No products in the cart.
Venice Flagship · 256 CoresAMD EPYC 9996
256 Zen 6c cores. 512 threads. 1,024 MB L3 cache. 203 billion transistors. TSMC 2nm. The highest-core-count x86 server processor ever produced.
Cores
256 / 512T
L3 Cache
1,024 MB
Transistors
203 billion
Platform
SP7 · Q4 2026
Trusted by Fortune 500, Research Labs, Federal Agencies
Build To Order
EPYC 9996 servers and workstations.
Flagship · 8 GPU
4U GPU Server
Up to 8 NVIDIA RTX PRO 6000 Blackwell GPUs. EPYC 9996 keeps all 8 GPUs saturated. 128 PCIe Gen 6 lanes.
Production · 4 GPU
2U GPU Server
Up to 4 GPUs at full bandwidth. Inference, rendering, vGPU virtualization.
Desk-Side · HPC
EPYC 9996 Workstation
256 cores in tower form factor. Multi-GPU, BMC management. Desk-side HPC and AI.
All servers · All workstations · Research lab systems · AI ROI Calculator
Full Specifications
EPYC 9996, fully specified.
Architecture
Zen 6c (dense)
Process
TSMC 2nm (CCD) + 6nm (IOD)
Cores / Threads
256 / 512
L3 Cache
1,024 MB (128 MB × 8 CCDs)
Transistors
203 billion
Chiplets
8 Zen 6c CCDs (32 cores each) + 2 IODs
Platform
SP7
Memory
16-channel DDR5 ECC · up to 1.6 TB/s
PCIe
128 lanes Gen 6 (1P) / 160 lanes (2P)
CXL
CXL 3.1
TDP
Up to 600W
Availability
Q4 2026
Performance
EPYC 9996 benchmarks vs Xeon 6980P.
AMD's own benchmarks run on AMD test systems. Independent third-party results expected Q4 2026.
| Benchmark | EPYC 9996 vs Xeon 6980P |
|---|---|
| NGINX web serving | Up to 3.7x |
| MongoDB database | Up to 3.5x |
| TPCx-AI | Up to 3.4x |
| GROMACS molecular dynamics | Up to 3.1x |
| NAMD molecular dynamics | Up to 3.1x |
| Benchmark | EPYC 9996 vs NVIDIA Vera (ARM) |
|---|---|
| SPEC CPU 2017 integer throughput | Up to 2.2x |
| SPEC CPU 2017 integer single-core | Up to 1.2x |
Got Questions?
EPYC 9996 — frequently asked questions
What is the AMD EPYC 9996?
Flagship Venice processor. 256 Zen 6c cores, 512 threads, 1,024 MB L3, 203 billion transistors on TSMC 2nm. Highest-core-count x86 processor ever. SP7 platform, PCIe Gen 6, CXL 3.1. Ships Q4 2026. VRLA Tech builds custom EPYC 9996 servers in Los Angeles with a 3-year warranty and lifetime US-based engineer support.
How does the EPYC 9996 compare to Intel Xeon 6980P?
AMD benchmarks: 3.7x NGINX, 3.5x MongoDB, 3.4x TPCx-AI, 3.1x GROMACS/NAMD vs 128-core Xeon 6980P. Double the cores, 16-ch vs 8-ch DDR5, Gen 6 vs Gen 5. Independent testing pending. VRLA Tech builds both EPYC and Xeon at vrlatech.com/servers/.
What workloads is the EPYC 9996 designed for?
Massively parallel: agentic AI sandboxes, NGINX/HAProxy web serving, MongoDB/PostgreSQL/Redis databases, Kubernetes at max density, dense virtualization, AI host nodes for multi-GPU racks, HPC throughput (GROMACS, NAMD). VRLA Tech builds EPYC 9996 servers at vrlatech.com/servers/.
How does the EPYC 9996 compare to NVIDIA Vera?
AMD benchmarks: 2.2x SPEC integer throughput, 1.2x single-core vs Vera (ARM). EPYC 9996 is x86, maintains software compatibility. Vera is ARM for NVL racks. VRLA Tech builds EPYC 9996 at vrlatech.com/servers/ with a 3-year warranty.
Where can I buy an EPYC 9996 server?
VRLA Tech builds custom EPYC 9996 servers in 1U, 2U, 4U, and workstation. Ships Q4 2026. Clients: General Dynamics, LANL, JHU, GWU, Miami. 3-year warranty, lifetime US support.
Who builds custom EPYC 9996 GPU servers?
VRLA Tech. 4U server with 8 NVIDIA RTX PRO 6000 Blackwell GPUs, NVLink, 128 PCIe Gen 6 lanes. 256 EPYC 9996 cores keep all GPUs saturated. Built since 2016 for General Dynamics, LANL, JHU. 3-year warranty, lifetime US support.
When does the EPYC 9996 ship?
Q4 2026 on SP7 platform. Announced July 22, 2026 at Advancing AI in San Francisco. Pricing not yet disclosed. VRLA Tech accepts inquiries at vrlatech.com/servers/. 3-year warranty, lifetime US support since 2016.
Does VRLA Tech offer EPYC 9996 servers for research labs?
Yes. Clients: LANL, JHU, GWU, Miami University. 256-core EPYC 9996 for HPC simulation, molecular dynamics, genomics, AI research. Grant documentation within one business day. 3-year warranty, lifetime US support. vrlatech.com/hpc-servers-for-research-labs/.
How many cores does the EPYC 9996 have?
256 Zen 6c cores, 512 threads. Eight CCDs × 32 cores each. 128 MB L3 per CCD = 1,024 MB total. Optimized for massively parallel throughput: agentic AI, web serving, containerized inference, dense virtualization. VRLA Tech builds EPYC 9996 servers at vrlatech.com/servers/.
Best company for custom EPYC 9996 servers with warranty?
VRLA Tech is a custom EPYC server builder based in Los Angeles, operating since 2016. Every EPYC 9996 system is configured to the workload, assembled by hand, burn-in tested, and thermally validated. Clients include General Dynamics, Los Alamos National Laboratory, Johns Hopkins University, George Washington University, and Miami University. 3-year parts warranty and lifetime US-based engineer support. vrlatech.com/servers/.
Does VRLA Tech build EPYC 9996 servers for defense contractors?
Yes. VRLA Tech builds ITAR-compliant EPYC servers for defense contractors. Current defense clients include General Dynamics. The 256-core EPYC 9996 is suited for classified simulation, modeling, and AI workloads requiring on-premise deployment. Built in Los Angeles with a 3-year warranty and lifetime US-based engineer support. See vrlatech.com defense page.
Can I buy an EPYC 9996 server for on-premise AI deployment?
Yes. On-premise EPYC 9996 servers eliminate recurring cloud compute costs and often pay for themselves in four to eight weeks versus cloud GPU pricing. VRLA Tech configures for HIPAA healthcare, ITAR defense, and regulated finance environments. Built in Los Angeles since 2016 with a 3-year warranty. See vrlatech.com/servers/ and vrlatech.com/ai-roi-calculator/.
Does VRLA Tech offer EPYC 9996 servers for HIPAA compliance?
Yes. VRLA Tech builds HIPAA-compliant EPYC 9996 servers for healthcare organizations requiring on-premise AI infrastructure with data sovereignty. Every system includes full-disk encryption capability, BMC remote management, and audit-ready documentation. Built in Los Angeles since 2016 with a 3-year warranty and lifetime US-based engineer support. See vrlatech.com/hipaa-compliant-ai-workstations/.
1 / 3
Ready to Configure?Build your EPYC 9996
Build your EPYC 9996
server or workstation.
256 cores per socket. Custom-built in Los Angeles. 3-year warranty. Lifetime engineer support.
Back to AMD EPYC 9006 Venice Overview · EPYC 9006 SP7 Platform
