AMD EPYC 8005
AMD EPYC 8005 selection: cores alone do not define value
AMD’s EPYC 8005 “Sorano” family brings full Zen 5 cores to the single-socket SP6 server segment. ServeTheHome’s SKU analysis is not a blanket recommendation for one processor. Its useful point is that seven chips sharing the same socket can produce very different outcomes for virtual machines, web services, databases, edge nodes, and compute farms. Buyers moving from EPYC 8004 Siena should treat this as a workload-sizing exercise, not a simple generational refresh.
What every SKU includes
Every EPYC 8005 part has six DDR5-6400 memory channels, 96 PCIe Gen5 lanes, and Zen 5 CPU cores. AMD does not reserve the platform I/O features for premium models. That removes one common purchasing concern, but it does not remove the core trade-off: core count, power, cache per core, and memory bandwidth per core move in different directions. The socket is still SP6, yet a drop-in upgrade must be verified against the server vendor’s BIOS, board support list, cooling design, and support policy.
Five concrete findings
- The range spans 8 to 84 cores. The 84-core EPYC 8635P is rated at 225 W by default and targets aggregate throughput.
- All parts can reach 4.5 GHz for single-core boost, but all-core boost differs. The 24-core 8225P reaches 4.1 GHz, while the 8635P reaches 3.45 GHz.
- Total memory bandwidth is fixed at roughly 307 GB/s. That works out to about 38 GB/s per core on the 8-core 8025P and about 3.7 GB/s per core on the 8635P.
- L3 cache per core varies with CCD configuration. The 8435P and 8225P have about 5.3 MB per core; several lower-count models have 8 MB per core. The 64-core 8535P has about 4 MB per core.
- At list price, 16- and 24-core models are below $50 per core. The 64-core model is around $86 per core, while the 84-core model is about $69 per core.
How to interpret those numbers
The 8635P makes sense where many independent tasks can keep 84 cores busy. Its default-TDP efficiency is also strong at roughly 2.7 W per core. The cost is greater contention for the fixed memory bandwidth and cache. A service with a smaller number of latency-sensitive workers, or one that needs abundant memory bandwidth per worker, may benefit more from fewer cores. The 32-core 8325P is a credible balanced candidate because its mix of core count, clocks, and cache is broadly useful; it is not a substitute for testing.
Do not infer application speed solely from the model number. The 16-core 8125P is an exception to the simple “fewer cores means higher all-core clock” pattern: it trails the 24-core 8225P. Benchmark with representative data, concurrency, and power settings. Capture tail latency such as p95 and p99, not only average throughput.
Deployment checklist
- Profile CPU use, runnable threads, memory bandwidth, cache misses, IOPS, and peak-period latency.
- Confirm SP6 server support, BIOS and BMC versions, DDR5-6400 DIMM population, cooling, and PSU capacity.
- Calculate full system cost: CPU, memory, networking, NVMe, power, warranty, and per-core software licensing.
- Run representative workloads with the intended DIMM population and power limit. Retain a baseline from the current fleet.
- Check application affinity, core limits, and licensing rules that could erase the benefit of an 84-core part.
Bottom line
EPYC 8005 simplifies SP6 feature selection but not CPU sizing. Choose the 8635P when density and total throughput are proven requirements. Evaluate 16- and 24-core parts for list-price value, and include the 8325P for balanced deployments. Compatibility, measured workload behavior, and multi-year operating cost should decide the final configuration.
Operational notes
Capacity planning should include a burn-in period under production-like traffic. Observe thermals, memory error reporting, firmware logs, and the behavior of backup, monitoring, and maintenance jobs. Keep the existing platform available until the new configuration meets its service-level objectives. Document BIOS settings, DIMM placement, power capping, and benchmark versions so that later capacity changes remain comparable. This discipline is especially valuable when a result is affected by memory population rather than the processor itself.
Source
Original reporting: https://www.servethehome.com/amd-epyc-8005-sku-list-and-value-analysis-mainstream-soranos-big-stride/.



