You are using an out of date browser. It may not display this or other websites correctly. You should upgrade or use an alternative browser.
GPUs & CPUs & Enthusiast hardware: Questions, Discussion and fanboy slap-fights - Nvidia & AMD & Intel - Separe but Equal. Intel rides in the back of the bus.
GN has published the failure analysis on this now and spoiler alert: The redditor was a fucking retard and killed the CPU by misaligning it in the socket.
Between this and the melting 14th gens, I'm wondering what Intel's testing process for their desktop CPUs looks like, since "we ran the CPU fully loaded for 36 hours" obviously wasn't part of it.
In other news, real-world training MI300 performance is about half of the H100 (link | archive). I'm not in AI world, so many of the things NVIDIA does to get really good AI performance aren't as relevant to me, but I can say that with AMD, you have to test each new release of ROCm to see if it actually works. It's not uncommon for them to push out a new version that won't compile.
GN has published the failure analysis on this now and spoiler alert: The redditor was a fucking retard and killed the CPU by misaligning it in the socket.
GN has published the failure analysis on this now and spoiler alert: The redditor was a fucking retard and killed the CPU by misaligning it in the socket.
If you don't need the SFP port (and you probably don't), look into the hEX refresh (E50UG). The hEX S (RB760iGS) has the same MMIPS CPU as the old hEX (RB750Gr3) you've linked, while the refresh has an ARM CPU that as seen in these tests, manages to do over 200Mb/s on Wireguard. At the same price. If you feel like spending a bit more, the RB5009 is probably the best homelab router Mikrotik has on offer. My hAP ac2 is just fine for my use cases so far so I wasn't compelled to upgrade just yet. My ISP closes out at 300Mb/s and the CPU is utilized at 50% when doing a speedtest. "Some" words of advice from someone that fucked around with RouterOS at home for a few years now and made a whole bunch of mistakes, including a couple of device lock-outs.
1. Do a regular backup of your configuration, both by /system/backup and /export show-sensitive. The former is device dependent data blob and the latter is just text. The reason I say that is because I used to do just the blob one, until at some point after upgrading to ROSv7 it bricked my router, and for some reason NetInstall failed to properly reimport it, and I had to manually dig out bits of the configuration from the backup and then manually recreate some. Spare yourself the trouble and keep a plain text backup on hand as well.
2. Be very careful when you configure the firewall, you have no training wheels here. If you delete all of them and you put the router as your gateway, be ready to see just how vulnerable you are when connected to the Internet, and why Windows XP computers that weren't behind a NAT got so ass-fucked back in the day. Some anecdotes from my initial days with RouterOS.
When I wasn't properly filtering out the input chain I'd be getting countless failed attempts at accessing the SSH from various IP's around the world. Another time I had no idea what I was doing while configuring Pi-Hole and ended up opening up my DNS to the world. This resulted in countless requests from random devices on the Internet and at least a few cases where the DNS request was hijacked and redirected me to some dodgy porn site. Honestly I'm shocked it only ended on that, and not on a phishing scam or a full-blown malware compromise.
Later on I fucked it up again, but this time I knew to not open up my DNS, because I didn't need to do so even for Wireguard, and to lock down the service ports to local IP's only. I got failed access attempts again, and as it turned out I messed something up in the default firewall rules, and at that moment I got the gist of how they worked. Well, that, and after another lock-out and config reset, but by this point I was prepared for it.
So, here's a little taste of how configuring RouterOS is like.
First, let's explain what the input chain is. It's the filter chain of the traffic coming to your router. If there are no rules in place, anything that gets thrown at the router passes through. If the packet didn't get filtered at any point in the chain it also passes through. This means that if your router is the gateway of your network, or in other words it's the device that faces the Internet directly, it's very fucking important to get this right, or else you'll be opening up your network to every Ivan, Zhang and Manjeet. Similarly, it's very important you don't cut out too much, as that's also the chain through which you manage your devices.
So, the main configuration on the default input chain on the hAP ac2 looks like this:
Code:
add action=accept chain=input comment="defconf: accept established,related,untracked" connection-state=established,related,untracked
add action=drop chain=input comment="defconf: drop invalid" connection-state=invalid
add action=drop chain=input comment="defconf: drop all not coming from LAN" in-interface-list=!LAN
First, it lets in whatever was already established by devices from your network, it opens up little pathways for what's surely something you requested and not someone from outside of your network.
Second, it drops any invalid packets. If it's dodgy and it doesn't belong, it's out. Simple.
Third, it drops anything that's not coming from interfaces on the LAN list. If it didn't originate from any of your local devices as defined by the LAN device list, it's dropped, such as packets from the WAN device list.
While I could write another paragraph about the issues relating to the interface lists and how much of a hierarchical clusterfuck RouterOS is, but let's focus on the most important aspect of this configuration: this rule is the only thing keeping your router away from the Internet. If a packet doesn't get caught by that, it passes through. This isn't really secure, and for some reason Mikrotik uses this as the default config for SoHo routers, while on their documentation they present a much safer solution.
As an example, here's the gist of my current configuration:
I moved the "Drop invalid" rule as the first one just so that gets caught before anything else, removed the "untracked" connection states since if you don't know what they are, you don't need them, established and related are all that matter, then I explicitly allow my LAN devices to access the router, after which I drop every single packet.
If you've managed to learn anything from this autistic tirade so far, you might notice what was the moment when I finally got the hang of the input chain. It was when I disabled the router access rule with the "drop everything else" rule by the end to check if it did anything and then I logged out of WinBox. Well, it did something important, and that's allowing you back in to the router. WinBox didn't autodisconnect after disabling this rule, which made me think it was useless. But the reason it didn't was because the "accept established,related" rule was maintaining that connection. Once it was closed, I wasn't able to open another one. It also explains why the default SoHo configuration is like it is. It is less secure, but it has less points of failure for newcomers.
So, when you do this approach, it's very important to have a way to get back into the router set up like that. You might also notice that you have a fine granulity of how you filter traffic. I only allow ports for SSH, HTTP and WinBox, over TCP, from LAN interfaces, and from an address list where I defined the local subnet and WireGuard addresses. You don't have to go so far but it feels nice to do so. I also locked down the router services themselves like that, just an extra safety measure that helped when I did that last misconfiguration.
Anyways, with this configuration you now have to explicitly allow traffic to your router, which makes it much safer. You can then check your connection with GRC's ShieldsUP!. If it passes all green, it means you have a properly configured router, and nothing you didn't explicitly allow is seen on the web. Technically it's also wise to do the same on the forward chain, but I'm yet to get the hang of that. Last time I've tried I've cut out all the traffic between LAN devices due to the lack of proper allow rules so I left them at default for now, the input rules matter more anyways.
Now, if you thought all that was difficult to wrap your head around, then you should avoid MikroTik routers, or get a cheapo one to play with in a home lab to get comfortable with it before using it as your gateway. The plus side of how granular and convoluted it is is that it forces you to learn the basics of how networking works.
3. Adding to what was said before, consider configuring a management (MGMT) port. The neat part about MikroTik is that you can configure any port to do anything in a myriad of ways, so for example you can dedicate a single port to serve as an emergency config port. RouterOS has something called a MAC WinBox server which lets you connect to WinBox even if there is no working TCP connections, but by default it applies to all LAN devices, so it's not very secure since anyone on your WiFi could use it as well. You can create a new device list called MGMT, set MAC WinBox to that, and now it'll only work on that port. This is a well documented topic so you'll find plenty of examples.
Remember: with RouterOS, the way you set up your network is up to you.
4. SoHo MikroTik devices, such as the hAP series or the Chateau series come with a default configuration that includes things like firewall, LAN and WAN device list separation and others to make them easier for newcomers, as well as a MikroTik Home app for Android meant for an easier first-time setup. I strongly encourage you to instead recreate that configuration from scratch with full understanding of what each thing does, especially with the aforementioned less-than-optimal firewall defaults. Understanding RouterOS like that is crucial, these are not beginner devices.
5. MikroTik sucks dick when it comes to WiFi. For the longest time they've used their own WiFi drivers which led to heavily degraded performance, no matter how well you configured it. Things have improved once they brought in their 802.11ax devices, while the competition was already moving to 802.11be. By that point RouterOS utilized Qualcomm drivers, but only for ax devices. With RouterOS v7.13, they've finally gave ac device users those drivers, so my hAP ac2 is actually serviceable now.
Speaking of ax, MikroTik half-assed it as well. Even their highest end Chateau routers offer less MIMO and bandwidths than the competition, and not a single device of theirs has the 6GHz radio for ax, which is probably the biggest draw of upgrading to ax if you have something like a Steam Deck. For that reason people tend to use wired MikroTik routers as the base of their network and then supplement WiFi through other vendors like Ubiquiti, with all the issues that entails. Or get some chink router, put OpenWRT on it and use that for your WiFi, whatever you believe will work best.
As for me, the hAP ac2 is good enough after the switch to wifi-qcom-ac and configuring it, I don't own any ax devices anyways. Again, it's up to your needs. If you just need some wireless connectivity it's not a big deal. If you're into the bleeding edge tech, you'll want to do a proper network design instead of an AIO magic box.
6. Avoid WinBox 4 if you can. If you're on Windows, it's piss easy, just run WinBox 3 and enjoy it's simplicity. If you're on Linux and already have Wine set up, same ordeal. If you're not willing to rebuild the Win32 environment on non-Windows OS', you probably will do so once you compare how botched WinBox 4 experience is. Unfortunately with MikroTik, once they fuck something up, they don't haste to fix it. You'll quickly learn it's a very abusive relationship, but if you're using Linux or tard-wrangling Windows already you'll get used to it.
7. Whenever you mess around in WinBox, remember to enable Safe Mode. In case your connection to WinBox gets severed, all the changes you were making get unrolled. In theory it sounds foolproof, but in practice, as you've seen in my case, it can be deceiving, understanding of RouterOS is still a must. Be sure to try and connect via your phone to your router before disabling Safe Mode if you don't have an alternative way of router management and you don't want to reset your router.
8. Get familiar with NetInstall, if shit hits the fan it has the potential to unbrick your router, such as when running v6 > v7 upgrades, or if you have an older device with limited storage space. You should do the v6 > v7 upgrades with NetInstall anyways, the clean slate will help you avoid any odd issues in the future. And do remember that MikroTik has two points of updates, the packages, which is the RouterOS software and it's packages, and RouterBOARD firmware. But I'm pretty sure every MikroTik beginner tutorial will mention it.
9. Do not ignore the RouterOS console. Sooner or later you'll be learning it for RouterOS scripting, because dear lord you can do a LOT with it. For example, due to the specifics of my ISP, I've created a script that'll automatically re-roll my IP by restarting the PPPoE session until I roll a new one, useful for bypassing shit like 1fichier IP limits. Sky's the limit when it comes to scripting, and once you get the hang of the rest of RouterOS like Netwatch, you'll be able to bend it all neatly to your needs.
10. Even though technically MikroTik has superb update support, be aware that with every major version release, older models end up performing worse, and in some cases lose functionality, as maintaining the gigantic feature set of RouterOS while still making sure it'll install on the 16MB of FLASH on the oldest RouterBOARD models is no small feat.
So, some main takeaways:
-MikroTik and RouterOS are user hostile, but once you know what you're doing you can do a lot. It takes three separate WinBox menus to set up a DHCP server compared to a static web UI box on other routers, but soon you'll realize that this complexity gives you great configuration granulity.
-MikroTik's configuration is much like networking itself. Segmented, convoluted, hierarchical. Both a blessing and a curse. If you're used to traditional training wheel routers, it'll be a hellscape to navigate. If you tackle it head on, you'll force your brain to understand networks by the very nature of using RouterOS.
-If you're looking for the latest in WiFi, get something like an RB5009 and a separate AP. It'll be an equivalent of getting the best headphones, mic and amp separately instead of an all-in-one USB gaming headset. It'll be ever so slightly less convenient, but you'll be able to get much more quality with your purchases in every segment, without being forced to change everything if only one part of it becomes a weak point.
-If you still don't know basics such as the separation between your local network and your ISP's network like @Betonhaus, do consider that perhaps you're not in the position to use a MikroTik just yet. Get through networking fundamentals, and at most get a cheap MikroTik, even on the used market, to crash that theory with practice in a home lab before you put a MikroTik into production.
And a quick buyer's guide:
-do not buy older models brand new. For example, the hAP ac2 got a price hike despite being ancient by this point, and nowadays you have the hAP ax2 and ax3 to choose from, which are much better choices. If you really want the older ones, look at the used market first.
-if you want a decent AIO SoHo router, get a hAP ax3.
-if you want a wired home lab powerhouse, get an RB5009.
-avoid the L009. It has the CPU performance of an ax lite and is one of those MikroTik mistakes that people don't tend to be aware of. The ax3 and the 5009 are still the GOATed two.
There's more I could say on the matter and my experiences or cool shit I've done w/ ROS like finely tuned auto-disabling port forwarding on a dynamic IP, or point out other cool shit you can do with it like using the USB port w/ an Android smartphone as an LTE modem fallback, but I spewed out enough drivel no one will care to read anyways.
Dear nerds,
I need to upgrade my GTX 1060 3GB, it can run RDR2 at mid/high settings surprisingly well but is pretty old now.
Is the 6750 XT a good replacement? Am hoping to get something that's a major upgrade, should last a few years, and cost less than a PS5 Pro. I'm just an occasional gamer so don't wanna spend too much.
PC specs are:
i7-8700
16GB RAM
Nvme SSD
On AMD's side, you're probably going to see RX 9070 XT and 9070 non-XT (Navi 48) first, and the 9060 and below (Navi 44) later e.g. March. Navi 48 adds "B-frame" support to AV1 hardware encode, while Navi 44 lacks hardware encode entirely according to Igor's Lab. That brings to mind the 6500 XT video situation although that was worse.
Is the 6750 XT a good replacement? Am hoping to get something that's a major upgrade, should last a few years, and cost less than a PS5 Pro. I'm just an occasional gamer so don't wanna spend too much.
It sounds fine to me. @The Ugly One owned a 6700 XT and is going to tell you it was one of the worst GPU purchases of his life. The 6750 does have 12.5% higher bandwidth than that card.
I still have it. At the time, it was $450, and it was hands-down the worst GPU in its price band. Of course, this is an era where a little shimmer at 1440p & 75 fps gets me very MATI, while it seems just yesterday I was happy to play Doom at 320x200 and 25 fps.
@Sexual Chocolate, unless you are getting a steal on it, like $350 or less, I'd avoid it. Raytracing performance is uselessly bad on that generation of cards, FSR is awful, and Adrenaline software is always half-broken. In the $400-ish range, other current-gen cards are better.
Also, @Sexual Chocolate, relevant to not buying an AMD card, DLSS is a lot better than FSR. The still images (DLSS left, FSR right) do a decent job of showing how much crisper DLSS images are, but what they don't show is all the ghosting and shimmering FSR is prone to. The consequence is that an NVIDIA card delivers image quality on par with AMD cards an entire tier up, i.e. I'd put a 3060 on par with a 6700 XT, despite the raw specs and synthetic benchmarks putting it equal with a 6600 XT. DLSS is just that big a deal. Reviewers like Tom's Hardware tend to pan DLSS, but the fact is, why would you not get yourself a 1.5x-2x FPS bump and run at higher settings? It just plain looks better, even if half the pixels on the screen are inferenced rather than rasterized.
However, NVIDIA's frame generation is no longer a big deal with Lossless Scaling, a simple $7 Steam app that adds frame generation to any game you want. In my experience, if your game is already running at 45-ish fps, the experience is near flawless.
Dear nerds, View attachment 6815124
I need to upgrade my GTX 1060 3GB, it can run RDR2 at mid/high settings surprisingly well but is pretty old now.
Is the 6750 XT a good replacement? Am hoping to get something that's a major upgrade, should last a few years, and cost less than a PS5 Pro. I'm just an occasional gamer so don't wanna spend too much. View attachment 6815060
PC specs are:
i7-8700
16GB RAM
Nvme SSD
Buy a 4070 non-super. It's cheaper than a PS5 Pro by $150 and runs circles around it in terms of graphical performance. Also you get proper DLSS and it only needs one 8-pin PCIe power connector.
The rest of your setup will probably hold it back a little but the 4070 is one of the most popular cards of this gen and will likely stay relevant until the next console generation.
If you're willing to shop secondhand, an RTX 3080 Ti can be found for like $100 less than a new 4070 and offers comparable performance while still giving you access to DLSS upscaling (no framegen but who cares lol)
Supposedly, FSR 4 will actually use inferencing. Tensor performance on the 7000 series is still poor, but at least on par with low-end 30 series GeForce. If FSR 4 is at least as good as XeSS in DP4 mode, it should significantly close the gap.
I’m predicting “DLSS 4” is going to be the same shit Nvidia pulled with frame gen except DLSS 4 will be what everyone associates with the rumored “neural rendering” feature that will likely be locked to the 50 series only.
I’m predicting “DLSS 4” is going to be the same shit Nvidia pulled with frame gen except DLSS 4 will be what everyone associates with the rumored “neural rendering” feature that will likely be locked to the 50 series only.
5090 Chad is coming, are you ready to be 4090 Manlet?
I agree. I don't want to speculate too much since we'll know everything in days. There may be clues in previous research material but the marketing department can call anything "neural rendering". It will not make a splash if it only works in 2 games at launch.
Yea I will be coping and seething until I buy the halo card for the 60 series. But realistically having 24 GB of VRAM still puts my 4090 in the chad tier. The 5080 will encounter issues with VRAM saturation at 4k.
I don't know how bad 16 GB could possibly be for 4K since for some games it clearly isn't. I would have to ponder graphz again. But I do expect a 5080 refresh with 24 GB using the 3 GB GDDR7 modules within a year.
I don't know how bad 16 GB could possibly be for 4K since for some games it clearly isn't. I would have to ponder graphz again. But I do expect a 5080 refresh with 24 GB using the 3 GB GDDR7 modules within a year.
The main issue with the vram at 4k comes down to upping the raytraced settings and also using frame gen. Both of those at 4k will eat a lot of vram on top of the high textures someone who is spending over $1,000 on a new GPU is going to want to use.
I honestly wouldn't be so sure that Nvidia will do a 24GB refresh of the 5080, mainly because it would cannibalize most of the potential buyers for the 5090. A 24GB 5080 would be priced similarly to a 4090 anyway most likely.
I honestly wouldn't be so sure that Nvidia will do a 24GB refresh of the 5080, mainly because it would cannibalize most of the potential buyers for the 5090. A 24GB 5080 would be priced similarly to a 4090 anyway most likely.
I don't think they'll ever have any trouble selling the 5090, which will attract attention for AI/compute. It has double the memory bus and cores than the 5080.
The 4090 clearly suppressed 4080 sales because buyers were disgusted by the 4080 being so expensive with similar price/perf to the 4090. I believe they are releasing the 5080 first this time to try to sell a few more.
Whatever mess Nvidia makes with the lineup this time, they have the opportunity to fix it up using the 3 GB GDDR7. 5060 12 GB, 5070 18 GB, etc. 5090 48 GB is the forbidden product.
* Doubles down on 7900 XT / 4080 raster, 4070 Ti (or better) raytracing for RX 9070 XT.
* Probably 16 GB for both 9070 XT and non-XT, so not 12 GB, 192-bit for the weaker card.
* 5080 will be expensive.
* Source claims H2 2025 launch for 5060 12 GB, and a 5060 8 GB would have less cores (like the 3060 8 GB).
* Battlemage = low supply, fake MSRP, B770 might not happen.
