AMD Previews Preliminary EPYC Venice Performance…
0 reactions
AMD Previews Preliminary EPYC Venice Performance Projections vs Nvidia Vera and Intel GNR-AP
AMD released a blog: Agentic AI Needs Rack-Scale CPU Performance – AMD EPYC Delivers It Today https://www.amd.com/en/blogs/2026/agentic-ai-needs-rack-scale-cpu-performance-amd-epyc.html with preliminary performance figures on Venice.
*Keep in mind*, these are AMD's 1st party numbers and assumes a single rack on a 2P node configuration normalized to a 100kW rack, scaling above Nvidia Grace by a 1.63x factor (from Phoronix's Vera test):
|Metric|Nvidia Vera (88C)|Intel Xeon 6980P (128C)|AMD EPYC 9965 (192C)|AMD EPYC "Venice" (256C)|
|:-|:-|:-|:-|:-|
|Cores Per Node (Normalized)|1.0|1.45|2.18|2.90|
|Cores Per Rack (Normalized)|1.0|1.24|1.86|2.08|
|Nodes Per Rack (Normalized)|1.0|0.85|0.85|0.71|
|Single Node Performance Derivation|Estimated as \~1.63x of the Nvidia Grace|N/A|N/A|Estimated as \~1.7x of the AMD EPYC 9965 SPECrate®2017\_int\_base and internal testing|
|Rack-Level Performance Derivation iv|Number of Nodes x Node level performance of Nvidia Vera|Number of Nodes x Node level performance of Intel Xeon 6980P|Number of Nodes x Node level performance or AMD EPYC 9965|Number of Nodes x Node level performance of AMD EPYC "Venice" 256C|
**Performance Projections**
|Workload|Nvidia Vera (88C)|Intel Xeon 6980P "GNR-AP" (128C)|AMD EPYC 9965 "Turin" (192C)|AMD EPYC "Venice" (256C)|
|:-|:-|:-|:-|:-|
|SPECrate®2017\_int\_base|1 (est.)|1.47|1.60|2.40 (est.)|
|Server-side Java® multi-JVM max|1|2.34|2.93|3.76|
|Web Serving (NGINX)|1|1.18|2.37|3.30|
|Key-Value Store (Redis)|1|1.31|2.23|3.10|
|In-Memory Caching (Memcached)|1|0.93|2.49|3.47|
|Relational Databases (TPROC-C)|1|1.99|2.91|4.05|
|**Geometric Mean**|**1**|**1.46**|**2.37**|**3.30**|
**Estimated Per Core Perf:**
|Metric|Nvidia Vera (88C)|AMD EPYC "Venice" (64C)|AMD EPYC "Venice" (96C)|
|:-|:-|:-|:-|
|Cores Per CPU|88|64|96|
|Node Level SPECrate®2017\_int\_base\*|1.0|0.92|1.21|
|Per-core SPECrate®2017\_int\_base\*|1.0|1.27|1.11|
\*2p server estimates
Rest of the details are here: https://www.amd.com/content/dam/amd/en/documents/solutions/ai/methodology-description.pdf
#hardware #technology
AMD released a blog: Agentic AI Needs Rack-Scale CPU Performance – AMD EPYC Delivers It Today https://www.amd.com/en/blogs/2026/agentic-ai-needs-rack-scale-cpu-performance-amd-epyc.html with preliminary performance figures on Venice.
*Keep in mind*, these are AMD's 1st party numbers and assumes a single rack on a 2P node configuration normalized to a 100kW rack, scaling above Nvidia Grace by a 1.63x factor (from Phoronix's Vera test):
|Metric|Nvidia Vera (88C)|Intel Xeon 6980P (128C)|AMD EPYC 9965 (192C)|AMD EPYC "Venice" (256C)|
|:-|:-|:-|:-|:-|
|Cores Per Node (Normalized)|1.0|1.45|2.18|2.90|
|Cores Per Rack (Normalized)|1.0|1.24|1.86|2.08|
|Nodes Per Rack (Normalized)|1.0|0.85|0.85|0.71|
|Single Node Performance Derivation|Estimated as \~1.63x of the Nvidia Grace|N/A|N/A|Estimated as \~1.7x of the AMD EPYC 9965 SPECrate®2017\_int\_base and internal testing|
|Rack-Level Performance Derivation iv|Number of Nodes x Node level performance of Nvidia Vera|Number of Nodes x Node level performance of Intel Xeon 6980P|Number of Nodes x Node level performance or AMD EPYC 9965|Number of Nodes x Node level performance of AMD EPYC "Venice" 256C|
**Performance Projections**
|Workload|Nvidia Vera (88C)|Intel Xeon 6980P "GNR-AP" (128C)|AMD EPYC 9965 "Turin" (192C)|AMD EPYC "Venice" (256C)|
|:-|:-|:-|:-|:-|
|SPECrate®2017\_int\_base|1 (est.)|1.47|1.60|2.40 (est.)|
|Server-side Java® multi-JVM max|1|2.34|2.93|3.76|
|Web Serving (NGINX)|1|1.18|2.37|3.30|
|Key-Value Store (Redis)|1|1.31|2.23|3.10|
|In-Memory Caching (Memcached)|1|0.93|2.49|3.47|
|Relational Databases (TPROC-C)|1|1.99|2.91|4.05|
|**Geometric Mean**|**1**|**1.46**|**2.37**|**3.30**|
**Estimated Per Core Perf:**
|Metric|Nvidia Vera (88C)|AMD EPYC "Venice" (64C)|AMD EPYC "Venice" (96C)|
|:-|:-|:-|:-|
|Cores Per CPU|88|64|96|
|Node Level SPECrate®2017\_int\_base\*|1.0|0.92|1.21|
|Per-core SPECrate®2017\_int\_base\*|1.0|1.27|1.11|
\*2p server estimates
Rest of the details are here: https://www.amd.com/content/dam/amd/en/documents/solutions/ai/methodology-description.pdf
#hardware #technology
| 4,000 mlx | |
| 4 | |
| 0 |