Interesting fact: Aldebran is the brightest star in the constellation Taurus. It is 44 times the size of the sun and 400 times the size of light. It is also called the upcoming AMD Instinct Accelerator.
Aldebaran, the PC part, has appeared several times in Linux patch notes. The two-level scheme was implicitly implemented in February and confirmed last month: "In Aldebaran, only the primary format receives valid power data. Displays power/power values as 0 in the secondary format",> This is the first AMD/GPU accelerator to use an MCM design (Multi-Layer Module), a method that has been patented for years but only recently achieved. The use of multiple closely-connected chips/blocks allows for greater scalability than traditional integrated designs, but also reduces the performance of the computed module on the card, especially in workloads that are not parallel to each other. Next year build your own MCM designs based on the Hopper architecture. Meanwhile, Intel is about to launch a Ponte Vecchio accelerator based on the Xe HPC architecture and the initial implementation of MCM.
The patch notes indicate that Aldebaran, like Ponte Vecchio, uses MCM to expand memory capacity. A note from last week described Aldbaran as having two formats, four UMCs per die and eight channels connected to 2 Gb HBMs per UMC. Or 128 GB in total.
AMD is an amazing MI100 virus. The patch notes also mention support for a "new type of HBM2 memory", possibly HBM2e. If AMD were to use the new HBM2e SK Hynix memory modules with 3.6 Gbps bandwidth and a 4,096-bit bus per format, Aldebaran could have 3.6 TB/sec of memory bandwidth.
For comparison, the Instinct MI100, precursor to Aldebaran (codenamed Arcturus, another star), has 32GB of HBM2 and 1.2TB of memory bandwidth. Nvidia's A100 Accelerator can be configured with up to 80GB of HBM2e memory at 3.2Gbps for a bandwidth of 2TB. Dr. Lisa Su, CEO of AMD, said earlier that the CDNA2 architecture, i.e. Aldebaran, will be released later this year. It will likely be released as the MI200 Accelerator.
“This year, as I said, we are building the next generation of CDNA architecture. This is a key component that will enable us to win the largest supercomputer. Showcases in the United States,” Su said at a conference in May. "...and we'll be launching the next generation of this architecture later this year. We're very excited about this. I think it's made a lot of progress."
Distinguished Teacher by Timothy Dix
The next AMD GPU can have two disks and 128GB of memory