AMD Takes Aim At Intel And Nvidia With New Sixth Generation EPYC Server And Embedded Processors - SMBtech

The new Venice lineup includes the EPYC 9996, the highest-core-count model with 512 cores, a default CPU power rating of 600W, and a per-unit price of $14,904 in quantities of 1,000.

AMD is positioning the 6th Gen EPYC generation squarely around the demands of agentic AI, arguing that a single agentic workflow can span running agents, calling models, retrieving data, executing code and interacting with databases and applications.

The company has structured its portfolio around three distinct CPU roles within AI infrastructure: general-purpose services, GPU host nodes, and agent sandboxes.

For GPU host nodes, AMD offers high-frequency processors capable of reaching up to 5.3 GHz, which the company says can deliver up to 1.8 times the LLM output throughput of previous-generation EPYC processors in CPU-offload inference scenarios.

For agent sandboxes, AMD claims its high-density processors provide approximately 3.3 times the agents-per-CPU-watt over Nvidia Vera and approximately 2.5 times over Intel Xeon 6980P.

The EPYC 9996 scored 5,470 on the SPECrate 2017 integer benchmark, which AMD says represents up to 3.7 times the performance of an Nvidia Vera CPU-based system and up to 2.2 times the performance of an Intel Xeon 6980P system.

On a per-dollar basis, AMD claims the EPYC 9996 delivers up to 2.0 times the performance per dollar of the Intel Xeon 6980P.

Further down the stack, the EPYC 9965 offers 384 cores at 500W with a 1,000-unit price of $11,988, scoring 3,260 on the same benchmark.

For comparison, Intel’s Xeon 6990E+ (576 cores, 450W, $14,995) scored 3,050, while the Xeon 6980P (256 cores, 500W, $13,955) scored 2,510, according to published SPEC results.

The company has also detailed how power management can increase agent density at the rack level.

Using a feature called Unified Power Provisioning (UPP), EPYC 9996-based server racks operating within a 35-kilowatt power envelope can fit an estimated 18.8 servers per rack, compared with 16 servers using default power management settings.

AMD says this translates to approximately 1.18 times as many agents per rack, with an estimated average performance reduction of just one per cent at the more aggressive power setting.

Cloud and ecosystem support

Microsoft Azure is bringing the processors across its AI infrastructure stack.

Azure HDv2 instances are now in preview for AI data processing, agentic AI and HPC workloads.

Azure HXv2 instances target electronic design automation and technical computing.

And Azure ND MI455X v7, based on AMD’s Helios rack-scale solution, is designed for production-scale AI inference.

OCI will power its Compute E7 instances with 6th Gen EPYC processors, supporting enterprise applications and the CPU services surrounding AI models.

Meta is working with AMD as a validation partner and has been preparing 6th Gen EPYC-based platforms for deployment at scale.

Beyond individual servers, AMD is extending its reach to rack-scale deployments with the Helion agentic AI rack platform.

Three configurations are available: Helion Liquid, designed for large-scale deployments alongside AMD’s Helios rack-scale solution; Helion Air, for expanding agent capacity within existing air-cooled data centres; and Helion Air Inference, which combines CPU execution and local GPU inference in a single rack.

AMD has worked with server ecosystem partners on high-density Helion rack configurations, with some reaching up to 73,728 threads in a single rack.

Embedded processors join the lineup

Alongside the server launch, AMD has introduced the EPYC Embedded 9006 Series SP7 processors, bringing the Zen 6 architecture to embedded applications.

The embedded line offers up to 256 cores and 512 threads, PCIe Gen 6 and CXL 3.1 connectivity and seven-year product longevity.

AMD is targeting semiconductor automated test equipment (ATE), enterprise storage and network firewall applications.

The embedded processors deliver up to 1.8 times the throughput of the previous-generation EPYC 9005 Series, according to AMD’s benchmarking.

They also offer up to 1.5 times higher I/O bandwidth and up to 2.7 times higher memory bandwidth compared with the prior generation.

Sixteen DDR5 memory channels support RDIMM speeds up to 8,000 megatransfers per second (MT/s) and MRDIMM speeds up to 12,800 MT/s.

Up to 96 PCIe Gen 6 lanes operate at 64 gigatransfers per second.

For network security applications, AMD claims the embedded processors deliver up to 1.3 times higher firewall throughput than the previous generation, measured using DPDK L3 forwarding benchmarks.

In enterprise storage, the processors achieved up to 1.5 times the performance on the Intelligent Storage Acceleration Library (ISA-L) benchmark, which measures how quickly a processor handles functions used in storage services such as compression, caching and RAID.

For semiconductor test equipment, the processors support production-scale validation of new memory and I/O devices, managing workloads across DDR5 RDIMM and emerging MRDIMM technology as well as PCIe Gen 6 and CXL 3.1 devices.

The embedded processors incorporate multiple layers of hardware security.

AMD Root of Trust (ART) establishes an immutable silicon root of trust and enables secure boot, with support for post-quantum cryptography for firmware signing intended to guard against future quantum computing threats.

The AMD Security Processor (ASP) acts as the boot orchestrator and provides runtime security services, including a trusted execution environment and secure debug capabilities for authorised diagnostics.

The processors also integrate Caliptra, an open-source silicon root of trust standard used for device identity and attestation, allowing remote verification that a device is genuine and running expected firmware.

Seven-year product longevity is a central feature of the embedded line, aimed at customers who require long-term support for mission-critical deployments.

Ongoing embedded-focused firmware updates and software support are included.

Partners highlight integration possibilities

Samsung Vice President of Memory Product Planning Jangseok Choi highlighted the integration possibilities with Samsung’s memory and storage products.

We look forward to working with AMD to bring these capabilities into our manufacturing test infrastructure.

Lanner Associate Vice President of Product Planning Fred Chang pointed to the collaboration around network appliances.

What excites me most is seeing what our customers will build with it, from tackling today’s toughest challenges to creating new possibilities for tomorrow’s networks.

IEI Integration Corp Vice President YT Lee underlined the benefits for storage device testing.