working on new designs for the next generation of CDNA-based GPU architecture, which is slated to offer both the CDNA and Zen architectures on one board design. It is anticipated that the company will release the third generation of accelerators in 2023 to compete with NVIDIA's current AI-focused components.