Home » Semiconductors & Computing Hardware » AMD Ships 6th Gen EPYC 9006 Server Chips, Up to 256 Cores

AMD Ships 6th Gen EPYC 9006 Server Chips, Up to 256 Cores

Servers built on Zen 6 are now coming from server makers, but nearly every performance number so far sits in AMD’s own tables. Buyers should wait for independent test results…

Close-up of a computer motherboard, illustrating server hardware like the AMD EPYC 9006

AMD says a single server socket can now hold 256 cores. On November 10, the company announced general availability of its 6th Gen EPYC server processors, built on the new Zen 6 design, and said platforms are shipping from leading server makers. The new AMD EPYC 9006 family targets cloud, enterprise and AI servers.

Here is what is verified. AMD’s own release confirms the launch, the shipping status and the 256-core figure for the EPYC 9996. The FPS Review lays out the wider platform specs. What is not verified: independent benchmarks. Every performance claim below comes from AMD.

What AMD Says Is Shipping Today

The announcement covers the SP7 server socket, the larger of AMD’s two planned platforms. AMD says SP7 systems are shipping from server makers now. Cloud instances are a step behind. AMD says Microsoft Azure and Oracle Cloud Infrastructure instances “will be generally available soon,” and lists an Azure HDv2 preview as already running.

Meta is named as a lead customer. AMD says Meta is validating platforms ahead of deployment at scale. That is a testing statement, not a deployment announcement.

The smaller SP8 platform is not here yet. According to The FPS Review, relaying AMD’s presentation, the rest of the family arrives across 2027.

Inside the 256-Core Part

AMD lists the EPYC 9996 at 256 cores per socket. A core is one independent processing unit, and more of them let a server run more tasks at once. The FPS Review reports that the 256-core ceiling comes from compact “Zen 6c” cores, with up to 512 threads and clock speeds up to 4.1 GHz. Standard Zen 6 parts top out at 96 cores and 192 threads, per the same report. AMD also cites high-frequency parts reaching up to 5.3 GHz for servers that feed GPUs.

The memory and connection specs, as The FPS Review relays from AMD’s press deck, are large for a server chip:

  • 16 channels of DDR5 memory, with standard modules up to 8000 MT/s (megatransfers per second, a speed rating).
  • Up to 1.6 TB/s of memory bandwidth per socket.
  • PCIe Gen 6, the newest generation of the link that connects GPUs and fast storage.

Power is the cost of all that density. AMD lists the EPYC 9996 at 600 watts of “Default CPU Power,” a term it now uses in place of TDP, its older power rating.

How AMD Built Its Benchmark Claims

AMD’s headline claim is that a two-socket EPYC 9996 server delivers up to 2.2 times the performance of a two-socket Intel Xeon 6980P system, and 2.0 times the performance per dollar. The test is SPECrate2017_int_base, an industry benchmark that measures how much parallel work a server finishes. AMD says it used published scores and prices as of November 5, 2026.

Measure (AMD’s footnote table) 2P EPYC 9996 2P Xeon 6980P
SPECrate2017_int_base score 5470 2510
List price used (per 1,000 units) $14,904 $13,955
Score per list-price dollar 0.367 0.180

Source: AMD press release, November 10, 2026, footnote ii. These are AMD’s figures.

AMD also claims up to 3.7 times the performance of a two-socket Nvidia Vera system. AMD says the Vera scores are its own estimates. Its 1.8 times claim for AI host servers is a projection from measured PCIe Gen 5 results to Gen 6 bandwidth, not a measured Gen 6 result.

What Is Still Unproven

List prices are not what large buyers pay, so the per-dollar figure is a rough guide at best. AMD’s footnote on the Intel comparison also does not say that the 2.0 times figure holds across other workloads. In its first footnote, AMD says its “world’s best server CPU portfolio” claim does not imply leadership in every workload.

The FPS Review’s piece includes no benchmarks, and its figures trace back to AMD’s press deck. A Phoronix review of the EPYC 9996 appears in search results, but we could not open it, so we cite no test numbers from it. Power draw is another open question. At 600 watts per chip, cooling and electricity limits may decide how many of these a data center can actually rack.

AMD also gives no list of every 9006 model with prices in its release, and cloud instance dates are still “soon.”

Who Benefits and How

Cloud customers are the likeliest winners. More cores per socket means a provider can fit more virtual machines in the same rack, which can lower the cost of renting compute. That depends on providers passing savings along, and no prices have been announced.

Data centers with tight power limits face a harder choice. A 600-watt chip does more work per socket, but it demands strong cooling. For everyday device users, the effect is indirect: faster, cheaper web services, if the numbers hold. Most buyers should expect real availability through server makers and cloud instances in the coming months, not overnight.

Plenty of businesses will not buy these chips at all. They will rent them. Watch for the first Azure and Oracle instance prices, then for independent test results that show whether 2.2 times survives outside AMD’s table.

Sources

  • AMD Investor Relations, “AMD Showcases 6th Gen AMD EPYC Processors to Lead Next Era of AI and Enterprise Computing,” November 10, 2026.
  • The FPS Review, “AMD EPYC 9006 ‘Venice’ CPUs & Zen 6 Architecture Launch,” November 12, 2026.

Related Reading