Overclocking Threadripper 7000 on the Gigabyte TRX50 Aero D!
Vložit
- čas přidán 6. 12. 2023
- www.newegg.com/gigabyte-trx50...
**********************************
Check us out online at the following places!
linktr.ee/level1techs
IMPORTANT Any email lacking “level1techs.com” should be ignored and immediately reported to Queries@level1techs.com.
-------------------------------------------------------------------------------------------------------------
Intro and Outro Music By: Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 3.0 License
creativecommons.org/licenses/b... - Věda a technologie
Wendal, bobbing head on case was so awesome. Do it always plz. And thx for being you.😊
Looking good Mr Wendell sir. Very nice build . I'm considering Arc for my next GPU. Only because it's cheap lol. I'm sure it will work. If not I will have fun trying
Looking good wendel
New to the channel, i love your content :=)
Like you said kind of confused by this board but Thunderbolt 4 and the PCIE slot layout also sets it apart from the others. Personally more interested in the ASrock board but it's nice to see something aside of the Asus get reviewed.
This was incredibly useful as usual!
Awesome content and thanks once again.
I’m looking to refresh my 3970X, AORUS Master, 128GB 3200MHz workstation to a 7960X, Gigabyte AERO D, TBD on the RAM. I didn’t realize the sTR5 uses RDIMMs, so memory seems to be a bit more expensive.
On the last HEDT mobo video I complained that there was no Thunderbolt, and look at this! Actually pondering giving this platform a chance.
Are you currently using thunderbolt, or is it a feature you *want* to use/experiment with?
@@boneappletee6416I am using an Apple Studio Display as my main monitor, so I am using TB every day.
@@boneappletee6416 Thunderbolt/USB4 is actually pretty handy. Pretty much the only way you can connect an external NVMe SSD with half-decent speed (I reach a bit over ~2 GB/s with an old recycled 2TB 970 EVO but could be a bit faster with a faster drive).
I use it for occasional dual boot if I need to run Windows when I want to do some gaming.
I also have 2 hot-pluggable NVMe slots using U.2 connections (so this is full speed) but these are really too expensive to waste them on only running a game now and then.
TB/USB4 is really good enough and fast enough for the purpose for me. And it is at least twice as fast as any USB 3.X Gen X
I have a benchmark request, and it's for my personal "extreme" use case:
I have been setting myself up to start a content creation business that involves RED cameras, thus meaning I need a machine that can handle R3D RAW files with the native resolution being 6K.
The build I started with had an RTX 2080, a Ryzen 2700X, 16GB of CL14 3200MHz DDR4 and a 4TB Gen 3 NVMe drive. That was enough to playback the files smoothly, but it would get choppy after an hour or so and proxy files wouldn't help. I then resorted to getting an RTX 4090, silly combination I know but it did get the job done at least in the playback department. It would, however, slow down once some effects and noise reduction are added.
Since then I stumbled upon some other major issues my build, and any other regular PC build I could make right now:
1 - As the R3D files take up a whole lot of space (a 1TB disk can be filled in just 30 minutes of recording) and require a decent read speed to be worked on, I bought an external Thunderbolt 4 storage device which can hold up to 16TB of NVMe drives and filled it up.
2 - You might wonder why I bought a Thunderbolt device having an x470 motherboard, this is because some tech channel had reccomended Intel platforms for creators as having Intel Quicksync through the iGPU in video editing programs is kind of a must have. I therefore was looking to switch to an Intel platform and bought a Thunderbolt device ahead of time. That however poses a problem that is going to be my next point.
3 - Thunderbolt and iGPUs use up 4 PCIe lanes each, considering I already have an internal NVMe drive and the RTX 4090 eating up 4 and 16 PCIe lanes already the total lanes I need jump up to 28. That means neither Intel's or AMD's current desktop chipsets will suffice, considering they offer 16+4 and 20+4 respectively. That's even more important if you consider that the card reader I must use to transfer the video files from the external cards my RED records on to, to the pc, uses a Type C connection that will need to pass through the chipset and I also want to run the editing program and OS on a second NVMe drive that's faster than the 4TB I am currently using. That pushes the needed PCIe lanes to 36, without considering the fact I might need to add a second GPU in the future.
Note: passing the files from the card reader to the PC is a dreadful procedure right now, as both the drive I'm writing to and the card reader use the same PCIe lanes to communicate through the chipset.
4 - While the RTX 4090 is a beast of a graphics card it is still not enough to work on the R3D files I have without running into issues. As of now I'm using the card in DaVinci resolve to deal with the decompression + debayering setting, which will eat up to 70% of the card's processing power alone. That means if effects and noise reduction are added to the equation the card will not be able to stand the workload. The only way to resolve that is if the CPU was capable of handling decompression and debayering on its own, leaving the graphics card free to do the rest.
All things considered I've come to realise the only possible solution is to either buy a Mac Pro or go for a Threadripper build with USB4 just like the one you've shown in the video. Every tech channel I know of usually just uses the PugetBench benchmarks for DaVinci Resolve, however the benchmark tells very little on how a system would behave while handling R3D files and even less on how it would handle PCIe lanes communication.
I would therefore like to see some kind of "real case scenario" test of this platform, with an RTX 4090 added in.
I could supply the R3D files as well as a project that's been worked on, let me know.
Thank you.
Side note: it seems like companies like Gigabyte know damn well there's people out there, just like me, who need this kind of platform and its features. I suppose that's why there's only two 16x slots, I probably wouldn't ever need more. It's also strange to see how desktop chipsets are so limited, considering the most recent AMD and Intel motherboards prices.
Same usecase here, man
I know sod all about your usecase (wish I had it as an excuse frankly) but I'll throw in a quick 2 cents.
-AM4 is 16gpu+4nvme+4chipset lanes, AM5 is 16+4+4+4 for 28 total.
So a 7950x build could have NVME x4 boot/program drive, x16 rtx4090, x4 to asmedia USB4 controller which your "REDcardReader" connects through all from the CPU, plus there's a final x4 downlink to the B650 chipset for more M.2 etc.
The readers meant to be 20Gbps using a USB3.2 gen2x2 link but do you actually see anywhere near these speeds? Or is it like the "14GB/sec" speeds on NVME drives at unrealistic que depths? Because if you actually tend to sit at like 600MB's I might not worry about it fwiw.
Motherboards are higher despite the limited chipsets because AMD/Intel think we all want to run 5 separate pcie5.0 NVME drives for some reason, If used is an option and you wanna go cheap then x299 and X99? (power draw aside) with 40-44 pcie3.0 lanes, but you could just run the 4090 at pcie4x8 for same bandwidth to free up some lanes, don't know the performance impact in your use case though.
If you do go TRX50 I might suggest a quad NVME pcie adaptor rather than that thunderbolt 4 storage enclosure, anything you plug into the back ports of a PC always seems slower than any other option.
Again just my 2 cents, Sorry I can't help more.
@@CrimzonRX I feel you brother :D
@@tomstech4390 Thank you for your insights. I'll try to clear some things up so pheraphs you can help more:
1 - The reader is USB 3.2 Gen 2 but caps at 10Gb/s, so has a theoretical max transfer speed of 1.25GB/s. However the cards are CFast 2.0 with a claimed sustained read/write speed of 480MB/s.
I just tried to do a test transfer (221GB total) to see current behaviour, since I formatted a few days ago and all is freshly installed. Speed capped at 350-ish MB/s, held somewhere around 300 MB/s for little more of half the transfer then went down to 200 MB/s max and 110 MB/s average while dropping to 0 from time to time. It gets worse with longer/larger transfers and sometimes starts out with the drop to 0 behaviour already.
Mind you I'm currently only running a single NVMe to try and avoid the issue, as I had previously noticed the transfer would exhibit the same sluggish behaviour with both NVMes installed, especially while prompting the first drive (M.2_1) to do any reading or just running any program that would use the disk. I thought it were the drive having its cache filled at first, but it also happens when using a Gen 4 Samsung 980PRO alone by itself so not sure.
I really can't tell what's slowing down the transfer at this point, might as well be multiple factors.
2 - I was unaware the motherboards could offer more PCIe lanes, as to my knowledge even the 7950X would only offer up to 20 lanes through CPU and 4 through the chipset. Will have to do some digging into that.
All in all the new motherboards don't seem all that appealing to me as Gen 5 uses 8 lanes and filling up the M2 slots might just turn into a nightmare.
3 - Haven't tried switching the 4090 to 8x yet, however seeing as it can get loaded up to the point where even that slows playback down I suppose it would introduce more issues. WIthout having a CPU that handles decompression and debayering that is.
4 - X299/X99 would resolve the PCIe lanes population issues but also can't offer Thunderbolt 4. Waiting for a newer platform or refresh on those two is a better idea imo.
5 - Yes, quad NVMe through PCIe would be faster and it is something I have thought of. However, the external enclosure is a portable solution I specifically chose to bring around with me during working trips. Firstly because I can't bring my desktop everywhere and need plenty of data that is moved fast, so a quad NVMe device that can fit in a backpack is the only practical solution I could think of. Secondly because I might not only need the files myself, I've recorded 4 day events alongside third-party crews in the past and being stuck transfering files to SSDs was something that ate us in the back a couple times. I need to make sure I have an empty CFast card at any time, while also being able to bring my files home and share them to colleagues/collaborators.
I really believe TRX50 is the only solution that makes both financial and practical sense right now. I'll probably wait for CES on 9th January to make a decision though.
I look forward to CXL memory testing, if you managed to get some.
I can really see myself perhaps use a Threadripper (none Pro) with "only" 128-256GB RAM with 1TB+ CXL memory instead of 1TB RAM TR Pro.
Use case is compiling (mostly UE5 main dev & shaders) and running other memory intensive computations (latency dependent)
Gigabyte TRX50 Aero D is real nice
thunderbolt! im on this! thanks mr wendel!😀😀
Ty❤
Very good point on what quality market leadership should look like. I always got in too early on every AMD product i ever got, with very bad driver support and just plain weird bugs stemming from 'a single dev getting it to work somewhat'. After feeling betrayed by this bad quality a few times it got to a point that i am never buying an AMD product again. Even if it is better on paper, it will probably let me down in practice. If only the drivers and ecosystem support where anywhere close to Intel or NVidia...
How I wish I co8uld get a Torrent XL... I have a normal Torrent now and I really like it, but to get 3 180 fans in front, a bunch of HDDs, more room inside for threadripper and 2 GPUs, that would be epic.
Hey Wendell, can you check if S3 sleep (suspend to RAM) is available on this board? I built the same system a week ago and found that S3 sleep is not available. I don't think the feature is enabled in the UEFI with no way to toggle it on or off unlike the ASRock TRX50 board.
Question: Where is that interesting fan bracket mounted? I'm using the Aero D with a 7960X and Renegade Pro RDIMMS (QVL). There seem to be stability issues with that board using the EXPO 1 profile (6000 MT/s CL32). Those could be thermal issues, in my opinion.
So which is the best TRX50 board to get?
Question for those that might actually know (both of you), In what situation will you really require these pcie5x16 lanes (not just M.2) where pcie4 doesn't suffice?
Because allot of the cost is the pcie5 and even pcie4 signalling integrity and shielding, if they could make a pcie4+pcie3 version for $200 less that would be awesome, but given the platforms lifetime maybe with rtx5000 and rdna4 cards that pcie bus might actually be utilised in the future so worth the extra money, who knows?
Oh man I want this board now. Just need AMD to release 12 and 16 core non pro Threadripper now. I'm not going to pay the amount I should be able to get board + CPU + 64gb of memory (in HEDT land) for on just the CPU because HEDT is dead and now it's just Lite Workstation. I would content with just 16 cores or dropping down to 12 cores to get the board, CPU and RAM for the $1500 price point. I mean I do still shudder at the thought of $1500 as the entry price point for a base platform, but I could live with it.
If you only want 12/16 cores just get a 7900X or a 7950X, you don't need HEDT. AM5 is a much cheaper platform anyways.
@@PandaMoniumHUN Maybe I want more PCIe connectivity?
@@PandaMoniumHUN Stop telling people they don't need things they do. I would never overpay so much for too many cores on babies first CPU socket. Save that socket for the Dell and HP pre-builts. I want the ability to install an enterprise network card, to be able to install more than one M.2 SSD worth of bandwidth outside of the GPU slot.
@@garrettkajmowicz Yeah, there is no way I would pay more than $350-400 for a CPU that has just 20 PCIE lanes. The 12 and 16 core CPUs belong on HEDT, not the same socket that goes into dell and hp internet surfers. But hey, if AMD wants to make it available for both - I'd be OK with that - but not consumer socket only.
Pcie lanes situation is f terrible. I couldn't care less about 32 or 64 cores but consumer platform is so starved on pcie its horrifying. Recently had to buy SATA SSD's from 2017, that were same price as 5x faster m2 ssd, cause there was no way for me to connect more than two m2 sticks.
Anyone have link to the 3D printed RAM dedicated fan bracket?
i got two thermal sensors and a noise sensor. u can select the thermal sensors as fan regulators but not the noise sensors (x570 master) the noise sensors are the most useless features ive seen. (except cpu usage led onboard - they never worked)
12:02 Wow, 5.8 Ghz turbo on a Threadripper? Not even my 7950x turbos to 5.8 Ghz. I've seen it go up to 5.7 Ghz at most and that's very rare. It's mostly hitting 5.5 - 5.6 Ghz in single core tasks. That's with an Arctic Freezer II 360 AIO and plenty of airflow.
Thanks for the overview! Is there any way to connect u.2/u.3 drives to this board without giving up one of the pcie slots for an aib? Any minisas or other connectivity for storage besides m.2 or sata?
I hope to get something with thunderbolt. Even if it's just some random used laptop.
"RAM-training" ... just the first boot? My X99-era system (old I know) does that with every boot, training 64GB's takes longer than a cold boot of Windows. I sincerely hope this issue has been taken care of over the years.
There is likely a setting in the BIOS to boot quicker by not training every time.
@@cracklingice If I recall correctly, I searched for that years ago, only to find reports that disabling it makes the system unstable.
@@bartgrefte I run an X99 Taichi. It takes several minutes to boot, but most of that is down to having external USB HDDs. I seem to recall having the skip memory feature enabled, but I also did not run crazy fast memory. 2400 CL 17 iirc. Tho it was two 2x4gb kits and two 2x8gb kits for a total of 48GB. Now it's eight 32GB 2133 RDIMMS. Using really fast memory for the time would likely result in the unstable comments. Stick to 2400 and you should be set.
@@cracklingice The X99 are able to use registered memory? That's intersting. I might have read it but don't have recollection of it at this point. Thanks for mentioning it :)
@@akirafan28 Have 256GB of RDIMMs on my X99 Taichi right now. The max potential supported. Do be aware the LRDIMMs are not supported and 32GB comes in registered and load reduced flavors with everything larger being load reduced only.
Does this cooler provide required pins and socket to attach at motherboard....??
👍👍👍👍
Damn Wendell looking good and handsome.
Should've had more than 4 memory channels, like EPYC Rome.
There are 4 boards for TRX50, Supermicro also has one.
Warming up the credit card, model #?
@@Level1Techs H13SRA-F
@@Level1TechsI wouldn’t bother with the H13SRA-F, its PCIe design is wastefully horrible.
Yeah thought it was a different model. The board layout and powr delivery oh noes
There are four lights!
title typo oops
Intentional mistakes increase community engagement because people can't resist to tell you when you screwed up
@@BGraves People say this but everytime i've seen a title typo its always fixed afterwards. I don't think most creators metagame that hard
@@snowwsquire Linus Would.
@@BGraves Linus metagames by swapping out the titles entirely. Hes on a different plane
Almost 2024 and we still have a car key battery on the motherboard - why not a little lithium ion battery that charges when needed?
Probably because the coin cell batteries have become an industry standard
Can this MB support standard DDR5 also?
Does anyone know of any driver problems on linux with the Gigabyte TRX50 Aero D? ANything that wont work?
No level 1 show today?
Does this motherboard have enough space to fit 2 RTX 4090s with the PCIe slots? Can anyone confirm?
Yes sure
@Negodyay77 is there space between for air or do I need to think of custom placement solutiom if both are air cooled ? Getting vertical riggings etc.?
Optane P5800X
How terrible is AMD software RAID on this platform? I’m gonna be building a Windows workstation and I’d like to run mirrored boot drives, but I’d rather not deal if I’m chasing raid issues.
Thunderbolt or nothing
Can I run dual 4090 on it?
If you make them physically fit, why not
@@sznikers
This is my question, Can I fit two regular 4090 in there :D
@@ThePatrvq waterblock is how you make em fit
Talk a little slower so I can process this outstanding information as it is coming at me.
AMD Threadripper 7960x 5,841ghz BIOS OC ASUS trx50 HEATKILLER IV ax1600i vs cpuz r15 r20 r23
3pci slots, that is a joke indeed. 100GB nic, quad port san hbas, graphics card, thats it..
AMD is _fantastic_ at making chunks of their customer base feel like second-class citizens. Just ask anyone who's ever tried to use Radeon in Blender.
01:49 JHL* ... there is always that ONE GUY
So glad this is informatinal instead of linus nonsense spam
Still can't stand his smug face again.. hate bullies
Loud Noises!!
Completely different content for equally different audiences. There's absolutely no reason to crap on LTT for adressing people on a different level - and tbh I think that Wendell would be the first to agree.
@@danieloberhofer9035absolutely. That's like saying Top Gear is bad because it's not in depth about the specs and engineering of cars.
(There's other, valid reasons to crap on LTT, but that's another topic...)
Go eat a snickers plz…..
The video title is missing in the content. Just bla bla bla. Where is the overclock itself in the video?
Its hard to imagine anyone paying the premium for threadripper if they don't need the pcie
Perhaps I'm just not seeing the use case
The thing is, there's always a usecase for any highly-specific niche setup... And you won't have any exposure to it unless it's something you stumble on or need to implement.
@@boneappletee6416 fair. I'm sure there's someone out there for whom it makes sense to pay an extra 1500$ for increased memory bandwidth versus am5.
The use case is you want 32/64/96 cores and high memory bandwidth.
I don't see anyone wanting 96 cores without memory capacity and io to feed it. There are much cheaper options for high core count low bandwidth cpus
@@christopherjackson2157 For 96 cores I agree that non-pro Threadripper might be an edgecase, but show me where else you can get 4 memory channels and 64 cores "much cheaper" and I'll eat my hat.