Key takeaways
- AMD said servers built on its 6th Gen EPYC processors, led by the EPYC 9996, are shipping from all leading server vendors as of Nov. 10.
- AMD claims a two socket EPYC 9996 server delivers up to 3.7x the performance of an Nvidia Vera system and up to 2.2x that of an Intel Xeon 6980P system.
- The 128 core EPYC Embedded 9006 model goes on sale Nov. 15, and AMD expects the remaining embedded models to ship by the end of the year.
In this story 4 sections
Advanced Micro Devices (NASDAQ: AMD) made its 6th Gen AMD EPYC server processors generally available on Nov. 10, led by the AMD EPYC 9996, and said systems built on the new "Zen 6" design are already shipping from every major server vendor. The company is aiming the lineup, known by the codename Venice in its press kit, at AI data centers where CPUs now run software agents, feed GPUs and handle some inference on their own.
Cloud access is the next step. AMD said instances from Microsoft (NASDAQ: MSFT) on Azure and from Oracle (NYSE: ORCL) on Oracle Cloud Infrastructure are coming soon, and an Azure HDv2 instance is already in preview.
What AMD claims for the EPYC 9996
All of the comparisons below come from AMD's launch release and its footnotes, and each one pits two socket servers against each other:
- Up to 3.7x the performance of a server using the Vera CPU from Nvidia (NASDAQ: NVDA) on the SPECrate2017_int_base test. AMD said the Vera result is its own estimate.
- Up to 2.2x the performance and 2x the performance per dollar of a server with a Xeon 6980P from Intel (NASDAQ: INTC).
- Roughly 1.18x as many AI agents per rack within a 35 kilowatt power budget when a new power setting is applied, which AMD also labeled an estimate.
AMD lists a dual socket EPYC 9996 server at 512 cores, or 256 cores per chip by our math (512 divided by 2). Its 5th Gen EPYC 9965 setup is listed at 384 cores, or 192 per chip. That puts the per socket gain at 33% ((256 minus 192) divided by 192).
Power climbs too. AMD rates the 9996 at 600 watts and the 9965 at 500 watts, using a new "Default CPU Power" figure that replaces the TDP label for this generation.
For GPU servers, AMD is selling high frequency host processors that run at up to 5.3 GHz. It projects they could deliver up to 1.8x the large language model output throughput of earlier EPYC chips in a setup with one Nvidia B200 GPU, and it described that number as a projection tied to doubled PCIe bandwidth rather than a measured result.
Helion racks built around agents
Alongside the chips, AMD introduced Helion, a set of rack designs for agentic AI that comes in three versions: Helion Liquid, Helion Air and Helion Air Inference. Helion Liquid is designed to stand beside AMD's Helios rack scale system. Some Helion configurations hold up to 73,728 threads in one rack, the company said.
AMD also claimed its high density parts fit about 3.3x more agents per CPU watt than Nvidia Vera and about 2.5x more than the Xeon 6980P in agent sandbox tests.
"AI is significantly expanding the role of the CPU across every layer of the data center," said Dan McNamara, senior vice president and general manager of Compute and Enterprise AI at AMD.
Early customers and partners
Meta Platforms (NASDAQ: META) is the lead customer and has been validating platforms ahead of a large deployment, according to AMD. The release names three Azure instance families from Microsoft: HDv2, HXv2 and ND MI455X v7, the last of which is based on Helios. AMD also said OCI Compute E7 instances from Oracle will use the new processors.
The 6th Gen EPYC press kit lists Lenovo, Quanta Cloud Technology, Hyve Solutions and Cornelis among the hardware partners with launch news.
Embedded chips start shipping Nov. 15
A separate embedded announcement covers the EPYC Embedded 9006 Series for networking gear, storage systems and test equipment. Those chips range from 64 to 256 cores and support 16 DDR5 memory channels, up to 96 PCIe Gen 6 lanes and CXL 3.1.
The 128 core model will be available starting Nov. 15, with the other models expected to begin shipping by year end, AMD said. The company promises seven years of supply for the embedded line and named Samsung Electronics, Lanner, Portwell and IEI Integration as early users.
AMD said the embedded parts deliver up to 1.8x the throughput of the 5th Gen EPYC 9005 Series, citing two socket SPECrate2017_int_base scores of 4,900 and 2,740. By our calculation that is a 79% gain ((4,900 minus 2,740) divided by 2,740).
The releases did not give a manufacturing process or list prices for the new processors.
The launch adds another option for AI buyers weighing custom silicon such as Google's TPU 8tx and 8i Max. More chip and data center coverage is on our tech page.
Sources
Share this story
Spotted an error? Tell us


