porina

At my last job, I had to capture Bluetooth audio data over the air to show where problems were. Source, USB, Bluetooth link, end device. To capture Bluetooth data over the air, we had a commercially available Bluetooth logger. Think of it like sniffing Wifi, but dedicated for Bluetooth. So you just turn on the receiver and wait for the data? Nope. To successfully get decoded audio data out of the transmitted data, we had to capture the pairing handshake as well as some details on one end, which we controlled.  With that there was sufficient data to capture the correct data packets and decode them. Unlike Wifi, Bluetooth is frequency hopping so you can't just monitor a single channel.
 
Even in a lab environment that process wasn't 100% successful all the time, and it can take several goes to get the right data logged. If you're 10m away, the nominal line of sight range of standard power Bluetooth, you'll be lucky it works at all.
 
Maybe smarter people with better optimised gear (directional antenna) could do better than that. But already this is far from trivial. If you're attracting that sort of attention, you have bigger things to worry about than Bluetooth.

On that page you link are synthetic tests. The same capacity MP700 is about the only drive that beats it more often than not on the data as presented.
 

 
Skip forwards and the MP700 barely beats it overall, with a lot of the drives there much of a muchness. If anything the takeaway from this is that even a mid range drive hardly makes a difference in practice.
 

The PS4 originally came with a slow (reportedly 5400rpm) 2.5" HD and the system is designed around that. Putting in faster storage doesn't make much difference. I have a SATA SSD in mine. It hardly makes a difference to load times. I don't have the numbers any more as I did it many years ago but I'll see if I can find it later - might be somewhere on this forum. Because the original HD was so slow, the system heavily uses compression to increase the effective transfer rates. You hit CPU limiting almost immediately.
 
So basically if you want a big drive to store all the games, go for it. Perf really doesn't make that much difference.
 
Edit: see new post below!

On that page you link are synthetic tests. The same capacity MP700 is about the only drive that beats it more often than not on the data as presented.
 

 
Skip forwards and the MP700 barely beats it overall, with a lot of the drives there much of a muchness. If anything the takeaway from this is that even a mid range drive hardly makes a difference in practice.
 

Any hard drive will support any file system. There is no hardware correlation.

I'm not aware of a (reasonably-priced) USB <-> SAS protocol adapter.

Using an external dock with a SATA drive is well within the realm of possibility though.

you can, if you find the right enclosure for it

I will have had my 980 Pro 3 years later this week. That's been fine. By chance I think I missed the bad firmware era.
 
I have a 990 Pro in my current system. It's still quite new. I had one weird thing where after leaving my system idle at the desktop, I came back a bit later and found it at BIOS. I noticed the 990 Pro had disappeared. Warm boot didn't help. It came back with a cold boot and has been fine since. In searching around it seems there are many reports of 990 Pro disappearing which was concerning to me. There weren't the same reports for SN850X which I would have bought had it been in stock at the time. I tried putting the SSD to "high performance" mode where it doesn't sleep in case that makes a difference.

I went and got some data on this with my own system. I don't have Speedway since that is paywalled, so I used Time Spy instead.
 
dGPU: 4070 FE
iGPU: whatever is in 7800X3D. I even updated the driver for it.
 
Scenario 1:
iGPU disabled, dGPU enabled, 1 monitor connected to dGPU
Time Spy run on 4070: 17080 https://www.3dmark.com/spy/47848613
 
Scenario 2:
iGPU enabled, dGPU present, 1 monitor connected to iGPU.
Time Spy run on 4070: 16913 https://www.3dmark.com/spy/47848350
 
Result 1 above is 0.99% higher than result 2.
 
I did do more runs to check repeatability.
Scenario 1 with 2nd monitor connected: 17060 (-0.12%), 17105 (+0.14%)
Scenario 2: 16853 (-0.35%)
 
3DMark did give a warning with display connected to iGPU when running on dGPU:


Time Spy renders at 2560x1440, which is also the same resolution as my display. One frame (8bpc) is about 11MB. PCIe 4.0 x16 has about 32GB/s bandwidth. To transfer one frame over that link takes about 0.35ms.
 
So, OP confirmed? At least on my system. I'd call the 1% difference as insignificant. I can't test end to end latency. 0.35ms over PCIe link I'd call insignificant to me, although there may be additional latency at each end to send, receive then display that data.
 
I agree that if you have enough connections on dGPU, why go through this? I did have some weirdness and the result with iGPU enabled was my 3rd attempt. The first attempt, 3DMark just got stuck on gathering system info. Next attempt ran but didn't have working hardware detection. Only the 3rd run did it give full data.
 
I'm back to normal now. iGPU disabled. I have no need for it and it adds another chance for Windows to do something stupid, as well as stealing some ram.

Before this post I didn't know they operated in various countries. I've ordered online from them once in UK. 7980XE. It works. I wish they cleaned it better as there was a load of old thermal paste on it, and under that paste was some liquid metal that nearly got in bad places.
 
It depends on the specific item but in my case that CPU was below ebay pricing plus they have warranty, which I wouldn't have got in private sales.

@da na
 
Don't exaggerate too much. I don't think the difference is THAT much but yes at least in competitive games, there will be a slight noticeable difference I think. But of course there is the inefficiency of the display signal to pass through the whole motherboard and into the I-GPU, when you have a display port just a few centimeters away on the board.
 
You could go like this all along and not realize it was plugged in in the I-GPU.

Most systems these days will allow it.
 
Can be a bit finnicky and stuff but its been around for a long time. Its what a lot of people did with the cheap mining gpu's, tesla's,... to game on them. Use the cruddy intel hd igpu and a dirt cheap portless card.

I just saw a Japanese tweet showing that the benchmark scores do not differ when the HDMI cable is plugged into the motherboard or graphic card.
 
I tried to read the replies, and the reason seems to be that Windows has been supporting to let users choose which GPU to use, iGPU or dGPU, since sometime in 2018.
 
I think somebody did the test and found no significant difference when setting the PCIe slot to x16 or x8 (4.0), meaning about half of the bandwidth is not used.  So here is a simple calculation: If half of the PCIe 4.0 x16's bandwidth is usable, that is about 16GByte/s. Even a 4k 144Hz refresh rate only uses something like 5GByte/s, so it is possible that this has no negative impact at all.
 
Since I'm using a Ryzen CPU, which has completely no iGPU with it, I cannot have a try on it. Is there someone who has already tried before?
 
Will there be any other difference that the benchmark cannot show?

Maybe Linus can make a video about it.
 

I will have had my 980 Pro 3 years later this week. That's been fine. By chance I think I missed the bad firmware era.
 
I have a 990 Pro in my current system. It's still quite new. I had one weird thing where after leaving my system idle at the desktop, I came back a bit later and found it at BIOS. I noticed the 990 Pro had disappeared. Warm boot didn't help. It came back with a cold boot and has been fine since. In searching around it seems there are many reports of 990 Pro disappearing which was concerning to me. There weren't the same reports for SN850X which I would have bought had it been in stock at the time. I tried putting the SSD to "high performance" mode where it doesn't sleep in case that makes a difference.

As shown that's about 1.5% difference in scores. It would ideally need repeat runs to see how much variation there is and see if that difference is really there or just run to run variation.
 
With rounding, one frame of 3820x2160 at 8bpc is about 25MB, or 3.6 GB/s at 144 fps. Looking at it another way, the time to transfer 25MB of data at 16GB/s is under 1.6ms, relative to a 144 fps frame time of 6.9ms.
 
If the GPU needs to transfer from system ram it would eat into that bandwidth though.
 
I'm assuming 8 bit per channel RGB. I have no idea what format it might actually be sent in. Data usage will go up if you go to 10bpc for example, or maybe HDR.
 
The biggest one for me is that Windows can be really stupid at times and put GPU load on the wrong GPU. I'd always disable the iGPU where possible to prevent that from happening. iGPU may also consume some system ram which wont be available to other applications. Modern AMD ones seem to reserve 2GB as standard, which is significant if you only have 16GB in your system for example.

You need to adjust PrimeGrid project settings. Only CUL counts towards the Pentathlon, so de-select the other CPU subprojects. You should then also adjust the threading. For that CPU (8272CL), 1 task using all cores, or 2 tasks each using half the cores running at the same time might be optimal. More info on how to work it out at 
 
Once you have that sorted out, it might be worth aborting existing work and grab new with the new settings. You should then be able to complete multiple units per day.

You need to adjust PrimeGrid project settings. Only CUL counts towards the Pentathlon, so de-select the other CPU subprojects. You should then also adjust the threading. For that CPU (8272CL), 1 task using all cores, or 2 tasks each using half the cores running at the same time might be optimal. More info on how to work it out at 
 
Once you have that sorted out, it might be worth aborting existing work and grab new with the new settings. You should then be able to complete multiple units per day.

I wrote the earlier post describing all this. I might have an idea what is going on 😄 
 
Ok, that might have been a point of confusion. Yes, I meant one task using all cores. Not one task per core. 
 
Timings for 2880K FFT length (8 cores, 1 worker):  0.85 ms.  Throughput: 1170.20 iter/sec.
Timings for 2880K FFT length (8 cores, 2 workers):  1.65,  1.64 ms.  Throughput: 1217.22 iter/sec.
Timings for 2880K FFT length (8 cores, 4 workers):  3.26,  3.22,  3.23,  3.23 ms.  Throughput: 1237.28 iter/sec.
Timings for 2880K FFT length (8 cores, 8 workers):  8.20,  8.13,  8.16,  8.11,  8.06,  8.02,  7.99,  8.09 ms.  Throughput: 988.30 iter/sec.
Timings for 2880K FFT length (8 cores hyperthreaded, 1 worker):  0.83 ms.  Throughput: 1203.82 iter/sec.
Timings for 2880K FFT length (8 cores hyperthreaded, 2 workers):  1.58,  1.58 ms.  Throughput: 1264.59 iter/sec.
Timings for 2880K FFT length (8 cores hyperthreaded, 4 workers):  3.10,  3.10,  3.07,  3.08 ms.  Throughput: 1295.83 iter/sec.
Timings for 2880K FFT length (8 cores hyperthreaded, 8 workers):  9.74,  9.35,  9.44,  9.35,  9.20,  9.30,  9.22,  9.35 ms.  Throughput: 853.96 iter/sec.
 
A fresh bench with last night's FFT size. I haven't checked again for now. The drop off in performance on the 7800X3D isn't that bad even if I were to run 1 task per core. Only a 20% loss in throughput compared to 4 simultaneous tasks.

I also show HT here. There's hints it might be low single digit % faster but again, on past testing with Intel CPUs I found it significantly increased power usage for near margin of error differences in performance. I haven't repeated that testing on AMD yet. There was also the danger that HT can give the impression of increasing performance, when it was effectively reducing loss from other things using CPU cycles.
 
BTW I was running 2 tasks earlier because that balances individual run times and gives me flexibility of suspending one task if I want to free resources for other activities on the PC. I want to try 1 task for a more direct comparison.
 
Yup, Windows does that. Hence my mention of Process Lasso as a workaround some have used in situations like this. This can also happen with multiple-CCX AMD CPUs.

I disabled HT, problem solved 🙂

So basically worst case it is 2 real cores so it could be very roughly around 2 days run time for 4 vCPUs.
 
 
Elsewhere: now that some CUL units have gone through, I can see the relative throughput of various systems. Intel systems: 1 task using all cores. 7800X3D 2 tasks of 4 cores each, which I'll use as reference performance = 100%:
7980XE: 21 to 30% faster
7920X: 3 to 7% slower
11700k: 54% slower
 
I'll try 1 task all cores on the 7800X3D later.
 
Take above relative performance to be VERY rough since it can and will vary with the FFT used in the task, and I may have used the systems while running which would impact timings.

Speak for yourself 🙂 I just started off some system and one was estimating finishing a little before the start. Pausing it a little to make sure that doesn't happen. That reminds me, I also need to switch to LTT team for PrimeGrid as I'm normally with another there.
 
Also, these units are bringing out coil whine in my PCs that I've never heard before. What is it doing? 😄 

Thanks a lott ! It works perfectly fine and it back to 6 core 12 threads

The number of cores could be limited in either the BIOS or Windows. If you have tried resetting BIOS and that doesn't help, let's check Windows.
 
In Windows, run msconfig and go to "boot" tab, and press the Advanced Options button. You should see the above. Is there a CPU limit applied in top left?

I *think* it is a combination of:
1) Nvidia working with far more system integrators and thus getting a much larger part of the prebuilt system buyers (which is the majority).
2) AMD only being competitive in certain aspects but being behind in others. 
 
I also think that people on this forum overestimate the value proposition of AMD. From what I have seen, although this varies a lot by price point and country, Nvidia and AMD have fairly similar price-to-performance ratios. I think the real myth is the whole "Nvidia is expensive and AMD are the price:performance kings".

Generally you only want to run on real cores. You can either set the max threads per task accordingly, or adjust the use CPU % in boinc settings to lower it. You'll have to do it after this unit completes as it can't be changed once sent out. 
 
One task. This falls under the mesh cache special case. They have a lot of L2 so it'll be fine. If they didn't have the L2 it'll still be 1 task, but it would be more memory bandwidth impacted.