I bought a 1700 in 2017 when they launched, upgraded to a 3800X 2 years ago and now I might even get to run a 5800X on my B350 board that I paid CA$130 for.
This is what good competition looks like. As a customer I love it!
This is what good competition looks like. As a customer I love it!
Competition is fantastic, but the people who rise to the competition are, in my opinion, the real heroes.
https://wccftech.com/asrock-amd-ryzen-5000-cpu-bios-support-...
Then AMD told them to stop it. Which is why that BIOS never got an official release and never appeared on ASRock's website.
There are people that have been running 5000 CPUs on ASRock for over a year now.
Conversely, you can run first gen Ryzen on B550/X570. Which many people seem to bizarrely think is impossible. Possibly due to AMD's marketing.
Or am I just thinking about an entirely different thing...?
It really does not fit the narrative of what AMD did with Zen so far, that's for sure.
Most of the stack has a similar TDP to Zen1/Zen+/Zen2. The 5950X has the highest TDP, 105W, which is the same as the 3950X, which is allowed. In practice the 5950X actually pulls a bit less power. Furthermore, the A320 boards (even lower-end) got official support, despite a lot of those boards having utter trash VRMs.
(furthermore, there's nothing that guarantees that every A520 board has a VRM that is capable of supporting a 5950X either! it's not like board generation is some guarantee of VRM quality, there are X370 boards that can handle a 5950X and A520 boards that can't. That's not how support is determined.)
16MB bios isn't a limitation either. B450 boards are the literal exact same silicon - B350 is to B450 as Z170 is to Z270, basically, it's rebranded silicon - and B450 boards with 16MB BIOS got support. There are B450 boards with 16MB bios that support literally the entire range of chips. And once again, so did A320 boards with 16MB BIOS.
Releasing official support for A320 really just blew a hole in every retconned justification that people tried to make for AMD. It's uncouth to say it here on HN, but there really is a large number of people who are super emotionally attached to the AMD brand and willing to stretch to absurd lengths to justify what are clearly business decisions.
And now they're just reversing their previous policy. Wasn't a problem at all, actually.
https://wccftech.com/amd-warns-motherboard-makers-offering-r...
The reality is that AMD fanboys haven't woken up to the fact that AMD is no different than Intel. All the same marketing and segmenting games exist on AMD. They were the underdog and then suddenly, they are on top and acting just like Intel.
From a technical point of view "Chipset" no longer exists. Every single chip going into Socket AM4 is a SoC able to work 100% standalone. Any differentiation and market segmentation you see around "chipsets", which are by the way just a fancy northbridges hanging off standard PCIE, is fake profit maximization theater.
For example, my B350 doesn't have a USB-C port, LAN is only 1Gbps, only one M.2 SSD slot.
- Caveats - These are software dev & light multimedia editing computers. By the time I get rid of one, it's had as much RAM added as possible. I've done this twice now. Would not work on a gaming rig.
Generally I'd buy midrange parts, from well known vendors, and pay a bit extra for SSD, ram, etc. Avoid anything with a small muffin fan (primarily chipsets these days), get a quality case, 140mm or bigger fans, quality power supply (a 5 year warranty is a good start). My current system is from 2015 and was a Xeon E3-1230 v5, 32GB ECC ram, GTX 1070, and a M.2 NVMe. The E3-1230 was cheaper with ECC then the similar i7 was without and just slightly slower (3.2 vs 3.4 GHz). Had no problems so far, and I'm just starting to look at a replacement. Looking closely at the expectations for Zen 4 CPUs and the current Alder Lake (or follow on ). I'm definitely sold on DDR5, especially since with a 2 dimm system (128 bits wide) you get 4 channels instead of 2 with DDR4. Should be that much better at not stalling your cores on cache misses.
If I bought today it would be tough decision between the AMD 5700x (8c/16T with a 65 watt TDP) vs the similar Intel Alder Lake. I do think AMD has the edge for systems that are low power and Intel the advantage for high power. Personally I'd rather have quiet, cool, and reliable instead of an extra bit of performance that nobody will care about in a few years. I'd consider a Apple M1 mini ... if it had the M1 pro or Max. Don't think the studio is quite worth it, but I am chipping in few $ a month to Marcan for a linux port.
Before that, MoBo manufacturers even made extensive use of additional chips for networking, USB, mora SATA ports, RAID and what not, which could make the whole system more complicated/fragile (all the additional drivers!).
I nowadays try to avoid that and aim for boards close to "reference" (whatever the chipset offers out of the box).
I just never bought a very expensive video card & I can't comment on how long a good one would last. I imagine it depends on to many variables from the games you play, the settings of the games you're okay with, the card, etc..
unfortunately now we get into the discussion about 3.5 slot GPUs that overhang all but 1 or 2 of your pcie slots, and the general lack of PCIe connectivity on consumer boards... you can sorta work around it sometimes with vertical mounts/etc. But it's annoying and takes work/planning.
I miss the days of X99 and having 28 lanes at my disposal, and every GPU being 2-slot or at most 2.5 slot such that I could actually use those lanes.
It's also been really annoying watching the arms race between GPUs and motherboard slot spacing. For a long time, GPUs were 2 and 2.5 slot, so motherboards went to 3-slot spacing for their slots. But then 3-slot and 3.5-slot GPUs became common, which overhangs the middle slot (or puts the GPU right up against a card placed in the middle slot). What we really need now is for motherboards to go back to 2-slot spacing so you get a slot in the #4 spot to give 3- and 3.5-slot cards some breathing room...
How far we've come. I remember having a semi-decent on board 100mbps LAN was only reserved for highest end.
Same for M.2.
These cpus are much more than 10x faster than the CPUs you would have used then.
Networking speeds in desktop devices haven't kept pace. Not just with cpu speeds but with storage speeds-- which is particularly obnoxious when you want to use network attached storage.
I assume this is due to a mixture of internet and wireless creating an extremely low bandwidth least common denominator and that running >1GB over copper is kinda problematic (finicky, power hungry, etc) -- and the industry seems to reason (perhaps correctly) the the customers in this segment can't handle fiber.
I personally have 40GBE attached desktops (and 100gbe attached stuff within the server rack)-- thankfully quite economically due to surplus enterprise hardware, so I'm well aware that it can be done... but to do it in a small formfactor system is still a problem, e.g. for any of these non-apu systems your normal mini itx systems will use all their slot space for a graphics card and not have room for a mellanox adapter.
Actually, it's not only about the CPU performance, because NIC performance drew a cosine wave w.r.t. processor power over the years. Earlier NICs had much more machinery inside them and were expensive. Then Realtek came and we witnessed the era of (almost) software LAN adapters, then the silicon became cheap, and some of the stuff moved back into the LAN controller. So modern, higher end cards do not do significantly more offloading, but they handle other things like virtualization of cards between many VMs, or other higher level tasks.
From what I've seen, making a faster CPU is easy, but a faster and wider fabric is harder to weave due to distance, switching related noise and other physics related phenomena. Cache and good memory controllers are also relatively easy, but expensive. Also, RAM and higher voltage doesn't go well together, because RAM uses a lot of energy for its size, so it comes with heat and stability problems and other lifetime related problems (yes, you can fry RAMs).
Storage is another hard technology. Spinning rust is limited by physics, vibration, rotation speed and other mechanical stuff. SSDs came crashing, but before they were cheap and Sun didn't got swallowed by Oracle, they did nice ZFS arrays with some very expensive SSD caches and spinning drives combined. I've seen whattheheckisthat amount of speed from an array with a size of 6-8U total and a couple of IB interfaces fitted to it at the factory. Currently we get that amount of speed from Lustre arrays with some SSDs and a lot of spinning disks.
With the current backbone capacity of a standard desktop computer, even on the high end, 40Gbps network is overkill unless you're going to ingest that data directly at the CPU. Yes, absorbing the data at the disks are possible, for what cost and what use case? 10Gbps is understandable, but I still consider 1Gbps as a sweet spot for general purpose computing. If you're going to transfer vast amounts of data from a storage array to a local system, yes you can go higher, but it's still a niche. Also as you've said, going over 1Gbps is problematic from a signal integrity point of view, and fiber is too fragile and expensive for average customer.
These Mellanox cards get quite hot when they're constantly utilized, so in a little ITX box, both the card and the system will be cooked. The process becomes faster if you use active fiber cables, because connectors sink the heat inside the case via heat sinks attached to the connectors on the cards. Even if you use them in a system room, building a large network with them requires expensive equipment which has adequate bandwidth and cable length. Also, running an IB network needs other software machinery to keep the network up. You can directly run ethernet, or TCP over IB, which kinda defeats the purpose and adds additional penalty and reduces the performance.
All in all, higher speed network is still doesn't bring too much of a value to the home of the average user, and is not still that cheap for the enthusiast. Also the hardware is not scaled down for the home. Cards are big & hot, connectors/cables are bulky and the fabric gear is noisy, bulky and high maintenance for home.
Yes they work rather nicely, but I'd not want them at home. I have enough of them at the system room already.
GigE is pretty annoying in that it can't keep up with a single spinning disk (let alone 5), so large files (family videos, movies, system images, backups, etc) can take annoying long. Hell even wifi access points are becoming bottle necked by GigE. Fortunately 2.5Gbe looks like a reasonable compromise. Rarely a bottleneck and support is common on cable modems (motorola among others), access points, motherboards (pretty much anything newish from AMD/Intel), and even small SBCs if you run one as a firewall, router, or pi-hole.
There are some switches and routers around and are often fanless till the port counts get high.
Even quality 10G cards seem to run pretty hot, have huge heat sinks, and the sfp+ I've worked with tend to get pretty hot. Seems like there is a newer gen 10G chip, but they are rare and expensive but do run substantially cooler.
Boards like the Intel quad port 10G X710-T4L (the cool one) vs the X710-T4 (the hot board) run at about half the power (14 vs 29 watts). I'm shopping around for the dual port version (X710-TSL) which peaks at 9.6 watts and uses 8.2 watts typically. No idea how other (non-intel) 10G boards do on power use.
And indeed, the heat from higher speed cards is a nuisance... though 10G copper is awful in that respect.
> too fragile and expensive for average customer
I think that's a bit overstated. It's more fragile but not enormously so. Terminations are a nuisance, which favors pre-made patch cables. I'd pretty much always prefer to use 10g over fiber than copper.
But I'll freely admit that having spent more than a decade as a network engineer my expectations may be weird. I will say that after many years of using fiber connected equipment at home for the desktop/server systems, my spouse has never broken a cable or connector (nor have I, of course).
I've never-- professionally nor personally-- had good luck with copper 10g and try hard to avoid it. If I have to use well made/tested pre-made patch cables to make it work I might as well use fiber (or dac cables) and trust that literally anything I buy will just work.
> Also, running an IB network needs other software machinery to keep the network up.
I used to use IB because the IB only nics were super cheap on ebay-- but I don't anymore. It's easy enough to use 10/40/100GBE. The extra software was a nuisance compared to the value it provided.
> and is not still that cheap for the enthusiast.
I mean, you can get 40gbe capable cards on ebay for like $50ish (usually generation or two old mellanox dual ib/eth cards)-- main cost to them beyond their power consumption is knowing how to use the utility to put them in ethernet mode. Dac cables and LR2 optics at similar prices. Particularly if most of your usage can be covered by dac cables, I think it's pretty affordable-- not 1gbe affordable, which is close to free on on a cost per port basis, but when you're bottlenecked on network bandwidth there isn't really an alternative, except for keeping data locally on the systems.
I think you're correct in that regard. I'm an HPC administrator for ~15 years, and I manage quite a few systems with a sizeable storage and related network machinery up to a certain point. While I can understand your point, I've written my comment from the eyes of a bona fide standard consumer which doesn't do anything higher end on their homes.
> 40GB is overkill, but 10GB still bottlenecks for me. I'm surprised to see anyone consider 1GB a sweet spot-- it can't keep up with a single conventional disk, and it's causes obvious slowdowns for fairly conventional operations.
My sweet spot comment assumes streaming content consumption, occasional system to system file transfers, in home entertainment streaming and some gaming maybe. If I'd have to move my photography or video archive to a NAS instead of internal drives, a fatter network would be indeed required. To be bottlenecked by 10gbps, you really need to have a busy network with a lot of going on. None of my personal networks carry that amount of data any given time, and they don't need to do it to be honest. Moving 10gbps fairly constantly is not fairly conventional, TBH.
> I think that's a bit overstated. It's more fragile but not enormously so. Terminations are a nuisance, which favors pre-made patch cables. I'd pretty much always prefer to use 10g over fiber than copper.
Yes, modern fibers can endure relatively unbelievable amount of abuse, but it's still well inside the boundaries of a normal user's limits, and the threshold is still lower compared to fiber. Many people I know can damage a fiber without intending or knowing it.
> I used to use IB because the IB only nics were super cheap on ebay-- but I don't anymore. It's easy enough to use 10/40/100GBE. The extra software was a nuisance compared to the value it provided.
When you're doing HPC, the value provided by the features are indispensable. Latency, MPI offloading, DMA are all essential things you need for a really performing cluster. So we prefer IB rather than Ethernet. So horses for courses, obviously. I'd also rather not use IB for a home network.
> I mean, you can get 40gbe capable cards on ebay for like $50ish...
Consider the price as the money + setup + maintenance. For you, the last two items have no cost virtually. For me they'd be negligible. For a regular consumer, they're much bigger items to wrestle with.
So, as a people who knows how to play with bigger tools is really a privilege and luxury. However, I have enough servers at work, and I don't want to manage stuff at home. So if that lower and, dumb stuff works as I want and doesn't make my life worse, I'm a happy camper. As a result, I over-engineer my home stuff very minimally.