AMD Zen 6: everything we know before the announcement on July 22, 2026

Mark Papermaster, CTO of AMD, does not do things by halves. At the RAISE Summit in Paris in early July, the man confirmed what everyone was expecting: the sixth generation of Zen architecture will be officially unveiled on July 22 and 23, 2026 , at the Advancing AI event at the Moscone Center in San Francisco.

His exact words: "We're now on our sixth generation. So at our Advancing AI event on July 22nd and 23rd, we're rolling out this new generation." Hard to be clearer. No "we'll see", no "if all goes well" — it's a done deal.

The event promises to be massive: Lisa Su's keynote on the 23rd, over 100 technical sessions, and an expo with Dell, Google, Lenovo, Nutanix among others. But the big piece, obviously, is Venice — the codename for the first EPYC processors based on Zen 6.

We've compiled everything we know: architecture, specs, AI platform, competition, and above all the burning question on gamers' lips — when is it coming to desktop?

Zen 6, it's not just a refresh. AMD is moving to the 2nm from TSMC — a first for a high-performance server processor. And the figures being circulated are staggering.

The pipeline switches to 8 instructions wide (against 6 on Zen 5). The instruction reorder buffer has been enlarged, the branch predictor redesigned, and AMD has added a new micro-op cache. In short, the entire front-end has been revised.

On the hidden side, standard CCDs incorporate 12 cores for 48 MB of L3 — 50% more cores than Zen 5, which peaked at 8 cores per chiplet with 32 MB. And for ultra-dense configurations, AMD also offers a Zen 6c variant with 32 cores per chiplet , same instruction set, comparable IPC, but slightly lower frequencies.

The gross result according to AMD: +70% performance and efficiency compared to Turin (Zen 5), and +50% on SPEC. That's something. Even dividing by two for marketing, it remains a hefty generational leap.

The flagship that will be announced on July 22nd is the EPYC Venice And the specs are impressive enough to impress even server segment regulars.

The maximum config goes up to 256 Zen 6c cores and 512 threads on a single socket. Yes, you read that right. 256. In standard version (Zen 6 "big cores"), we are talking about 96 cores max with higher frequencies.

The socket also changes: make way for the SP7 , about 12% larger than the current SP5. Normal, given what needs to be powered - TDP rises between 700 and 1400 W according to the SKUs. It's massive, but that's the price to pay for 256 cores.

On the memory side, AMD is switching to 16 DDR5 channels with MRDIMM support at 12.8 GT/s, resulting in a memory bandwidth of 1.6 To/s per socket . And for the I/O, it's PCIe Gen 6.0 up to 128 lanes in dual-socket. Doubling the CPU-GPU bandwidth will please AI workloads.

Venice does not come alone. AMD pairs it with its Helios platform , its first complete rack-scale system for AI — a frontal attack against Nvidia's NVL72.

The heart of Helios is the GPU Instinct MI455X (architecture CDNA 5). Some staggering figures: 432 GB of HBM4 per GPU, up to 40 PFLOPS in FP4 by chip, and 320 billion transistors spread over about 12 chiplets. All in liquid cooling.

On the rack scale (72 MI455X), it gives 2.9 exaFLOPS in FP4 and 31 To HBM4 To give you an idea, this is the power that an entire supercomputer needed to achieve five years ago. Now, it's just one rack.

Meta has already announced plans to deploy Helios in the second half of 2026. OpenAI is also in the loop, with a multi-generational contract for Instinct GPUs. In short, AMD is no longer playing the role of outsider in AI - they are at the table.

The timing is not insignificant. Nvidia has just released Vera , its 88-core ARM CPU tailored for agentic AI. Intel pushes its Xeon 6 (and prepares Nova Lake). And AMD arrives with 256 x86 cores.

Three different philosophies. Nvidia Vera is betting everything on single-thread performance — about 10-11% ahead of the current EPYC 9575F on agent-based workloads where latency is key. It's clever: for real-time inference with AI agents, it's the fast core that matters, not the number.

But AMD counters with brute force. According to their projections, Venice displays 3.3x the performance of Vera in a 100 kW rack When you run hundreds of thousands of queries in parallel, it's the throughput that speaks. And with 256 x86 cores compatible with the entire existing software ecosystem, migration is simpler.

Intel? Let's say that the Xeon 6980P is 55% behind Vera and 46% behind Venice on these benchmarks. Panterlake arrives on the desktop side, but on the server side, Intel plays the niche of economic inference rather than raw performance. A niche positioning, that's it.

The burning question. Because yes, Zen 6 on the server side is exciting, but you and I mainly want to know when it lands in our towers.

The code name for desktop is Olympic Ridge — which will become the range Ryzen 10000 . And according to the most credible rumors, the announcement is scheduled for CES 2027 in January , with a launch following in the first half of 2027.

Good news for AM5 motherboard owners: no socket change AMD has confirmed that the AM5 will be supported until at least 2029. So your current B650 or X670 should welcome a Ryzen 10000 with a simple BIOS update. No AM6 for this generation.

The desktop CCDs would use TSMC's 2 nm, and the IOD would move to 3 nm. The IPC gain should be significant — if the server announces +70%, the desktop should recover a good part, perhaps 30-40% in gaming thanks to the improvements in interconnection and cache.

Initially, there were rumors of a late 2026 release, but that slipped. Some reports mention potential issues with silicon's B0 stepping. Nothing dramatic at first glance — just AMD preferring not to rush the desktop while the server brings in more margin.

The Advancing AI 2026 should shed light on quite a few areas of shadow:

• The precise SKUs — how many Venice models, at what TDP, what prices
• The official benchmarks — AMD will surely showcase Venice against Vera and Xeon on AI and HPC workloads
• The desktop roadmap — Lisa Su could drop a teaser about the Ryzen 10000 during the keynote on the 23rd
• Helios in action — live demos of rack-scale with MI455X and Venice together
• The partners — who buys first (Meta, OpenAI are already mentioned)

In summary: AMD is coming out strong. Zen 6 is not a quiet evolution—it's a node change (2 nm), an architecture redesign (8-wide pipeline), and a new strategic positioning (rack-scale AI). On July 22, we will find out if the promises hold up against independent benchmarks.

For gamers, patience. CES 2027 is in six months. By then, your Ryzen 7 9800X3D is still doing a great job.

The EPYC Venice server processors (Zen 6) will be announced on July 22-23, 2026 at the Advancing AI event in San Francisco. The first systems should be available in Q3 2026. For the desktop (Ryzen 10000), rumors point to CES 2027 (January) with a release in S1 2027.

Yes. AMD has confirmed support for AM5 until at least 2029. The Ryzen 10000 (Olympic Ridge) will remain on AM5, so your B650/X670/X870 motherboard will be compatible with a BIOS update. No AM6 for this generation.

AMD announces +70% performance and efficiency on server compared to Zen 5 (Turin). In desktop gaming, the IPC gain should be significant (estimated between 25-40%) thanks to the widened pipeline and cache improvements, but the precise figures will only be known after independent benchmarks.

It depends on the workload. Nvidia Vera (ARM, 88 cores) excels in single-thread performance and latency for agent-based AI. AMD Venice (x86, 256 cores) dominates in rack-scale throughput with 3.3x the performance of Vera per rack according to AMD. For gaming desktop, only AMD offers Ryzen — Nvidia does not have consumer CPUs.

If you're building a PC now, the Ryzen 7 9800X3D and Ryzen 9 9950X3D remain excellent choices on AM5. Since the platform will remain compatible with Zen 6, you will be able to upgrade the CPU later without changing the motherboard. No reason to sit idle for six months.