Next-generation computing system
Next-generation computing system
The ThreadRipper 3990X is the most advanced CPU available, yet supercomputers might not upgrade to it because it doesn’t match their current performance needs. Switching would require significant changes in architecture and infrastructure. Additionally, while a 3990X offers better value compared to top Xeon models, its power still falls short of what supercomputers typically demand.
You can't fit several 3990X chips on one motherboard, so you'll need significantly more room for the same number of processing units. Additionally, newer CPUs are far more powerful than the 3990X. Consider using AMD EPYC processors.
it depends on the specific components. Several models rely on Xeon Phi, an evolution from Larabee, while others are shifting toward greater GPU utilization, leveraging newer EPIC processors. a 3990x server model can fit into two motherboards.
Missing multi-socket capability. System may accommodate up to four EPYC CPUs or just one TR. Insufficient 8-channel memory availability. HEDT architecture typically restricts to four channels. Limited support for large memory capacities. EPYC can handle up to 2TB RAM, while TR cannot. Fewer PCI-E lanes compared to server components (128 vs 72 in this instance). Certain functionalities are absent—may involve proprietary instruction sets or hardware-assisted virtualization, ECC memory, etc., depending on the platform. Reliability matters. Server modules are built for extended, continuous operation, whereas consumer and HEDT units aren't designed for critical missions. Raw performance is only one aspect.
Supercomputers are specifically designed—budget plays a minor role. The upcoming Crays system will run on AMD EPYC processors and is expected to be the most powerful ever, with an estimated cost of around $600 million. https://en.wikipedia.org/wiki/Frontier_(supercomputer)
High performance at 1.5ghz is impressive. These cores definitely stand out. Are they truly exceptional?
EPYC versus Intel represents a distinct conversation. Supercomputers aren't meant for everyday tasks; their software is tailored specifically for them, often leveraging every performance boost available. Deep optimization for Intel differs significantly from AMD approaches. If you currently use an Intel system and later switch to AMD, developers must adjust the software to achieve optimal results. Longevity and reliability with a vendor also matter—purchasing large quantities of CPUs and GPUs tends to favor Intel due to proven stability over decades. Establishing trust with consumers takes time. Lastly, while Intel may cost more upfront, in server environments AMD doesn't always cause catastrophic failures, making it a viable alternative despite higher prices.