# AMD EPYC Venice

> Source: https://aiwiki.ai/wiki/amd_epyc_venice
> Updated: 2026-06-07
> Categories: AI Hardware
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

**AMD EPYC Venice** is the codename for [AMD](/wiki/amd)'s sixth-generation [EPYC](/wiki/epyc) server processor, built on the [Zen 6](/wiki/zen_architecture) core architecture and positioned as the successor to the fifth-generation EPYC "Turin" (Zen 5). Venice is notable on two counts: AMD describes it as the first high-performance computing (HPC) product in the industry to be manufactured on [TSMC](/wiki/tsmc)'s N2 (2nm-class) [process node](/wiki/semiconductor), and it serves as the host CPU of AMD's "Helios" rack-scale [AI](/wiki/ai_chip) system, where it is paired with Instinct MI400-series GPUs. AMD says Venice delivers more than 70% higher performance and efficiency than Turin, more than 30% greater thread density, and scales up to 256 cores per socket. It is expected to reach the market in the second half of 2026. Venice is a server CPU and should not be confused with AMD's Instinct accelerators, which are the GPUs it hosts.

## Overview

Venice is the CPU half of AMD's 2026 [data center](/wiki/data_center) platform refresh. In AMD's product hierarchy it is the general-purpose host and orchestration processor, while the heavy AI math runs on the Instinct GPUs. AMD first attached the Venice name to a manufacturing milestone in April 2025, when it announced that the chip was the first HPC product to be taped out and brought up on TSMC's N2 node. It previewed Venice as the host CPU for its [Helios](/wiki/amd_helios_rack) AI rack in June 2025, displayed working silicon publicly for the first time at CES 2026, and announced the start of volume production in May 2026.

The headline figures come from AMD itself: more than 70% higher performance and efficiency versus the current EPYC Turin generation, more than 30% more thread density, and more than a doubling of per-socket memory bandwidth to roughly 1.6 TB/s. Finer details such as the chiplet layout, core-per-die counts, socket dimensions, and memory speeds come from leaks, engineering samples, and hands-on press coverage rather than full official datasheets, and are attributed as such below. AMD has also named a companion sixth-generation part, codenamed Verano, that it positions for performance-per-dollar-per-watt leadership using LPDDR memory for power-constrained workloads.

## Zen 6 architecture and process

Venice is the server implementation of AMD's Zen 6 microarchitecture. As with prior EPYC generations, AMD ships two core flavors: a standard Zen 6 core tuned for clock speed and per-thread performance, and a density-optimized Zen 6c core that trades clock headroom for more cores in the same area. The two share an instruction set and socket but use different core complex dies (CCDs).

The compute chiplets are fabricated on TSMC's N2 process, the foundry's first 2nm-class node to use gate-all-around (nanosheet) transistors. AMD's claim to being the first HPC product on N2 is specific: Apple and others were expected to use N2 for mobile and consumer silicon, but AMD says Venice is the first high-performance computing product to tape out, bring up, and then enter volume production on the node. Reporting indicates the standard Zen 6 CCD carries 12 cores (up 50% from the 8-core Zen 5 CCD) with around 48 MB of L3 cache, while the dense Zen 6c CCD packs 32 cores; both reportedly keep a die area close to their Zen 5 predecessors, with the density gains coming largely from the N2 shrink.

Press analysis of the CES 2026 hardware describes a significant packaging change. Rather than routing CCD-to-I/O traffic through the organic package substrate as earlier EPYCs did, Venice appears to use a more advanced 2.5D-class package with two central I/O dies flanked by up to eight CCDs, a layout closer to AMD's MI-series accelerators than to past server CPUs. AMD has not published the full packaging details, so this remains reported rather than confirmed.

## Specifications (cores, memory, sockets)

A top-end Venice part combines eight 32-core Zen 6c CCDs for up to 256 cores and 512 threads, while standard Zen 6 configurations using 12-core CCDs reach up to 96 cores. The platform moves to a new socket family: the high-end SP7 socket (reported at about 123.6 x 100.6 mm, roughly 12% larger than today's SP5) for flagship Venice parts, and a smaller SP8 socket (about 123.9 x 80.9 mm) aimed at entry and edge systems and the value-oriented Verano line. AMD says per-socket memory bandwidth more than doubles to about 1.6 TB/s, up from roughly 614 GB/s on Turin; reporting attributes this to a shift from 12 to 16 DDR5 channels with support for fast DDR5 and MRDIMM modules. The I/O complex adds PCIe Gen6 and CXL 3.x connectivity.

The table below summarizes the configuration. Figures marked "AMD" are from AMD statements; figures marked "reported" come from leaks, engineering samples, or press coverage and should be treated as provisional until launch.

| Specification | EPYC "Venice" (6th Gen) |
| --- | --- |
| Microarchitecture | Zen 6 (standard) and Zen 6c (dense) |
| Compute die process | TSMC N2 (2nm-class), AMD |
| Max cores, standard Zen 6 | Up to 96 cores via 12-core CCDs (reported) |
| Max cores, dense Zen 6c | Up to 256 cores / 512 threads via 32-core CCDs (AMD: 256 cores) |
| Compute dies | Up to 8 CCDs around two central I/O dies (reported) |
| L3 cache | ~48 MB per standard CCD; ~128 MB per dense CCD, ~1 GB on full dense part (reported) |
| Memory | 16-channel DDR5; up to DDR5-8000, MRDIMM-12800 (reported) |
| Peak memory bandwidth | ~1.6 TB/s per socket (AMD), vs ~614 GB/s on Turin |
| I/O | PCIe Gen6 and CXL 3.x; up to 128 PCIe lanes in 2P (reported) |
| Sockets | SP7 (high-end); SP8 (entry/edge and Verano) (reported) |
| Performance vs Turin | More than 70% higher performance and efficiency; more than 30% more thread density (AMD) |
| Expected launch | Second half of 2026 |

## Role in the Helios rack and AI servers

Venice's most visible role is as the host CPU of AMD's [Helios](/wiki/amd_helios_rack) rack-scale AI platform, the company's answer to integrated GPU racks from [Nvidia](/wiki/nvidia). Helios is a double-wide, liquid-cooled rack built to the Open Compute Project's Open Rack Wide specification. It combines 72 [Instinct MI400-series](/wiki/amd_mi400) GPUs (the AI-tuned MI455X variant) with Venice CPUs acting as the hosts, and AMD-Pensando "Vulcano" AI network cards. Each compute tray pairs a Venice CPU with the accelerators it feeds, so a full rack uses roughly 18 Venice CPUs alongside the 72 GPUs, with the CPUs handling the operating system, data staging, scheduling, and the parts of a workload that are not offloaded to the GPUs.

In this design the CPU and GPU play complementary roles. The MI455X carries the dense matrix math, the 31 TB of HBM4, and up to roughly three FP4 exaflops of compute per rack, while Venice supplies the high-bandwidth memory, PCIe Gen6 connectivity, and many-core throughput needed to keep those accelerators fed. The scale-up fabric between GPUs uses [UALink](/wiki/ualink)-class interconnect, and scale-out networking uses [Ultra Ethernet](/wiki/ultra_ethernet). Helios succeeds AMD's earlier MI300- and [MI355X](/wiki/amd_instinct_mi355x)-based systems and is the platform AMD is positioning against Nvidia's [Vera Rubin](/wiki/nvidia_vera_rubin) racks. Helios systems are targeted for the second half of 2026, aligning with Venice availability.

## Launch timeline

- April 14, 2025: AMD announces that Venice is the first HPC product in the industry taped out and brought up on TSMC's N2 (2nm) node; CEO Lisa Su shows a Venice CCD die publicly.
- June 12, 2025: At its Advancing AI event, AMD previews the Helios rack and names sixth-generation EPYC "Venice" as its host CPU alongside Instinct MI400 GPUs.
- October 14, 2025: AMD shows Helios as a reference design at the Open Compute Project Global Summit.
- Late 2025: On AMD's quarterly earnings call, Lisa Su says Venice silicon is performing well and will launch with Instinct MI400 in 2026.
- January 2026: At CES 2026, AMD displays Venice silicon, the Instinct MI455X, and Helios hardware publicly for the first time.
- May 2026: AMD announces the start of Venice production ramp on TSMC's N2 process, calling it the first HPC product to enter volume production on 2nm. Lisa Su frames it as "an important step forward in accelerating the next gen of AI infrastructure."
- Second half of 2026: Expected commercial launch and availability, in step with Helios systems.

## Competition versus Intel and Arm

Venice arrives into the most crowded server-CPU field in two decades, with viable x86 and Arm options all targeting AI [data center](/wiki/data_center) sockets.

On the x86 side, AMD's rival is [Intel](/wiki/intel). Intel's density-oriented answer is Xeon 6+ "Clearwater Forest," launched in June 2026 on the Intel 18A process with up to 288 efficiency (E) cores and 576 MB of L3 cache; it lacks simultaneous multithreading, so its 288 physical cores line up against Venice's 256 Zen 6c cores with 512 threads. Intel's performance-core flagship, Xeon 7 "Diamond Rapids" on 18A-P, brings up to about 192 P-cores, 16-channel memory, and PCIe Gen6, but is not due until 2027. That cadence gives Venice a roughly one-year lead over Intel's comparable P-core part, while Clearwater Forest contests the density segment in the same window.

On the Arm side, the competitive set has broadened sharply. Nvidia pairs its own Arm-based "Vera" CPU (88 custom "Olympus" cores) with Rubin GPUs in its racks, the direct analog of Venice's role inside Helios. Ampere Computing, now owned by Arm's parent SoftBank, offers AmpereOne at up to 192 cores with a higher-core "Aurora" part on its roadmap. Arm itself has begun shipping its first in-house production silicon, a data center CPU of up to 136 Neoverse cores co-developed with Meta. Against these, AMD's pitch for Venice rests on x86 compatibility, very high core and thread counts, a large installed EPYC base, and the timing advantage of being first to high-volume 2nm. The competitive question is less about raw core count, where several vendors now exceed 200 cores, and more about per-core performance, memory bandwidth, total cost, and how tightly each CPU integrates with the GPUs in a rack.

## References

1. AMD, "AMD Achieves First TSMC N2 Product Silicon Milestone" (April 14, 2025). https://ir.amd.com/news-events/press-releases/detail/1245/amd-achieves-first-tsmc-n2-product-silicon-milestone
2. AMD, "AMD Announces Production Ramp of Next-Generation AMD EPYC Processor 'Venice' on TSMC 2nm Process Technology" (May 2026). https://www.amd.com/en/newsroom/press-releases/2026-5-20-amd-announces-production-ramp-of-next-generation-a.html
3. Tom's Hardware, "AMD begins production ramp of 256-core EPYC Venice -- first 2nm HPC chip claims 70% performance leap." https://www.tomshardware.com/tech-industry/semiconductors/amd-begins-production-ramp-of-256-core-epyc-venice-on-tsmcs-2nm-node
4. Tom's Hardware, "AMD EPYC Venice boasts 256 cores and bandwidth galore -- next-gen server CPUs arrive in 2026." https://www.tomshardware.com/pc-components/cpus/amds-256-core-epyc-venice-cpu-in-the-labs-now-coming-in-2026
5. Tom's Hardware, "AMD's Enterprise CPU and GPU roadmap: Venice, Verano, Zen 6, Helios, and CDNA." https://www.tomshardware.com/pc-components/cpus/amds-enterprise-cpu-and-gpu-roadmap-venice-verano-zen-6-helios-and-cdna
6. ServeTheHome, "AMD's EPYC Venice, Instinct MI455X, & Helios Hardware On Display for First Time at CES 2026." https://www.servethehome.com/amds-epyc-venice-instinct-mi455x-helios-hardware-on-display-for-first-time-at-ces-2026/
7. Chips and Cheese, "CES 2026: Taking the Lids off AMD's Venice and MI400 SoCs." https://chipsandcheese.com/p/ces-2026-taking-the-lids-off-amds
8. Wccftech, "AMD SP7 & SP8 Platforms For Next-Gen EPYC 'Venice' & 'Verano' CPUs Detailed." https://wccftech.com/amd-sp7-sp8-platforms-epyc-venice-verano-cpus-12800-mtps-16-channel-memory-128-pcie-6-0-lanes/
9. TechRadar, "Helios rack design to support MI400 GPU and PCIe Gen6 EPYC Venice chips." https://www.techradar.com/pro/amd-gets-ready-for-nvidias-vera-rubin-and-2026-with-432gb-mi400-gpu-monster-paired-with-256-core-epyc-venice-and-i-cant-wait-to-see-the-sparks-fly
10. TweakTown, "AMD's next-gen Zen 6 EPYC Venice CPU: over 70% more performance, 30% more thread density." https://www.tweaktown.com/news/108874/amds-next-gen-zen-6-epyc-venice-cpu-over-70-percent-more-performance-30-percent-more-thread-density/index.html
11. Igor's Lab, "AMD Zen 6 leak cites 12 cores per CCD and 48 MB of L3 cache with virtually unchanged chiplet area." https://www.igorslab.de/en/amd-zen-6-leak-cites-12-cores-per-ccd-and-48-mb-of-l3-cache-with-virtually-unchanged-chiplet-area/
12. Tom's Hardware, "AMD Zen 6 'Venice' ES chips break cover with up to 192 cores, 32 per CCD." https://www.tomshardware.com/pc-components/cpus/amd-zen-6-venice-es-chips-break-cover-with-up-to-192-cores-32-per-ccd-in-early-stress-test-kenya-congo-nigeria-platforms-leaked
13. DataCenterDynamics, "AMD unveils full MI400 product lineup, claims MI500 chips will deliver 1,000x increase in AI performance." https://www.datacenterdynamics.com/en/news/amd-unveils-full-mi400-product-lineup-claims-mi500-chips-will-deliver-1000x-increase-in-ai-performance/
14. Tom's Hardware, "Intel Xeon 6+ 'Clearwater Forest' puts 18A in the data center with up to 288 cores, 576 MB of L3 cache." https://www.tomshardware.com/pc-components/cpus/intel-xeon-6-clearwater-forest-puts-18a-in-the-data-center-with-up-to-288-cores-576-mb-of-l3-cache-new-xeon-6990e-is-30-percent-faster-per-thread-than-192-core-amd-epyc-9965-says-intel
15. Tom's Hardware, "Intel Xeon 7 'Diamond Rapids' CPUs officially launching in 2027 on Intel 18A-P." https://www.tomshardware.com/pc-components/cpus/intel-xeon-7-diamond-rapids-cpus-officially-launching-in-2027-on-intel-18a-p-next-gen-p-core-xeon-features-pcie-6-0-50-percent-higher-core-counts-and-twice-the-memory-bandwidth
16. Tom's Hardware, "Arm moves beyond IP with AGI CPU silicon -- 136-core data center chip targets AI infrastructure with Meta as lead partner." https://www.tomshardware.com/tech-industry/semiconductors/arm-launches-its-first-data-center-cpu
17. Tom's Hardware, "Nvidia unveils details of new 88-core Vera CPUs positioned to compete with AMD and Intel." https://www.tomshardware.com/pc-components/gpus/nvidia-unveils-details-of-new-88-core-vera-cpus-positioned-to-compete-with-amd-and-intel-new-vera-cpu-rack-features-256-liquid-cooled-chips-that-deliver-up-to-a-6x-gain-in-cpu-throughput
18. HPCwire, "AMD Tapes Out 1st HPC Product on TSMC 2nm Process with 'Venice' EPYC CPU." https://www.hpcwire.com/off-the-wire/amd-tapes-out-1st-hpc-product-on-tsmc-2nm-process-with-venice-epyc-cpu/