Getting access to a 16-core processor for under $650 is an impressive feat, but that configuration becomes trivial when compared to the firepower that some supercomputers demand. The case of the Cerebras Andromeda stands out not only for its staggering number of cores, but for the type of work it does, aiming directly at artificial intelligence.
One of the most common problems in computing is scaling, or rather, its lack. I mean throwing twice the hardware at a problem and expecting to solve it in half the time, only to discover that things don't work that way. But parallelism has improved enormously, allowing us to reach extraordinary performance levels. Not to go too far, the Frontier supercomputer was the first to cross the exaFLOP barrier, adding almost nine million cores across processors and GPUs.
With that in mind, we arrive at the gates of Cerebras, which has just officially unveiled its Andromeda supercomputer. The first thing that surprises about Andromeda is the total number of cores, 13.5 million to be precise. However, that doesn't mean it beats Frontier in performance. In fact, we are dealing with two different creatures, both in application and cost. While Frontier can simulate the lifetime of a nuclear reactor or tackle genetics problems, Andromeda's priority is artificial intelligence.
Cerebras Andromeda: Nearly Perfect Scaling
The official Andromeda page dedicates an entire section to the supercomputer's scaling capability in "large language models" (read: LLM) such as GPT-3, GPT-J, and GPT-NeoX. Its configuration is based on 16 Cerebras CS-2 systems, delivering between 14.5 and 15.8 times the performance of a single CS-2. Cerebras speaks of more than an exaFLOP in "artificial intelligence operations", and 120 petaFLOPs of "dense operations", combining a huge number of cores, RAM, and solid-state storage.
Another very interesting aspect of Andromeda is the design of its processor, the WSE-2. It is a 'wafer-scale' chip that uses almost the entire silicon wafer and has three physical planes. The first is for communications, the second accumulates 40 gigabytes of SRAM (much more expensive and faster than the DRAM in our computers), and the third is for arithmetic operations, with 850,000 independent cores and 3.4 million floating-point units. Each of the nodes is connected to a backplane with a bandwidth that reaches 96.8 terabits per second.
And as if that were not enough, Andromeda also has a bank of AMD EPYC processors (18,176 cores in total) that prepare and coordinate the data to be analyzed. How much does all this hardware cost? In money, $35 million for its construction. In energy, approximately 500 kilowatts under full load.
Official site: Click here Source: ExtremeTech.
