How to check FAH stability after undervolting?
Posted: Tue Dec 02, 2025 12:01 pm
- I was aiming for 725 mV to maximize efficiency.
- I was able to get 2100 MHz target limit (around a similar stable limit from a famous overclocker guide affiliated with MSI)
- Checked OCCT with steady extreme for 1 hour with no errors (add even more clock and a single error exists, more clock, more errors, too high and it will crash/reset immediately, this was basically the highest stable), reality clock around 2092 MHz and 725 or occasionally 730 mV, temp is quite good, 50-ish GPU and 60-ish VRAM.
- Tried game GOWR half the entire story on max (a few dozen hours), no issue.
- Tried FAH, 3 different projects (18263, 18245, and 15223), all 10 WU completed, no error or issue.
- But there's another 4 WU from a single project (15225) that keeps getting automatically dumped around 60-ish % completion (a few percent difference), now I don't see it again in the Projects tab. Is it bad WU? I did not check on the stock GPU config because I left it overnight.
- I foolishly deleted that username/account and started a new one with an even more conservative target of 1972 MHz (reality 1965 MHz) after much consideration, all the same three project successes (total 6, one almost done), I don't get any 15225 project, nor see it in the Projects tab.
Everything indicates that it wasn't undervolting instability that caused the dumped WUs.
Or maybe it's 50:50 that the project WU was the issue for some reason? If not,
Is FAHBench version 2.3.1 (latest from website) with 2017 file creation date any useful for the current core/engine?
- There's only OpenCL (0x21 core), no CUDA, which was used in the latest core (0x27)
- Three WU options (dihydrofolate reductase or dhfr, dihydrofolate reductase implicit or dhfr-implicit, and voltage-gated sodium channel or nav), the first two "common system for benchmarking molecular dynamics" and the latter "a large, membrane-bound ion channel system").
- Two precision, single and double.
- Accuracy check options, NAN check increment 10 checks, and runtime with 1-minute increment options.
If the WU wasn't the issue and this benchmark is still relevant, any suggestion on which option I should use?
Check every WU with each single and double-precision (GPU supports two different modes), but it seems too excessive and time consuming, accuracy check enabled, NaN check 10 is enough?, and runtime 15 minutes or I should set it to 1 hour?
Thanks!
CPU = Ryzen 7 8845HS (FAH only set to RTX 5060 Ti active so mostly idle), LPDDR5-7500, ReBar disabled (limited BIOS option though above 4G decode enabled by default)
OS = Windows 11 Pro 24H2, VBS disabled.
GPU = RTX 5060 Ti 16GB, driver 581.80 WHQL
eGPU dock = ADT-Link F43SG (running at PCIe 4.0 x4), never had fatal PCIe errors reported by HWiNFO or any crashes during usage.
- I was able to get 2100 MHz target limit (around a similar stable limit from a famous overclocker guide affiliated with MSI)
- Checked OCCT with steady extreme for 1 hour with no errors (add even more clock and a single error exists, more clock, more errors, too high and it will crash/reset immediately, this was basically the highest stable), reality clock around 2092 MHz and 725 or occasionally 730 mV, temp is quite good, 50-ish GPU and 60-ish VRAM.
- Tried game GOWR half the entire story on max (a few dozen hours), no issue.
- Tried FAH, 3 different projects (18263, 18245, and 15223), all 10 WU completed, no error or issue.
- But there's another 4 WU from a single project (15225) that keeps getting automatically dumped around 60-ish % completion (a few percent difference), now I don't see it again in the Projects tab. Is it bad WU? I did not check on the stock GPU config because I left it overnight.
- I foolishly deleted that username/account and started a new one with an even more conservative target of 1972 MHz (reality 1965 MHz) after much consideration, all the same three project successes (total 6, one almost done), I don't get any 15225 project, nor see it in the Projects tab.
Everything indicates that it wasn't undervolting instability that caused the dumped WUs.
Or maybe it's 50:50 that the project WU was the issue for some reason? If not,
Is FAHBench version 2.3.1 (latest from website) with 2017 file creation date any useful for the current core/engine?
- There's only OpenCL (0x21 core), no CUDA, which was used in the latest core (0x27)
- Three WU options (dihydrofolate reductase or dhfr, dihydrofolate reductase implicit or dhfr-implicit, and voltage-gated sodium channel or nav), the first two "common system for benchmarking molecular dynamics" and the latter "a large, membrane-bound ion channel system").
- Two precision, single and double.
- Accuracy check options, NAN check increment 10 checks, and runtime with 1-minute increment options.
If the WU wasn't the issue and this benchmark is still relevant, any suggestion on which option I should use?
Check every WU with each single and double-precision (GPU supports two different modes), but it seems too excessive and time consuming, accuracy check enabled, NaN check 10 is enough?, and runtime 15 minutes or I should set it to 1 hour?
Thanks!
CPU = Ryzen 7 8845HS (FAH only set to RTX 5060 Ti active so mostly idle), LPDDR5-7500, ReBar disabled (limited BIOS option though above 4G decode enabled by default)
OS = Windows 11 Pro 24H2, VBS disabled.
GPU = RTX 5060 Ti 16GB, driver 581.80 WHQL
eGPU dock = ADT-Link F43SG (running at PCIe 4.0 x4), never had fatal PCIe errors reported by HWiNFO or any crashes during usage.