# EPYC Venice 9006, 512 threads and new hardware with SP7 socket

> AMD opens orders for the EPYC Venice 9006 with 256 cores, 1.6 TB/s and SP7 socket, while Intel leaves Diamond Rapids for 2027.
> By José Maldonado · Technology · 2026-11-21

AMD has begun accepting orders for the EPYC Venice 9006 processors, the generation it announced in the summer and which now arrives with list prices. The top of the family is 256 Zen 6c cores and 512 threads on a single chip, double the 256 threads of Intel's Xeon 6, with 16 memory channels reaching 1.6 TB/s versus the 614 GB/s of the EPYC 9005 Turin, according to the [EPYC Venice 9006 datasheet](https://www.techpowerup.com/353209/amd-publishes-full-epyc-9006-venice-specs-and-pricing-8-core-at-usd-700-256-core-at-usd-14-904). The price of that leap is a new socket, SP7, and up to 600 watts per processor.

That last figure decides more purchases than the core count. A server with SP7 is not upgraded: the entire platform is replaced, including the motherboard, cooling and probably the chassis. For a company that bought Turin in 2025 and has its workloads covered, the reasonable calculation is to wait for SP8, which will arrive in 2027, and AMD knows it, because much of Venice's sales pitch lies in AI racks rather than generic server replacement.

The comparison with Intel is not the usual one either. While AMD opens orders, the Xeon 7 Diamond Rapids keeps piling up delays: Intel pushed its launch to 2027 in August and brought no demonstration of its own to the SC26 trade show. The first public samples of Diamond Rapids were seen at other companies' booths, with a machine running in a memory controller maker's space and two engineering processors in Bull's server.

## What is inside Venice?

The design combines compute chiplets manufactured on TSMC's N2 process with an input/output block at 6 nanometers. The maximum configuration is 256 Zen 6c cores with 512 threads, and for those who prioritize per-core speed there is a variant of up to 96 Zen 6 cores that boosts to 5.3 GHz. L3 cache reaches 1,024 MB, double the previous generation.

In memory, the leap is the most solid technical argument. Venice supports 16 channels of DDR5-8000 or MRDIMM-12800, up to 8 TB per socket and 1.6 TB/s of bandwidth, but only with one module per channel. That 1DPC restriction is what forces 32 modules to squeeze a two-socket server, and it arrives at the peak of the DRAM memory shortage, with the price of DDR5 multiplied in recent months.

Connectivity rises to 96 PCIe 6.0 lanes plus 8 PCIe 4.0 and CXL 3.1 lanes, and there is a correction here regarding what was published in the summer: the 128 lanes attributed to the socket are actually from SP8, the low-end platform. Each lane is twice as fast as in PCIe 5.0, so the total is still higher, but the round figure that circulated was incorrect.

Specification

EPYC Venice 9006

EPYC 9005 Turin

Xeon 6 6980P

Maximum cores and threads

256 and 512

192 and 384

128 and 256

Maximum L3 cache

1,024 MB

512 MB

504 MB

Memory channels

16

12

12

Bandwidth

1.6 TB/s

614 GB/s

About 845 GB/s

PCIe

96 PCIe 6.0 lanes and CXL 3.1

128 PCIe 5.0 lanes and CXL 2.0

96 PCIe 5.0 lanes and CXL 2.0

Maximum power

600 W

500 W

500 W

## Performance on paper

AMD publishes 33% more cores, double the cache and 20% more instructions per cycle versus Turin. In SPECrate, the benchmark that measures overall multithreaded performance, the flagship model scores 5,470 points versus 3,260 for the EPYC 9965, roughly 1.7 times more. These are manufacturer figures and should be treated as such until independent measurements arrive, something that in servers usually takes weeks.

The price of the top of the range is $14,904 for the EPYC 9996, and the actual purchase price usually differs from the list price. The family includes six models between 64 and 256 cores, with configurable power from 400 to 600 watts and seven years of support, a longevity that in data centers weighs as much as raw performance.

There is an asymmetry that benchmarks do not capture: the chip that performs best is not always the one that gets bought. In servers the decision is usually made by memory cost and power consumption, not the score.

The business figure accompanying the launch explains the haste. AMD grew data center revenue 97% in the third quarter, to $8.5 billion, and expects about $17 billion in the fourth, with Venice and Helios racks as the pitch. GF Securities analysts project 400,000 Venice units in 2026 and 5.4 million in 2027, a ramp that depends on memory being less scarce than expected.

## The embedded version and the detail of the signature

The same day came the EPYC Embedded 9006, the variant for equipment with long life cycles such as 400 and 800 gigabit firewalls, storage controllers and memory and SSD test benches. There are six models from 64 to 256 cores, with the same SP7 socket, power from 400 to 600 watts, seven years of production and firmware updates.

There a detail appears that will go unnoticed and that is relevant for the next decade: these chips include a hardware root of trust with post-quantum firmware signing. That is, the migration to quantum-resistant cryptography starts with the boot of the machine, not with the applications, and the first partners are Samsung, Lanner, Portwell and IEI.

AMD's [concurrency in the Linux kernel](https://observatorioblockchain.com/en/technology/linux-7-4-rc2-fat32-162x-torvalds-ia/) and its driver work for the AI Engines of the Versal chips show the other side of the same strategy: sell silicon and make sure free software supports it from day one, something Intel also does but with less commercial urgency in this cycle.

## What decides the purchase in 2027

With Venice in the catalog and Diamond Rapids without a date, the purchase decision comes down to three variables that do not depend on benchmarks. The first is the price of memory: 32 modules per two-socket server with scarce DDR5 can cost more than the processors. The second is cooling, because 600 watts per chip pushes air cooling to the limit and forces liquid to be planned.

The third is Intel's own schedule. If Diamond Rapids arrives in the second quarter of 2027, as analysts estimate from what was seen at SC26, Venice's window of advantage will be two or three quarters, enough to capture 2026 renewals and nothing more. If it is delayed again, AMD will have a full year without real competition in the segment above 128 cores.

That is the scenario that explains the ramp of 5.4 million units that analysts project for 2027 and also the risk: a market that buys out of need for AI capacity does not wait for the alternative supplier to catch up, but it also does not forgive a socket that forces the whole server to be changed if the price of memory does not cooperate.
