Ryzen vs Intel Input Lag/Latency

[urikawa]

thank you for your response where do you work ?
On my motherboard "Zen5 gaming mode" have 3 settings 'level 1, level 2, legacy" i think you are talking about level 2 for gaming mode right?
What agesa version are you on?

[peak]

ZEN5 Gaming Optimizations: Enabled
DFE Read Training: Enabled
Gear Down Mode: Disabled
UCLK DIV1 Mode: UCLK=MEMCLK
SoC / Uncore OC Mode: Enabled
SoC Voltage VDDCR_SOC: Auto
VDDCR_SOC Load-Line Calibration = Level 3
Infinity Fabric Frequency: 2000Mhz
DRAM voltages left on Auto
Above 4G Decoding: Enabled
Suspend to RAM: Disabled
XHCI Hand-off: Disabled
Global C-state Control: Enabled
Power Supply Idle Control: Typical Current Idle
Local APIC Mode: xAPIC
ACPI _CST C1 Declaration: Enabled
MCA FruText: False
SMU and PSP Debug Mode: Disabled
MONITOR and MWAIT disable: Disabled
SVM Lock: Disabled
SVM Enable: Disabled
ACPI SRAT L3 Cache as NUMA Domain: Disabled
PSP Error Injection Support: False
Power Down Enable: Disabled
Disable Memory Error Injection: True
TSME: Disabled
Memory Context Restore: Enabled
IOMMU: Disabled
PCIe ARI Support: Disabled
PCIe All Port ECRC: Disabled
Advanced Error Reporting (AER): Not Supported
PCIe ARI Enumeration: Disabled
dGPU Only Mode: Enabled
UMA Version: Non-legacy
GPU Host Translation Cache: Disabled
NB Azalia: Enabled
PCIe Loopback Mode: Disabled
Persistence mode for legacy endpoints: Disabled
Retimer margining support: Disabled
FCH Spread Spectrum: Disabled
CPPC Dynamic Preferred Cores: Cache
GFXOFF: Disabled
Pluton Security Processor: Disabled
DRTM Support: Disabled
SMM Isolation Support: Disabled
ABL Console Out Control: Disabled
App Compatibility Database: Disabled
Unused GPP Clocks Off: Enabled
Clock Power Management (CLKREQ#): Enabled
PM L1 SS: L1.1_L1.2
DDR5 Nitro Mode: Enabled
DDR5 Robust Training Mode: Auto
Nitro RX Data: Auto
Nitro TX Data: Auto
Nitro Control Line: Auto
Nitro RX Burst Length: Auto
Nitro TX Burst Length: Auto
PBO: Auto
PPT / TDC / EDC: Auto
Performance Preset (in OC Tweaker menu): Auto
Boost Override: Disabled
Curve Optimizer: Not forced
Manual CPU OC: Not used

SMT does stay on. Modern engines are multi-threaded and if I disabled SMT, Windows tasks compete with game threads which means more context switching and worse 1%/0.1% lows. X3D+SMT is explicitly optimized by AMD too. The old disable SMT/HT comes from an old Intel era with fewer cores, bad Windows scheduling, older engines and CPUs with limited cache. None of that applies anymore.

Some other BIOS settings with my comments:

AMD fTPM: Disabled (but can leave this enabled if you need it. AMD fully fixed the stutter issue people had when fTPM was enabled in later AGESA updates. Disabling this might introduce more background activity in Windows but haven't confirmed this yet. fTPM should have zero imapct on latency or latency as it doesn't run on game threads and does not poll during gameplay. None of the games I play require TPM though)

PSS Support: Enabled (I've read elsewhere to keep this disabled but I disagree. This enables fast boost transitions, it works together with CPPC & Preferred Cores and theres really no downside on modern systems. I could not find a performance penalty with this on)

NX Mode: Enabled (Has no impact on gaming performance or latency. It is enforced at the hardware page-table level and does not run during gameplay).

Re-Size BAR Support: Disabled (The GPU bandwidth of the 4090 is already far beyond what esports engines need. Frametime consistency, 1%/0.1% lows and input-to-photon latency were slightly hurt when I had this enabled. YMMV)

DF Cstates: Enabled (I've seen other guides that have this disabled but I disagree on a few points. DF Cstates control low power idle states for the Data Fabric and on AM5 DF Cstates are very light and the exit latency is extremely fast. They are designed to work with CPPC, boost and Windows scheduling. This doesn't work like old "deep sleep" behavior. Disabling this will hurt your latency consistency instead of helping it)

ECC: Disabled (My RAM is non-ECC DDR5 UDIMM, it has on-die ECC only which is not system ECC. Explicitly telling AGESA the "Disabled" setting says there is no system ECC and to use normal, lowest latency memory path)

And that's pretty much it. I've seen other guides telling people to max out tREFI but I think that's bad advice (only makes sense in short benchmark runs/LN2/XOC/etc). Auto already sets it high but safe and dynamically adjusts with temperature, plays well with Nitro Mode and it gives you 99% of the latency benefit already with 100% of the stability.

Windows Settings (still using 23H2)

Windows Game Mode: On (This prioritizes game threads and reduces background interference but more importantly works correctly with Ryzen CPPC + X3D)

Hardware-Accelerated GPU Scheduling (HAGS): On (This is controversial I know as there have been reports from different people that their esports games run worse, usually in the instance where there isn't a GPU bottleneck but supposedly Microsoft has made some fixes since. I haven't encountered any issues with stuttering that was reported. HAGS should in theory reduce CPU scheduling overhead, lower render queue latency which is especially beneficial with an RTX4090)

Xbox Game Bar: Off (prevents overlay hooks, avoids background capture logic, reduces DPC latency risk, etc). In addition to this all Background Capture options in Windows Settings has been turned off.

Notifications: Off

USB Power in Device Manager every USB Root Hub's Power Management tab unchecked “Allow the computer to turn off..."

Timer Resolution: W11+Game Mode+HAGS already manages this properly. No need to do registry hacks anymore or use ISLC.

Power plan: Balanced (and then in Power Mode set it also to "Balanced" in Windows Settings on W11), This preserves AMD CPPC logic and keeps cores ready and has the best 1%/0.1% lows in my testing. AMD supposedly tunes for Ryzen under Balanced. I believe AMD has a "Ryzen Balanced" power profile as well you can download with the chipset software but I never tried it. Setting a high performance or ultimate performance works against the X3Ds. Boost responsiveness is already maxed under Balanced. And a higher performance plan can increase boost oscillation, DPC latency and make frametimes less smooth.

The only things I did not touch are assigning things like affinities to cores as my benchmarks didn't demonstrate in actual meaningful difference (unlike prior Intel systems I had with say pining the last core to my GPU resulted in significant FPS gains).

If a game has NVIDIA Reflex I would keep NVCP Low Latency Mode OFF. If a game does NOT have Reflex I use NVCP Low Latency Mode ON (not Ultra). ULLM forces “just-in-time” frame submission which can cause frametime jitter, inconsistent pacing, input microstutter. At 600Hz consistency beats theoretical queue depth.

I could go into more depth in NVCP settings if anyone cares but most of the stuff I turned off.

Any questions or further elaboration on why I chose these particular settings please ask, willing to go through why I chose these.

I understand the majority of the selected BIOS options, but I'm curious how you ended up with the RAM related settings.

SoC / Uncore OC Mode: Enabled
SoC Voltage VDDCR_SOC: Auto
VDDCR_SOC Load-Line Calibration = Level 3
Infinity Fabric Frequency: 2000Mhz
DRAM voltages left on Auto
UCLK DIV1 Mode: UCLK=MEMCLK
Gear Down Mode: Disabled
DDR5 Nitro Mode: Enabled
DDR5 Robust Training Mode: Auto
Nitro RX Data: Auto
Nitro TX Data: Auto
Nitro Control Line: Auto
Nitro RX Burst Length: Auto
Nitro TX Burst Length: Auto

I have an ASRock X870E, I currently loosened my EXPO's AGESA primaries to be sure that I'm not unstable and out of curiosity I scanned the list to see how it differs from me and my default approach.
Wouldn't the EXPO Competitive timings be considered XOC, and do the provide any meaningful improvements at all? To my understanding RAM timings is a grey area.
Is it possible for you to auto your voltages, or does this just mean EXPO defaults?
My EXPO's SOC voltages are set to 1.2, so I was curious if you ended up setting it to auto by yourself?
I also listed Gear Down Mode because I wouldn't want to touch that if I already expected EXPO being a risk of instability.
UCLK Div1 Mode gets set to /2 for me when enabling the competitive mode, I was curious if that also happens for you and if you manually set it back yourself.
The Nitro settings default to Enabled and fixed values instead of Auto, once again - just curious if you manually Auto them by design.
I haven't seen anyone do this and I was curious why you did that, if it's defaults or if there were any meaningful gains behind it.

My personal goal is absolute stability and I'm currently roughly stuck to default on almost the same setup.

[yehaw]

CPU is too complex to say one is better than the other for latency. They both excel in latency, while one or the other might edge out the competitor in specific scenarios. For the average/competitive gamer, focus on getting a CPU that has features you want. Once you put them under load of a game, you'll never notice any latency difference with all the other latency noise stacking.

Upgrading your peripherals/hardware will have a larger perceptual difference on latency:
- 4k-8k mouse
- 1k gaming actuation sensitive keyboard setup properly, it's not as easy as just switching to 4k profile like a mouse
- High refresh rate monitor
- High fps

If you asked me my personal preference, I like the Intel monolithic architecture better