Latest C7H BIOS

[Shamino]

https://www.dropbox.com/s/k2beo11w1kw7kus/ROG-CROSSHAIR-VII-HERO-WIFI-ASUS-0068.rar?dl=0

[gupsterg]

I think we'll still see duplication.

AMD CBS is the menu that was on 1xxx/2xx and now can be used on 3xxx.

The AMD Overclocking menu I saw on a slide by AMD and believe is just for 3xxx. Why I say that is I have used UEFI 2406 with 2700X and never saw that section, but on 3600 I do.



On 1xxx/2xxx CPU, it was sorta preferred to use the ASUS settings. If due to OC, board failed to POST, AMD "recovery mode" would reset settings. All the settings on ASUS sections of UEFI would still be there to reapply, the AMD CBS ones would be reset and need to be manually re-entered or profile would need reapplying.

[KeithMyers]

Hey Matt, thanks for the feedback. My first question is what kind of ram timings did you manage at 3600Mhz. I can do CL14 Fast timings. That certainly helps with the integer math. But you are correct in your assumption that Linux aids math calcs better than Windows. It does for almost all the comparisons I regularly come across. But that test system certainly wasn't optimized. It is my daily driver and has everything plus the kitchen sink running on it. My barebones systems that do nothing but crunch whoop it regularly when it comes to cpu task processing time simply because they don't have all the crap running on them that I do on this daily driver. It is what I browse on, handles all the network communications for all my solar generation and monitoring and weather telemetry gear to upload to the cloud and runs my BoincTasks server to constantly poll all the other PC's on the network for their running tasks and stats. All that stuff is running when I do a Geekbench 4 benchmark. The only process I stop for the benchmark is BOINC.

[MattTheTech]

KeithMyers wrote:
Hey Matt, thanks for the feedback. My first question is what kind of ram timings did you manage at 3600Mhz. I can do CL14 Fast timings. That certainly helps with the integer math. But you are correct in your assumption that Linux aids math calcs better than Windows. It does for almost all the comparisons I regularly come across. But that test system certainly wasn't optimized. It is my daily driver and has everything plus the kitchen sink running on it. My barebones systems that do nothing but crunch whoop it regularly when it comes to cpu task processing time simply because they don't have all the crap running on them that I do on this daily driver. It is what I browse on, handles all the network communications for all my solar generation and monitoring and weather telemetry gear to upload to the cloud and runs my BoincTasks server to constantly poll all the other PC's on the network for their running tasks and stats. All that stuff is running when I do a Geekbench 4 benchmark. The only process I stop for the benchmark is BOINC.

Crazy man! So My Timings were also Fast 3600, but I tuned them even further than the Fast preset in the new Ryzen DRAM Calculator, specifically For Timings that differ from the preset is trcdrd:14-trc:38-trrdl:4-tfaw:12-trfc:256-tcwl:12-trtp:7, mainly I have found the one that seems to make the most difference in latency (and really we are only talking about .1 of a ns, but its consistent, is dropping the trfc as low as you possibly can. However, I have all 4 DIMMS populated with 2 16gb Flare X 3200CL14 Kits, so its a tiny bit harder to tune these than just with one Kit, at least for me. And when I first Validated this, I was able to pass 1000% of Karhu's RAM Test no problem, and I am not sure what has changed (Probably the night I spent throwing 1.75v at the DIMMS trying to get 3800MHZ to boot I might have degraded one of the DIMM Modules a bit) but I kept getting weird program crashes last night some time after those benchmark runs, and all roads pointed to the RAM, so I did another Ram Test, and this time I couldn't get past 177% with the same exact timings I had validated before. I spent all night trying to get it stable, and even with 1.6v it just wasn't happening, so I dialed it back to 1.45v and went back to the Fast Preset, which would still fail at around 300% consistently, So I ended up having to raise the trfc to 294 before I was finally passing 2000% without a problem. So now I am 100 Percent stable again, but unfortunately my timings are a tiny bit looser than what you were running (if you stuck with Fast). I was able to beat you in both Memory Bandwidth and Latency (how ever slightly) before, and now I bet if I did the same exact test you would probably beat me slightly.

So I don't think the reason why you crushed me so bad was down to RAM, it has to be something in my Windows System that is just crushing me, or I am possibly not as stable as I thought or something. I am going to boot up to one of my Ubuntu Drives when I have a chance and run the same test again and see how I do. That will at least tell me if Its Windows or my Overclock that is the issue.

Regardless this was a fun and revealing exercise. I hope more people post their Geekbench 4 scores to give you something to compare too.

[MattTheTech]

KeithMyers wrote:
Hey Matt, here is the link to my Google Drive for the old MLC 3.6.7z that I used. You can run it for direct comparisons against my full scores.

https://drive.google.com/open?id=1eH6YKPXvn8mNe40Q1AUqdJO2w-jeu3jg


Hope this helps you. We are not that far apart with regard to memory performance.

Thank You for the MLC 3.6.7. I Ran it, and am about to post my results below.

Also I am curious if you have run The Blender Benchmark for your system. They have it on Linux (You can get it from this url: https://opendata.blender.org/ ), and its a great realistic workload to compare against Platforms. Linux is also supposed to have the upper hand here, but I am wondering by how much. Below I am posting 2 of my Results. The First one is the Benchmark I just ran, its with my system at 4225Mhz all core, With memory at 3600Mhz and the not as tight CL 14-14-15-14-30-42-294 timings. Then the result below that was with my tighter timings that I mentioned earlier (The Primaries being CL 14-14-14-14-28-38-256), and My Per CCX OC which was CCX0:4450Mhz/CCX1:4400Mhz/CCX2:4250Mhz/CCX3:4300. This gave me a pretty decent little boost, which this benchmark reflects. Anyways Here's all the data, I will list Blender First, since they are just links.

Blender 4250 All Core/ Ram 3600 (Slightly Looser Timings): https://opendata.blender.org/benchmark/7e1de4e8-2e5f-4e00-825d-87058058e955

Blender Per CCX OC / Ram 3600 (Tighter Timings): https://opendata.blender.org/benchmark/bfb17b14-465d-43eb-ab67-2e9771f94161

And then here is the MLC Test Results using the same version You Did:

M:\Desktop\Programs\Tools\Memory Tools\MLC v3.6.7\Windows>mlc
Intel(R) Memory Latency Checker - v3.6
Measuring idle latencies (in ns)...
Numa node
Numa node 0
0 51.0

Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 52202.7
3:1 Reads-Writes : 45943.8
2:1 Reads-Writes : 44857.6
1:1 Reads-Writes : 43866.5
Stream-triad like: 46852.8

Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 52219.9

Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 311.15 52462.6
00002 315.54 52413.7
00008 302.79 52484.3
00015 301.74 52479.9
00050 293.57 52511.5
00100 279.07 52483.3
00200 120.75 51526.8
00300 68.97 37011.6
00400 63.24 28865.1
00500 61.16 23653.2
00700 57.92 17542.7
01000 56.09 12789.7
01300 55.86 10145.4
01700 55.09 8076.6
02500 54.84 5881.2
03500 54.48 4553.1
05000 54.41 3543.7
09000 54.32 2495.9
20000 54.01 1779.6

Measuring cache-to-cache transfer latency (in ns)...
Using small pages for allocating buffers
Local Socket L2->L2 HIT latency 14.4
Local Socket L2->L2 HITM latency 24.6

[EZ_PC_TECH]

The BIOS with new AGESA 1003 ABBA that must improve boost for Ryzen 3000 available for download through the BIOS EZ-Flash utility for C7H. Not published on the website yet. Share your results!

https://i.redd.it/l2kzj084aln31.jpg

[KeithMyers]

Thanks for the link to the Linux Blender. Was not aware of it at all. Downloading now. Will be curious to compare my Linux results against your Windows results.

[MattTheTech]

KeithMyers wrote:
Thanks for the link to the Linux Blender. Was not aware of it at all. Downloading now. Will be curious to compare my Linux results against your Windows results.

No Problem! Can't wait to see them too!

[KeithMyers]

http://opendata.blender.org/benchmark/753d27b5-7908-4b5a-a86b-2ec774759a2c





scenes

namebmw27stats
resultOKtotal_render_time132.431device_peak_memory140.74device_memory_usage140.73render_time_no_sync130.97pipeline_render_time133.24

nameclassroomstats
resultOKtotal_render_time416.08device_peak_memory296.7device_memory_usage296.64render_time_no_sync415.412pipeline_render_time416.68
timestamp2019-08-10T02:04:47.255071+00:00device_info
device_typeCPUcompute_devices

nameAMD Ryzen 9 3900X 12-Core Processor
num_cpu_threads24
system_info
systemLinuxbitness64bitdevices

nameAMD Ryzen 9 3900X 12-Core ProcessortypeCPU

nameGeForce RTX 2080typeCUDAis_displayfalse

nameGeForce RTX 2080typeCUDAis_displayfalse

nameGeForce GTX 1080typeCUDAis_displayfalse
machinex86_64cpu_brandAMD Ryzen 9 3900X 12-Core Processordist_nameUbuntudist_version18.04num_cpu_cores12num_cpu_sockets1num_cpu_threads24
blender_version
version2.79 (sub 0)build_date2018-03-22build_hashf4dc9f9d68bbuild_time14:39:03build_commit_date2018-03-22build_commit_time14:10
benchmark_client
client_version1.0b2

So on the face of it with less cpu clock and less memory clock, the advantage goes to Linux. The cpu was locked at 4150Mhz and the memory was running at 3533CL14 Fast timings.

[MattTheTech]

KeithMyers wrote:
http://opendata.blender.org/benchmark/753d27b5-7908-4b5a-a86b-2ec774759a2c





So on the face of it with less cpu clock and less memory clock, the advantage goes to Linux. The cpu was locked at 4150Mhz and the memory was running at 3533CL14 Fast timings.

Damn that is Beautiful! I know you are running Ubuntu, but what specific version are you running, I want to boot up to it and see if I see those exact gains with my setup. You are KILLING IT!

[KeithMyers]

MattTheTech wrote:
Damn that is Beautiful! I know you are running Ubuntu, but what specific version are you running, I want to boot up to it and see if I see those exact gains with my setup. You are KILLING IT!

I am running the Ubuntu 18.04.3 LTS version https://ubuntu.com/download/desktop

Just grab the ISO and rufus from https://rufus.ie/ and burn it to a 16GB USB stick. It will create what is called a Live USB installation. Then reboot the computer and point the BIOS at the UEFI bootloader on the USB stick and in a minute or so you will be running Ubuntu. You can investigate how it looks and works. It is running solely from the stick and doesn't install anything to your drives.

There is an icon on the Desktop for installing permanently to your storage when you want later. I would recommend just adding a new cheap small 128GB SSD drive to your machine for the target for the future Linux installation. That way you don't have to deal with dual booting or deal with grub installers later. You just reboot the machine when you want to change the operating system by choosing which OS to run from the BIOS selection.

You could just run the Blender benchmark from the stick to but suspect the slow USB stick access might hinder that. Don't know if it runs mostly in memory or not. I know it has to load the files from the Blender folders to do the rendering.