Memory timings

Why is this important?

Until this version, users with any NVIDIA GPU with GDDR5X had to risk using EthEnlargementPill to reach better speeds and efficiency. EthEnlargementPill is a protected closed source binary. For the operation it does not need internet connection. Therefore it is especially suspicious because it does connect to some servers when you would expect it not to use any network resources. The issue was raised some time ago on Bitcointalk Forum. We are not saying that EthEnlargementPill is malicious, but considering being closed source, heavily protected and with only up to 100 lines of code that performs actual task while everything else is there for some unknown reasons, there are reasons to believe that not everything has been exposed what exactly this software does. Therefore we suggest you should NEVER run EthEnlargementPill on a computer that contains any sensitive information such as any logins or passwords or on a computer that you intend to use for any work that requires any kind of privacy. We were aware of the issue and worked hard to integrate this performance feature directly into NiceHash QuickMiner. Future versions will provide you with an automatic memory timings tuning.

From version 0.4.2.0 - ability to modify certain memory timings

You can already start using this feature manually by editing commands.json file. Example of commands.json file which modifies timings for GPU0:

[{
		"time": 0,
		"commands": [{
				"id": 1,
				"method": "device.set.memory.timings",
				"params": ["0", "__MT0=16", "__MT1=4"]
			}]
	}, {
		"time": 20,
		"commands": [{
				"id": 1,
				"method": "workers.reset.all",
				"params": []
			}]
	}, {
		"time": 30,
		"loop": 30,
		"commands": [{
				"id": 1,
				"method": "worker.print.efficiencies",
				"params": []
			}]
	}, {
		"time": 1,
		"loop": 4,
		"commands": [{
				"id": 1,
				"method": "devices.smartfan.exec",
				"params": []
			}]
	}, {
		"event": "on_quit",
		"commands": []
	}, {
		"event": "on_quickminer.start",
		"commands": []
	}, {
		"event": "on_quickminer.stop",
		"commands": []
	}]

What we have discovered so far.

Note: we do not have access to any private NVAPI where everything is explained. We are doing reverse engineering and these are our findings.

Code Name	Real Name	Known values	Remarks
__MT0	FAW	24, 20, 16	GDDR5X has this value 24 or 20 (1080 Ti or 1080). Set to 16 for better performance. Multiple of RRD.
__MT1	RRD	6, 5, 4	GDDR5X has this value 6 or 5 (1080 Ti or 1080). Set to 4 for better performance.

Excavator v1.6.8c build 764+ / upcoming QuickMiner (not available yet)

Download from here: https://github.com/nicehash/excavator/releases/download/v1.6.8b/excavator_v1.6.8b_build761_Win64.zip

Timing	Remarks
RC	considerable boost possible
RFC	considerable boost possible
RAS	considerable boost possible
RP	considerable boost possible
CL
WL
RD_RCD
WR_RCD
RPRE
WPRE
CDLR
WR
W2R_BUS
R2W_BUS
PDEX
PDEN2PDEX
FAW	EthEnlargementPill sets this to 16 (--revA sets to 20)
AOND
CCDL
CCDS
REFRESH_LO
REFRESH	considerable boost possible
RRD	EthEnlargementPill sets this to 4 (--revA sets to 5)
DELAY0
ADR_MIN
WRCRC
OFFSET0
DELAY0_MSB
OFFSET1
OFFSET2
DELAY0

Unfortunately, changing memory timings works only on Pascal series. If anyone has any tips that would leave us to make this work on Turing and Ampere... there is a 0.2 BTC bounty for this piece of information!

Following cards are fully supported (tested):

• TITAN V
• GTX 1080 Ti
• GTX 1080

From this version on, simply use timing name equal value to set timing. You can set multiple of them; example to set FAW to 16 and RRD to 4:

[{
		"time": 0,
		"commands": [{
				"id": 1,
				"method": "device.set.memory.timings",
				"params": ["0", "FAW=16", "RRD=4"]
			}]
	}, {
		"time": 20,
		"commands": [{
				"id": 1,
				"method": "workers.reset.all",
				"params": []
			}]
	}, {
		"time": 30,
		"loop": 30,
		"commands": [{
				"id": 1,
				"method": "worker.print.efficiencies",
				"params": []
			}]
	}, {
		"time": 1,
		"loop": 4,
		"commands": [{
				"id": 1,
				"method": "devices.smartfan.exec",
				"params": []
			}]
	}, {
		"event": "on_quit",
		"commands": []
	}, {
		"event": "on_quickminer.start",
		"commands": []
	}, {
		"event": "on_quickminer.stop",
		"commands": []
	}]

Calling device.get or devices.get also returns actual timings for each GPU device, example:

...
"gpu_memory_timings":
{
    "bEditable": false,
    "timings": {
        "RC": 78,
        "RFC": 210,
        "RAS": 52,
        "RP": 26,
        "CL": 24,
        "WL": 5,
        "RD_RCD": 26,
        "WR_RCD": 16,
        "RPRE": 0,
        "WPRE": 1,
        "CDLR": 9,
        "WR": 27,
        "W2R_BUS": 7,
        "R2W_BUS": 7,
        "PDEX": 12,
        "PDEN2PDEX": 2,
        "FAW": 16,
        "AOND": 0,
        "CCDL": 2,
        "CCDS": 2,
        "REFRESH_LO": 5,
        "REFRESH": 4,
        "RRD": 4,
        "DELAY0": 12,
        "ADR_MIN": 6,
        "WRCRC": 16,
        "OFFSET0": 39,
        "DELAY0_MSB": 0,
        "OFFSET1": 13,
        "OFFSET2": 7
    }
},
...