32,768 cores and 74 petaflops: AMDs monster new rack for AI agents could be too much for traditional data centers

Date:
Fri, 16 Oct 2026 20:35:00 +0000

Description:
AMD's Helion Agentic AI Rack combines high-core-count processors, accelerators, and programmable networking, potentially requiring data centers to upgrade existing infrastructure.

FULL STORY ======================================================================Copy link Facebook X Whatsapp Reddit Pinterest Flipboard Threads Email Share this article 0 Join the conversation Follow us Add us as a preferred source on Google Newsletter Subscribe to our newsletter AMD's proposed rack combines 32,768 CPU cores with 16 AI accelerators Traditional data centers may need
new power and cooling infrastructure AMD divides agent workloads into personal, enterprise, and technical computing categories AMD is developing a rack-scale computing system for artificial intelligence agents, with configurations designed for personal, enterprise, and technical workloads.

Dubbed Helion Agentic AI Rack, it supports up to 32,768 Zen 6 x86 CPU cores, 16 Instinct MI350P accelerators, and 74 petaflops of FP4 computing performance. The proposed platform also supports 102 Tbps of networking bandwidth and uses modular hardware designed for different agent computing requirements. Latest Videos From TechRadar Watch full video here: A rack
built around three kinds of AI agents AMD groups agent workloads into three categories: personal assistants, enterprise applications, and technical computing, each with different processing requirements.

Personal agents need dense, energy-efficient servers; enterprise applications require a balance between agent capacity and performance; while technical workloads prioritize processing speed. You may like 'Same compute, fewer resources - More compute, same energy': AMD says it is on track to make AI four times more energy efficient AMD Threadripper Halo Station is probably
the world's most expensive desktop PC thanks to its 2.7TB memory and four AI GPUs a GB300 rival likely to cost $150,000 with just 1 CPU Meta enlists tiny Korean startup to build 'one-chip-like datacenter' CXL architecture introduced by Facebook's parent company and Panmnesia can handle almost 1000 AI GPUs per domain

AMD allows operators to customize CPU core density, processing performance, and power consumption according to their workloads.

The hyperscale configuration uses CPU-only infrastructure, while the enterprise version combines processors and graphics accelerators within an integrated rack. Are you a pro? Subscribe to our newsletter Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed! Contact me with news and offers from other Future brands Receive email from us on behalf of our trusted partners
or sponsors By submitting your information you agree to the Terms &
Conditions and Privacy Policy and are aged 16 or over.

For personal agents, the system uses 256-core Venice processors operating at 400W, prioritizing processing density and power efficiency.

Enterprise workloads use the 256-core Venice processors rated at 600W, while technical computing uses 96-core, high-frequency Venice processors.

The enterprise configuration separates agent execution, inference, and general-purpose computing into distinct server nodes. What to read next AMD's Venice is the first 2nm x86 server chip, but the stack doesn't fully ship until 2027 AMD's Helios beats Nvidia on memory at rack level and on compute only if you count it at the GPU level Nvidia paid $20 billion for SRAM decode - AMD just partnered for it instead

CPU sandbox nodes handle agent execution, Instinct MI350P accelerators
support inference, and general-purpose servers accommodate databases,
storage, and technical computing.

AMD also includes an integrated token router for directing agent model requests across frontier and open-source models.

The rack supports programmable networking, an open-source software ecosystem, and ORv3 44OU hardware compliant with Open Compute Project specifications.
Why traditional data centers may need substantial changes Traditional data centers are built for standard cloud or legacy enterprise computing, not
racks packing up to 32,768 Zen 6 CPU cores alongside heavy GPU
configurations.

Therefore, facilities built around existing server arrangements may lack the appropriate combination of power delivery, cooling capacity and network infrastructure for these configurations.

Workload diversity creates another challenge because personal agents, enterprise applications and technical computing require different processing characteristics.

AMD's proposed rack mixes dense 256-core processors with 96-core high-frequency systems, requiring infrastructure that can accommodate different server types and operating requirements.

This creates a planning challenge for data centers, which must allocate rack space, power, and cooling to different server configurations instead of deploying identical systems.

Operators may need to modify existing infrastructure if it cannot support the selected combination of high-core-count and high-frequency servers.

Networking introduces an additional consideration because the rack supports
up to 102 Tbps of bandwidth and integrates a programmable token router.

Existing facilities may need compatible networking equipment and software to support this throughput and route model requests across different AI systems.

The Helion Agentic AI Rack is scheduled for availability in November 2026,
but AMD has not disclosed complete facility upgrade costs or cooling requirements. Follow TechRadar on Google News and add us as a preferred
source to get our expert news, reviews, and opinion in your feeds.



======================================================================
Link to news story: https://www.techradar.com/pro/32-768-cores-and-74-petaflops-amds-monster-new-r ack-for-ai-agents-could-be-too-much-for-traditional-data-centers


--- Mystic BBS v1.12 A49 (Linux/64)
* Origin: tqwNet Technology News (1337:1/100)