The first Zen 5 nodes went into Amsterdam in January and the last of the initial wave landed in Frankfurt six weeks later. Every plan on the Ryzen line is now a Ryzen 9 9950X, sixteen cores and thirty-two threads per chip, boosting to 5.7 GHz, paired with DDR5-5600 and Gen4 NVMe in mirrored pairs.
FRA-02 opened on this generation and holds nothing else. It is the first site we have built without a single node of a previous generation in it, which makes the capacity planning simpler and the failure modes duller.
What we measured
Vendor slides are not evidence. These are our numbers, taken on production node builds against the Zen 4 parts they replaced, with the same image, the same kernel and the same storage layout.
| Workload | Change against the previous generation |
|---|---|
| Single-threaded integer, one hot core | About 15% faster |
| Full compile of a large C++ tree, all cores | 9 min 40 s down to 8 min 12 s |
| Single-connection database read benchmark | About 13% more transactions per second |
| Package build farm, sixteen parallel jobs | About 11% faster wall clock |
| Energy per completed job | About 7% lower |
The single-thread figure is the one that matters, because the workloads people buy this line for are almost never parallel. They are one hot thread waiting on one core, and the only thing that makes them faster is a faster core.
Per-core power under sustained load went up. Energy per job went down, which is the number that appears on our electricity bill and, indirectly, on your invoice.
Why the EPYC line did not move
The server line stays on the EPYC 9354P, which is Zen 4, thirty-two cores and sixty-four threads with twelve channels of registered ECC DDR5-4800. This surprises people who assume newer is a strategy.
Customers on the EPYC line are memory-bound, not clock-bound. They buy it because the working set stopped fitting in RAM on anything smaller, and the constraints that matter are memory channels, capacity per socket, ECC and NUMA behaviour. On our own memory-bound test set, the newer server silicon was worth about eight percent. Eight percent does not justify a fleet migration.
A migration is not free. Every node means a scheduled window, a live migration for the instances that support it, a cold start for the ones that do not, five days notice to every affected customer, and roughly two hours of engineer attention. Multiply that by the fleet and you have spent a quarter of a year of everyone’s time for eight percent on a workload where the customer will not feel it.
We will move the EPYC line when the numbers justify it, and the post explaining the decision will contain the same table format as this one. Nothing older than the current server generation is in production, so the choice is between two current parts rather than between new and stale.
What this means if you are already on Ryzen
Existing instances were migrated at no cost, at the price they were already paying, during announced windows. Nobody paid a generation surcharge, because we do not sell one. The old generation is simply gone from the fleet rather than being resold at a discount to somebody less fussy, which is what most of the industry does with it.
One thing to check: if your build pins to specific CPU feature flags, the flag set changed. A handful of customers running hand-tuned numerical code had to rebuild. Support flagged this in the migration notice, and the three people it affected had all noticed before the notice arrived, because that is the kind of person who hand-tunes numerical code.
The other thing that shipped this year
Dedicated forty-gigabit ports became available on the EPYC, storage and metal lines at every site with the upstream capacity behind them. That is described properly in the June post about what unmetered means, and the two changes are related: a faster core is useless if the port in front of it is the bottleneck.
Availability
Zen 5 is in every site that sells the Ryzen line, which is all thirty-four of them. Stock varies, Zurich is rationed for reasons that have nothing to do with silicon, and Johannesburg is not open yet. The rest is on the location index.