- Introduction and Hardware Configuration
- Overall performance
- Individual game benchmarks
- Ashes of the Singularity
- Civilization VI
- Cyberpunk 2077
- Factorio
- Final Fantasy XIV Endwalker
- Forza Horizon 5
- Gemcraft: Frostborn Wrath
- Horizon Zero Dawn
- Kerbal Space Program
- OSRS Runelite 117HD
- Rainbow 6: Siege
- Shadow of the Tomb Raider
- Starcraft 2
- Stellaris
- The Riftbreaker
- Tiny Tina's Wonderlands
- Total War: Warhammer II
- World of Warcraft
Guide v1.02.
o/ Aeryn here!
This is my first attempt at a more formal guide (covering the performance of the 5800x3d, 5900x, memory scaling for both and 6c vs 8c CCD scaling). It's been a surprising amount of work and so i'm releasing as-is and will continue to polish and add information over time.
I used an Arctic Liquid Freezer II 420 with the offset mounting position and a controlled ambient temperature of around 23c for the Precision Boost tests.
The motherboard is an x570 Aorus Master with the bios F34 (AGESA 1203B) for the 5900x / F36c (AGESA 1206B) for the 5800x3d. This second AGESA has a couple of problems including a 1900fclk hole, but is required for the x3d to work properly.
The RAM is a pair of dual-rank sticks, also known as 2x2R which is top tier for overclocking performance, best all-around and also the best performing option within stock specifications. It is G.Skill Samsung 8gbit B-die - F4-3200C14-16GTZN.
There are two profiles used:
A: JEDEC-3200 2x2R, the best memory performance supported within the CPU specification with as much polish as allowed to give it the best possible shot. Also somewhat comparable to a bad XMP/DOCP - it may have worse primaries, but the dual-rank-per-channel configuration and a few tweaked options under the hood give it an edge.
When this profile is used on the charts below, it will be called " - JEDEC".
B: My daily overclock which has every accessible timing and several other parameters tweaked for optimal performance while retaining rock solid stability and reliability. It is capable of passing tests such as TM5 w/ Anta777's Extreme1 profile for 12+ hours and shows pretty much the best safe and reliable performance which is available from the DRAM.
This configuration was proven stable at 3800 and then stepped down to 3733 without changes.
When this profile is used on the charts below, it will be called " - RAM" or specifically mention the overclock in order to indicate that overclocked RAM is used.
The graphics card is an RTX 3080 with some tuning and the driver version 511.65.
I almost entirely avoided GPU bottlenecking by focusing on settings which were very easy for the graphics card to run. This generally means flat low settings, but some settings such as crowd density in Cyberpunk 2077 are turned up as these impact the CPU much more than other parts of the system. When these settings are used, they are explicitly mentioned. Avoiding bottlenecks elsewhere gives a better picture of what the CPU is truly capable of in order to better understand the microarchitecture and scaling both now and in the future. I also plan to gather data on higher settings which can have more CPU demand, but require careful monitoring and planning for correct measurement.
Turning up other settings may only hurt performance so you can consider the raw performance values in the game charts to be a ceiling for achievable performance when the mentioned settings are enabled. Which settings hurt CPU performance and by how much varies a lot game-to-game but is tuneable by the user.
I took measurements both at stock specifications and with a RAM overclock seperately. All gaming CPU's benefit massively from RAM overclocking at the moment, but some benefit many times more greatly than others. The 5800x3d stands out of the pack as a CPU which excels even with bad memory; the rest of the Vermeer lineup is much more reliant on a custom memory overclock and Cezanne or Alderlake are even more so still. Using different memory configurations shifts the absolute and relative performance of these CPU's quite massively! You can see below how the relative performance of even the 5800x3d and 5900x changes when this memory overclock is applied.
I focused on testing games which may respond to CPU/cache/mem performance and didn't bother testing some games which are known not to do so; this is not meant to be a perfect reflection of what the 5800x3d or 5900x does on every game; it's moreso designed to figure out where and what the scaling is with a limited amount of benchmark runs available. Several of these titles had never been tested with an x3d before i ran them, so nobody knew how they would behave. To interpret the data most accurately it should be treated as a subset of CPU/cache/mem sensitive games or looked at on a game-by-game basis to examine performance for that particular game.
As you can see later, more than a few also prefer an 8-core CCD to a 6-core one. This doesn't just create leads in a 6c12t vs 8c16t scenario (which is indicative of general core scaling), but it also frequently shows up in 8t16t vs 6c12t + 6c12t with the dual-CCD models.
There are regressions in clock speed which affect the cores (including their L1+L2 caches) and the L3 cache as well as an additional L3 latency penalty of around +9% when measured by clock cycles. These undoubtedly hurt everything on the list but at the same time, pretty much everything on there is finding at least a fringe benefit from the tripled L3 cache capacity which offsets this and claws performance back.
This clock speed reduction seems to favor the 5900x a bit over the 5800x3d. I think that the reasons for that are twofold:
1: The 5900x is more memory limited, but by slowing the core while leaving the RAM performance intact we reduce that bottleneck.
2: The 5800x3d is getting higher L3 cache hitrates, but the L3 cache runs on the core clock. By slowing the core clock we're also slowing its huge pool of L3 cache and hurting latency/bandwidth there.
This is evident in the way that the x3d loses to the 5900x in Civ 6 at a static 3.4ghz, but it manages a narrow win at stock clocks even though the 5900x is gaining more clockspeed than the x3d is.
The 5800x3d is ahead by 37% here with stock RAM or by 20% with overclocked RAM.
There's a 32% performance lead here for the 5800x3d with stock memory, down to +16% with overclocked.
I started up an endurance map with the modifier to send additional creeps and then queued 352 waves all at once, measuring the time that it took for them to kill me. This depends on the game simulation speed and scales with all manner of CPU performance.
Margin for error is rather high here due to the chaotic and non-deterministic nature of the simulation, but scaling trends are strong enough to establish favor for the x3d and for RAM overclocking. Don't take any specific value too seriously.
A picture with a simple average and a 1% low can't convey how much of a dream this is to play with. The x3d's larger L3 cache fails to elevate averages much sometimes, but it removes or heavily mitigates engine microstutter at all times which turns it into a completely different game.
Here's a frametime chart from CapFrameX which shows how radically the lows have been improved:
Riftbreaker is a game which scales onto 12 and even 16 cores, but core and cache/memory performance is of paramount importance and there's a clear 5800x3d>5900x>5800x tier list.
