AMD EPYC 9996 vs Intel Xeon 6980P: Full Spec Comparison
AMD says its new EPYC 9996 beats Intel's Xeon 6980P by 2.2x on AMD's own benchmarks. Here's how the two flagship server CPUs compare, spec for spec.
AMD's new EPYC 9996 is the fastest x86 server chip on paper, and on AMD's own benchmarks it beats Intel's current flagship, the Xeon 6980P, by roughly 2.2x on throughput and 2.0x on performance per dollar. But "fastest" and "best fit for your rack" are different questions: the EPYC 9996 draws up to 600W, costs $14,904 at AMD's quoted 1Ku price, and ships on a brand-new SP7 platform, while the Xeon 6980P is a known quantity that has been shipping to data centers since Q3 2024. Below is a spec-for-spec comparison pulled directly from AMD's and Intel's own product pages, plus what AMD's own numbers do and don't tell you.
The short version: EPYC 9996 has double the cores (256 vs 128), more cache, faster memory, and newer PCIe 6.0 I/O, and AMD's own SPECrate2017_int_base figures put it about 2.2x ahead of the Xeon 6980P in a 2-socket config. The Xeon 6980P is cheaper per chip ($13,955 vs $14,904), has been shipping for over two years, and draws less power per socket (500W vs 600W). Both are enterprise-only parts that nobody is buying for a home server.
Meet the contenders
The AMD EPYC 9996 is the flagship of AMD's sixth-generation EPYC "Venice" family, built on Zen 6 cores and TSMC's 2nm process. It's the top SKU in the new EPYC 9006 series, which AMD confirmed is in general availability as of its November 10, 2026 press release. The chip targets the same market Intel has chased for a decade: dense, AI-heavy data center racks where core count and memory bandwidth matter more than single-thread speed.
Intel's answer is the Xeon 6980P, the top part in the Xeon 6900P "Granite Rapids" P-core series and still Intel's highest-end shipping Xeon 6 processor as of this comparison. It launched in Q3 2024, which makes it a two-year-old design going up against AMD's newest silicon — a gap AMD is happy to point out, and one worth keeping in mind when you read AMD's performance claims below.
Spec-for-spec comparison
Here is how the two flagship parts compare using only figures published on AMD's own EPYC 9996 product page and Intel's own ARK spec page.
| Spec | AMD EPYC 9996 | Intel Xeon 6980P |
|---|---|---|
| Architecture | Zen 6 ("Venice"), TSMC 2nm | Granite Rapids P-core, Intel 3 |
| Cores / Threads | 256 / 512 | 128 / 256 |
| Base clock | 2.55 GHz | 2.0 GHz |
| Max boost / turbo | Up to 4.1 GHz | 3.9 GHz (3.2 GHz all-core) |
| L3 cache | 1,024 MB | 504 MB |
| Default CPU power / TDP | 600 W | 500 W |
| Memory channels | 16 | 12 |
| Max memory speed | DDR5-8000 (RDIMM); MRDIMM up to 12,800 MT/s | DDR5-6400; MRDIMM up to 8,800 MT/s |
| I/O | PCIe 6.0 x96 | PCIe 5.0, 96 lanes |
| Platform / socket | SP7 | LGA 4710 (Birch Stream) |
| Launch | General availability announced Nov. 10, 2026 | Q3 2024 |
| Price (1Ku / RCP) | $14,904 | $13,955 |
A decade-long rivalry, accelerated by AI
AMD's EPYC line and Intel's Xeon line have been trading the server-CPU crown back and forth since AMD re-entered the data-center market in 2017. For most of that stretch, the fight was about core count and price: AMD undercut Intel on dollars-per-core while Intel leaned on its software ecosystem, platform maturity and OEM relationships. The current round looks a little different. Both the EPYC 9996 and the Xeon 6980P are being marketed first and foremost as AI-infrastructure chips — the CPU that sits next to racks of GPUs, feeding them data and running everything the accelerators don't, from orchestration to agentic-AI workloads. That shift explains why memory bandwidth, PCIe generation and performance-per-watt show up as prominently in both companies' marketing as raw core count does.
It also explains the generational mismatch in this comparison. AMD is comparing its brand-new, just-launched flagship against a Xeon that has been on the market since Q3 2024. That's not unusual — new chips always launch against whatever the incumbent currently ships — but it does mean the "2.2x" framing below is really a comparison of 2026 silicon against 2024 silicon, not two same-generation rivals.
Cores, clocks and architecture
The headline number is cores: EPYC 9996 packs 256 of them against the Xeon 6980P's 128, exactly double. Both chips run simultaneous multithreading, so thread counts double too — 512 versus 256. AMD also wins on clock speed in both directions: a 2.55 GHz base versus Intel's 2.0 GHz, and a 4.1 GHz boost ceiling versus Intel's 3.9 GHz peak (3.2 GHz when all cores are loaded).
Cache tells a similar story. EPYC 9996 carries a full gigabyte (1,024 MB) of L3 cache, roughly double the Xeon 6980P's 504 MB. For workloads that are sensitive to cache misses — large in-memory databases, certain HPC codes — that gap can matter more in practice than the raw core-count difference.
What the AMD figure doesn't tell you on its own is core density per socket versus actual per-core performance, since Zen 6c (the dense/efficiency-optimized variant AMD uses to hit 256 cores) trades some per-core headroom for density. Neither company publishes a same-generation, same-core-count comparison, so the fairest reading is: AMD wins on aggregate throughput and raw core count; Intel's per-core clocks are in a similar ballpark once you account for the generation gap.
Power and thermal design
AMD lists the EPYC 9996 at 600W "Default CPU Power," the metric AMD now uses in place of a single TDP figure. Intel's Xeon 6980P is rated at 500W TDP on its ARK spec page. That's a 100W-per-socket gap, which compounds in 2-socket (2P) servers, the configuration both companies use for their flagship benchmarking. A 2P EPYC 9996 system can draw up to 1,200W from the two CPUs alone, versus up to 1,000W for a 2P Xeon 6980P system, before you count memory, networking, drives and cooling overhead.
That power difference is also why AMD's own performance-per-watt claims matter as much as raw throughput for operators who are power- or cooling-constrained — a growing concern as AI racks push past 35kW. See AMD's benchmark claims below for the specific figures.
Memory and I/O bandwidth
EPYC 9996 supports 16 memory channels per socket to the Xeon 6980P's 12, and AMD's top memory speed (MRDIMM at up to 12,800 MT/s) outpaces Intel's (MRDIMM at up to 8,800 MT/s) as well. More channels plus faster per-channel speed means meaningfully more memory bandwidth available to feed 256 cores — which is the point, since a core-dense chip is only as fast as the memory subsystem that can keep it fed.
On the PCIe side, AMD's product page lists PCIe 6.0 with up to 96 lanes for the EPYC 9996, while Intel's Xeon 6980P tops out at PCIe 5.0 with the same 96-lane count. PCIe 6.0 roughly doubles per-lane bandwidth over PCIe 5.0, which matters most for GPU-attached and high-speed-networking deployments — the same AI-server use case AMD is explicitly targeting with this launch. For a look at how server CPUs and accelerators are starting to talk to each other over newer interconnects, see our explainer on UALink, the open standard taking on Nvidia's NVLink.
Pricing and performance per dollar
At AMD's quoted 1Ku price, the EPYC 9996 costs $14,904 per chip when bought in batches of 1,000 — the highest price AMD has ever put on a single processor. Intel lists the Xeon 6980P's Recommended Customer Price at $13,955 on its ARK spec page, making the two flagships within about $950 of each other on paper. Neither figure reflects what any specific customer or OEM actually pays; both vendors offer volume and system-integrator pricing that isn't public, and retail/channel prices for both chips have moved well below list at various points since launch.
AMD's own pitch is that the $949 price premium of the 9996 is more than offset by the extra cores, cache and bandwidth: AMD claims roughly double the performance per dollar of the Xeon 6980P configuration in its own benchmark comparison, detailed next.
AMD's benchmark claims, and the asterisk that comes with them
In its November 10, 2026 press release announcing general availability of the EPYC 9006 series, AMD published a direct comparison between a 2-socket EPYC 9996 system and a 2-socket Xeon 6980P system, using the SPECrate2017_int_base benchmark, with data dated as of November 5, 2026:
- EPYC 9996 (2P): SPECrate2017_int_base score of 5,470
- Xeon 6980P (2P): SPECrate2017_int_base score of 2,510
- AMD's stated result: up to 2.2x the performance and 2.0x the performance-per-dollar of the Xeon 6980P system
- Performance per CPU watt: AMD cites 9.117 vs. 5.020 score per watt in the Xeon 6980P's favor for AMD
These are AMD-published, AMD-run figures, not independent third-party benchmark results, and they compare AMD's newest architecture against a Xeon generation that's now more than two years old — Intel's own newer Xeon 6+ ("Clearwater Forest") parts weren't part of AMD's comparison. SPECrate2017_int_base is a real, widely used industry benchmark, so the methodology is standard, but the specific test configurations, compiler flags and system tuning are AMD's to choose. Treat the 2.2x figure as AMD's best-case framing of its own hardware rather than a neutral verdict until independent labs publish their own numbers on both chips under the same conditions.
Which one should you actually buy
For net-new AI-inference and agentic-workload deployments where rack power and core density are the binding constraints, EPYC 9996's extra cores, cache, memory channels and PCIe 6.0 I/O make it the more future-proof pick — if your power and cooling budget can absorb the extra 100W per socket and your OEM has SP7 systems in stock. For organizations already standardized on Intel's platform, with existing Xeon 6900P-series infrastructure, firmware and support contracts, the Xeon 6980P remains a proven, shipping part with two years of production hardening behind it and a lower per-chip price. Buyers evaluating both should also factor in that AMD's $14,904 and Intel's $13,955 are list/1Ku prices, not what a given OEM server will actually cost once memory, storage, networking and support are included.
Neither chip is aimed at consumers or small businesses — these are 1P/2P data-center parts sold through OEM server vendors and cloud providers, the same category covered in our recent piece on Nvidia's Q3 earnings preview, where CPU-GPU pairing in AI racks is a recurring theme analysts are watching.
Software and platform maturity matter here too, in ways a spec sheet can't capture. The Xeon 6980P has had over two years in production, meaning BIOS, firmware, hypervisor and OS support across the major server vendors is thoroughly shaken out. The EPYC 9006 series, by AMD's own account, is only just reaching general availability on SP7 as of this comparison, with SP8-based systems and major cloud-instance types (Azure, OCI) still described as coming "soon" rather than available today. Buyers who need to deploy at scale this quarter, rather than next, should weigh that maturity gap alongside the raw spec advantage.
What's next
AMD says EPYC 9006 "Venice" SP7 systems are shipping now from OEM server vendors, with Microsoft Azure and Oracle Cloud Infrastructure instances based on the platform expected "soon," per AMD's press release. AMD's SP8 variants of the Venice family — covering sockets below the 9996 — are expected to roll out through the first half of 2027. On Intel's side, the Xeon 6980P remains the current flagship of the shipping Xeon 6 P-core lineup; Intel's next-generation Xeon 6+ "Clearwater Forest" parts, built on Intel's 18A process, are a separate line aimed at the E-core/efficiency side of the family rather than a direct, like-for-like replacement for the 6980P. Expect independent benchmarking of the EPYC 9996 against Intel's current and next-generation Xeons to start appearing from third-party labs in the coming weeks — worth checking before making a purchasing decision based on vendor-supplied numbers alone.
Frequently asked questions
What is the AMD EPYC 9996?
The EPYC 9996 is the flagship processor in AMD's sixth-generation EPYC "Venice" server CPU family, built on Zen 6 cores and TSMC's 2nm process. It has 256 cores, 512 threads, a 2.55 GHz base clock, up to 4.1 GHz boost, 1,024 MB of L3 cache, and a 600W default CPU power rating, per AMD's product page.
What is the Intel Xeon 6980P?
The Xeon 6980P is the top SKU in Intel's Xeon 6900P "Granite Rapids" P-core series and Intel's current flagship shipping Xeon 6 processor. It has 128 cores, 256 threads, a 2.0 GHz base clock, up to 3.9 GHz turbo, 504 MB of cache, and a 500W TDP, per Intel's ARK specifications page. It launched in Q3 2024.
How much faster does AMD say the EPYC 9996 is than the Xeon 6980P?
In its November 10, 2026 press release, AMD cited its own SPECrate2017_int_base benchmark results of 5,470 for a 2-socket EPYC 9996 system versus 2,510 for a 2-socket Xeon 6980P system, a claimed 2.2x performance advantage and 2.0x performance-per-dollar advantage. These are AMD-published figures, not independent third-party results.
How much do the AMD EPYC 9996 and Intel Xeon 6980P cost?
AMD lists the EPYC 9996 at $14,904 at its 1Ku (1,000-unit) price, the highest price AMD has set for a single processor. Intel's ARK page lists the Xeon 6980P's Recommended Customer Price at $13,955. Neither figure reflects actual OEM or retail pricing, which varies.
Which memory and PCIe standards does each chip support?
The EPYC 9996 supports 16 DDR5 memory channels (up to DDR5-8000, or MRDIMM up to 12,800 MT/s) and PCIe 6.0 with up to 96 lanes. The Xeon 6980P supports 12 memory channels (DDR5-6400, or MRDIMM up to 8,800 MT/s) and PCIe 5.0 with up to 96 lanes, per each company's own spec pages.
When can I buy a server built on either chip?
AMD says EPYC 9006 "Venice" SP7 systems are shipping now from OEM server vendors as of its November 10, 2026 announcement, with Azure and OCI cloud instances coming soon. The Intel Xeon 6980P has been shipping since Q3 2024 and is broadly available through Intel's server OEM partners.
Sources
- AMD: EPYC 9996 product page (specifications)amd.com
- AMD press release: AMD Showcases 6th Gen AMD EPYC Processors (Nov. 10, 2026)ir.amd.com
- Intel ARK: Xeon 6980P Processor specificationsintel.com
- Wikipedia: AMD Epycen.wikipedia.org
- Wikipedia: Intel Xeonen.wikipedia.org
Mara Lindqvist edits the hardware desk. She covers graphics cards, processors, memory and storage, the foundries and chip designers behind them, and what the numbers on a spec sheet mean for people choosing a PC. Specifications in her stories come from manufacturer spec pages and datasheets.


