AMD Launches 12 Desktop Renoir Ryzen 4000G Series APUs: But You Can’t Buy Them | Hacker News Reader

AMD Launches 12 Desktop Renoir Ryzen 4000G Series APUs: But You Can’t Buy Them

anandtech.com

140 pointsfrankjr138 comments

I just got a Lenovo V14 Ryzen 5 4500U laotop yesterday and this thing is amazing. It benchmarks almost as high as my Ryzen 5 1600 desktop PC, but it feels so much faster in actual everyday use.

To those who scoff at repurposed laptop parts being used in desktop systems, I don't think that's going to be an issue here because these 4000 series APUs have outstanding performance.

I'm not an Intel fan and have a tendency to root for the underdog so I may have an inherent bias here. The fact that Intel not only has competition, but has in alot of arenas been bested by AMD should only mean good things for the health of the overall market.

I've got an X13 AMD with the 4750U and 32GB of RAM and it just dramatically outperforms anything else I've used. It's crazy. I bought it because I wanted a step-up from Intel CPUs that was worthwhile from a 7th gen i5, and this is over 5 times as fast and feels like it.

I tried to customize an X13 on their website and it wouldn't give me an option for 32GB. Did you have to do something else to get that configuration?

You have to choose the Ryzen 4750U first.

I wonder why they added this limitation. Does somebody know if there another explanation other than Lenovo trying to maximize profit by forcing every customer who wants to go for 32GB of memory to get the most expensive CPU as well?

It appears that half of the memory is soldered to the board. That was also the case on my laptop mentioned above. It came with 4GB soldered, and 4GB in a slot. Why? I don't know. I was able to swap the 4GB with an 8GB and somehow it's still running in Dual Channel mode per CPU-Z so I guess it's no big deal, but I also find it odd.

The Lenovo Thinkpad X13 Gen 1 (AMD) has all memory soldered and no SODIMM slots according to its specification: https://psref.lenovo.com/syspool/Sys/PDF/ThinkPad/ThinkPad_X...

Probably the model built around the lower-specced CPU has a "lesser" motherboard + power supply, that might draw less power on idle (so, better battery life) and be smaller/ligher, but would render the motherboard unable to output the spike currents required for the higher-specced CPU, nor the sustained current required for DRAM refresh of that much memory.

The model with the better CPU necessarily has a more power-hungry motherboard feeding it, and so can also feed current to 32GB of DRAM; but this likely comes at the expense of on-idle battery life.

This causes me pain. I really, really want an AMD Thinkpad, but NONE of them has a 2k or 4k display. I tested the FHD ones and I just can't bare it/feel alright dropping 1.3k€ on FHD in 2020. No sure what to do now :(

I too want an AMD ThinkPad but none of the ones with the new Ryzens are 15". It sucks that they've decided to cripple their AMD line for no reason.

If you really want you can get the T14s and swap the panel, but it's expensive and a pain, especially on a brand new machine.

I keep an eye on that route. So far it hasn't been confirmed the AMD motherboards actually can connect to the high res displays.

I would love the X13 yoga 4k(?) oled version with AMD. That be my perfect machine as it has pen input too.

While I agree, one plus is that some of the FHD displays are now extremely low power, with about 1-2W lower idle power consumption.

But I'm not sure why Lenovo doesn't make the low power display the baseline like HP did with their Spectre line.

On this note, I got the 500nit e-privacy display and it's "fine". I don't love it, it's not quite as good as it sounds, but it's "fine". Viewing angle with the "privacy" part off is still pretty terrible.

A 15" 1080p panel has the same pixel density as a 32" 4K panel.

But you sit closer to it so it needs to be sharper. The issue with 1080p at 14 inch is that 150% scaling has blurriness issues with bitmap graphics and 100% scaling is a bit too small and 200% scaling is a little bit too big for most people. Linux is also terrible at 150% scaling across the board. 1440p is the perfect trade-off at 200% between crispness and battery life and well-supported in linux, and while 4k at 300% is very crisp it is a massive power drain.

Apple and microsoft consistently ship displays between 200 and 250 dpi, the sweet spot for 200% scaling. It is frustrating that nobody else does this. I’m on the fence between replacing my 1080p thinkpad with another thinkpad or a macbook pro, and this is a major advantage for the mbp.

Windows has 125% scaling. Linux vendors should look into adding that. :)

KDE does scaling at any 6.25% interval (for technical reasons), which hits 125%, 150%, 175%, 200% and so on.

The IdeaPad S540-13ARE is spec'ed with a 2560x1600 screen: https://psref.lenovo.com/Product/IdeaPad/IdeaPad_S54013ARE

But seems very hard to find a place that actually sells it around here (Norway).

Yeah, but IdeaPad is not ThinkPad... I would only do education offers and the non ThinkPads don't come with 3 years of onsite premium support. The IdeaPad does come edu discounted tho.

Do you know what's expandable in this machine? Below I see that some RAM is soldered onto the motherboard, so is there 1 DIMM slot available? What about the PCIe SSD - is that replaceable too?

In the X13 only the SSD is replaceable. I _think_ it has a slot for cellular too, but I'm unclear if you can use it for anything. Lenovo says that if you want cellular you have to specify at build because of antennas in the display.

The T14S has soldered memory and I think even the T14 does too now.

Thanks! I just looked up the T14 on the Crucial site [1] and it looks like there's a single DIMM on that machine as well as replaceable SSD. So nice to have options ...

[1] https://www.crucial.com/compatible-upgrade-for/lenovo/thinkp...

I remember having a similar impression with the introduction of past CPUs like Pentium-M and Core 2, and even recently with a Ryzen 7 Pro 3700U.

A new laptop platform suddenly seems faster than a desktop of the prior generation (in practice, often 3+ years old at the point of comparison). This was true even compared to my Xeon workstations of those eras. I can remember saying to myself, why would I get another desktop when this little thing is so fast and low power? Eventually, I rediscover the ways in which the laptop is confining, and revert to the mindset that desktops have a place too. Whether it is upgrading storage, RAM, GPUs, or other accessories, there is a flexibility to the desktop chassis.

Depending on whether it is an employer budget or my personal funds, I find utility in workstation-class or commodity uATX desktops. I can reliably find a sweet-spot configuration to get a fully new platform on an appropriate interval and then trickle in small upgrades as needed during its service lifetime, due to the ease of replacing parts piecemeal.

I also tend to keep the same monitors for a long duration and just procure new, headless PCs for office upgrades. I only upgraded monitors for format changes like 1600x1200 CRT, 1600x1200 LCD, 1920x1200 LCD, and 4K LCD. I suppose you can do the same even with docked laptops, but I find you are paying quite a premium for the built-in display even if you only wanted to upgrade the CPU...

APU's are great for the mini pc market. I don't need a laptop (currently) but I love using a SFF machine without much fan noise. These parts tend to get packaged into these machine. Lenovo M75q Tiny etc.

Based on these parts speed I may move to a laptop as my "main" machine in the future. Just dock it at home, office and use on road.

I've been waiting for this generation of Ryzen APUs to build a Small Form Factor PC for under my TV.

Same. I have a 2015-era home NAS that could really use new guts, but I'm waiting for a 12W-or-lower Ryzen CPU and matching Mini ITX motherboard to become available first.

I recently built a Celeron J4105 (10W TDP) system, but I would have loved to use an AMD board with similar power consumption.

They also work well as home servers and stuff!

I too just got a Renoir laptop (the 8C16T Ryzen 7 4800H) and its performance is truly impressive. It scores SC:1254 and MC:8269 on Geekbench 5 [1], within spitting distance of my desktop Ryzen 7 3700X workstation performance of SC: 1319 and MC: 9296 [2].

[1] https://browser.geekbench.com/v5/cpu/3002232

[2] https://browser.geekbench.com/v5/cpu/3007038

These are not repurposed laptop parts being used in desktop systems; these are desktop parts because they fit in an AM4 socket, laptop parts don't. The CPU parts (CCX, IO) are the same, but that is the commonality with laptops, Threadrippers and Epic server parts too.

I have a little 1.92 liter DeskMini A300W with a 2400 G APU and it works well as my desktop computer and also connects to my TV a few feet away. It actually sits on the little shelf under the desk. The only graphics game I have played recently is Rocket League and it handled that very well. Also was fine for GTA V.

I have been thinking about upgrading to 4000 when it becomes available but right now I actually don't really need more power.

I wonder what PC Engines will do to replace AMD's Jaguar APU/SoC. All of AMD's new SoCs seem to be too hot for passive cooling, at least without huge fins.

I can't find the link, but there were recent discussions (by Red Hat? GCC? clang?) about changing the default minimum supported ISA. There were several proposals, IIRC, with Jaguar right in the middle of the proffered choices, and possibly the best choice for now. But I imagine if that discussion were to happen 2 or 3 years down the road, even Jaguar might not make the cut.

At the low end of TDP, Intel still reigns supreme. But their Atom lineup makes far too many compromises, excluding various ISA extensions to price differentiate. And Soekris got burned with their net6501 and defective Atom chips, so I imagine Intel isn't that great a partner for small companies. (I doubt AMD is, either, given the firmware pain people have, but at least their products work.)

They'll use the R1102G which has the same power envelope as the T40E released 8 years ago?

https://en.wikipedia.org/wiki/List_of_AMD_accelerated_proces...

Ah, released February 2020. I missed that. Cool.

The last word on the matter from PC Engines was here, http://www.pcengines.info/forums/?page=post&id=7D919FB8-C4BB..., where it seemed even 12W might be too much, given the power draw of the other components.

Would they move from four physical cores (GX-412TC) to two and without even hyperthreading (R1102G)? I'm not convinced.

I've seen a few thin client manufacturers that moved to Intel Goldmont APUs such as 4c ~10W J4105 (Wyse 5040, Fujitsu s740). Interestingly HP continues to use a custom downclocked Carrizo in their t630 (GX-420GI, 4c, 16W).

My guess is PCEngines is gonna wait and see, as the four core jaguar is enough for what they need for now, and they care about the longevity of their product life cycles.

Two cores that are twice as fast would be preferable to four slower cores. Since PCEngines wouldn't be powering up the GPU side of the chip, they could probably make pretty good use of the new CPU's boost clock capabilities. That plus the microarchitecture advantages of Zen over Jaguar, and the higher DRAM bandwidth should make the two-core Zen parts competitive. But I could also see them waiting for 7nm embedded parts.

There is a delightful market of very small fanless boxes on AliExpress that typically just throws a 15W laptop CPU into an aluminum case. Turns out if you slap a couple pounds of aluminum onto the CPU, it will stay pretty cool. I have a small box with an i5-5200U (15W part) that includes two NICs. This runs as my router and firewal at home. It sits around 40C. It might throttle if I ran heavy workloads on it, but it's great as a small always-on PC.

Here's a search to get you started, and you can refine from there: https://www.aliexpress.com/wholesale?catId=0&initiative_id=S...

I wouldn't delve into AliExpress when there is https://pcengines.ch, all boards made in Taiwan and engineered in Switzerland.

But they don't seem to offer boards based on AMD's R1606G which is super power-efficient. Nor do they offer dual 10GbE NIC boards. And they don't have good ready-made cases with good cooling. :(

That's the competition that ultimately ate Soekris' market--they never could recover after the net6501 debacle, and maybe would have exited regardless.

But what Soekris offered and PC Engines still offers is ostensibly greater consistency and reliability, a better firmware story, and a proven commitment to long term availability and support of the platform. PC Engines still sells their ALIX model, and Soekris Europe still has net5501 boards in stock. The ALIX and net5501 are both ~15 years old, I think.

A substantial portion, if not most of PC Engines' business, comes from small vendors building custom embedded solutions, such as WiFi base stations. The developer/hobbyist community loves PC Engines because it takes the guesswork out of things. The hardware will always be well supported by FOSS--because of PC Engines' efforts, and because there are plenty of FreeBSD, Linux, OpenBSD, etc developers with a box--and there's solid brand recognition regarding quality. (Also, serial!) The downside is that the specs are antiquated almost from day 1, but for basic chores still adequate.

I hear you on the reliability of a known and long-tenured vendor.

I got one of the AliExpress ones, because I wasn't originally certain what it would be used for. At one point it was my primary PC, and it performed silently and as well as or better than a same-specced laptop.

The support is nonexistent, but compatibility is pretty good. These mini PCs are just taking generic laptop motherboard and CPU assemblies (probably the same you'd find in a Clevo or similar ODM machine). I wouldn't use one for something mission critical, but it worked great as a silent desktop and it has been a (super overpowered) router box as well.

Do you use pfSense as a router, or something else?

I am on the fence of not buying Mikrotik or Ubiquiti and just get one of these boxes you mentioned. But I am not a network admin and have no clue about network protocols (yet).

My main idea is to return to Android but have very strict blacklists on my router machine so as to preserve privacy the best I can. And while Mikrotik routers can import various blacklists -- and even auto-update them! -- I am still not 100% sure about it and I wonder if I should take the leap to a small pfSense box that I control completely.

Do you have any opinion on this?

I run the latest OpenBSD stable and several utilities for myself.

The router and firewall is composed of just standard PF and a config I've built myself based on the networking FAQ.

I also have a bastion host for my home LAN, local DNS caching, a syncthing server, and some useful cron jobs. My goal is to also use it for a home automation hub, but all of those projects are in the future.

Thank you. Do you have observations on how easy it is to add a WiFi access point to a pfSense box? I've read some obscure posts on Reddit claiming that it's a hassle and that's the main that scares me.

I've never used pfSense, so couldn't tell you.

If you are referring to adding a wireless NIC to the pfSense machine and having that box be a combination router/firewall/AP, the hardware side is easy. Wireless NICs are trivial to install - just plug in and run an antenna somewhere useful.

From a software side, it's really easy to configure an AP in OpenBSD - just a quick ifconfig. I have no clue if pfSense supports this sort of config through their management interface. I would assume yes, because it seems basic.

If, on the other hand, you are talking about having a discrete AP as a different piece of hardware on your network, this is also straightforward. I have both a DD-WRT flashed Trendnet router/AP acting as a non-routing AP and some Ubiquiti APs on my network.

Ubiquiti stuff is trivial - just plug it in and configure through their management interface.

The DD-WRT was a bit more work, because it wants to be a router by default. It's easy enough to turn off its router and just run as an ethernet switch + wireless bridge. I had to do some searching for a guide that indicated the right settings, but it was easy enough once I knew what needed to change.

All of my home networking setup was my education in networking. I did this with no significant experience prior, except some high level theoretical knowledge, and having messed around in a consumer router/AP's settings before.

Honestly, I'd say just give it a try. Don't be afraid to do some searching and some trial and error. And if it doesn't work out, return the stuff and buy the full gamut from a vendor that sells networking gear with a good interface. I like Ubiquiti equipment from my personal experience with it, but I haven't used other vendors to give you any comparison.

Recent Atoms (Goldmont) are pretty decent - out of order, x64, three instructions per clock (compared to Jaguar's two), SSE4.2 and seemingly free from the bus degradation flaw. You lose AVX, but get to keep AES-NI and gain accelerated SHA which is nice on a tiny server.

I think firmware on the intel's side might be a limitation, as IIRC it now requires UEFI, while PCEngines relies on their legacy coreboot implementation (which is actually open so yay).

AMD sells a 6 W Ryzen part that come with 2 10 GbE ports and a GPU. I believe these are 14 nm parts...likely the 7 nm parts will have 50-100% better perf/W.

https://www.anandtech.com/show/15554/amd-launches-ultralowpo...

I have searched high and low everywhere on the most popular outlets that I could and I never found one of these with the dual 10GbE NICs. If you go to AMD's website that advertises machines made with these boards, _all_ vendors backtracked on dual 10GbE and reverted to dual 1GbE for reasons that none of them explain.

If you can point me at a store that actually sells a mini PC, say based on R1606G, _and_ it having dual 10GbE NICs then I'd be ecstatic. :)

EPYC chips also come with integrated ethernet, but I don't think any vendors make use of that, either, not even the EPYC Embedded (3000-series). My Supermicro M11SDV-8CT-LN4F with an EPYC 3201 has 4x1Gbps Intel ports (+1x1Gbps Realtek for IPMI) even though the chip provides 4x10Gbps ports. I've read the issue is that motherboard designers can't be bothered to redesign around the integrated controller; easier to just continue slapping Intel or Realtek controllers on the board.

That seems to be a big hurdle for AMD--they don't have enough volume to motivate much of the industry to make the best of what AMD offers. And when it does happen, there's a much longer lag time.

Pretty good explanation. Thank you.

It's a real shame. The R1606G can go at 25-35W and it would practically be the most economic I/O heavy server design in history but as you said, the vendors can't be bothered. :(

Making use of the NIC/MACs in the CPU would require board makers to buy PHYs, which would be only used by AMD boards, or they could not do that and slap the same PCIe NICs on there as they do on all other boards.

They said they're going to be doing something for retail consumers soon.

Their standard desktop CPUs are flying off the shelves.

I imagine there's only so much capacity for chip building.

The main reason is simply market share - the DYI APU market is very small compared to the system builder market.

So if 80% of their APUs end up in prebuilt systems anyway, it makes sense for them to focus on that market.

These are down-binned laptop parts, and the OEM's are also laptop manufacturers, so this strategy gives both AMD and the OEM's more flexibility in selecting product mix.

Nothing that I do requires a discrete GPU. I want an integrated GPU that can drive a 4k screen without input lag at 60fps.

It kills me that there are no good AMD offerings.

So the 4000G series should be great for you, right?

Yup. Kind of bummed to see that it is mostly for OEMs though.

I feel like they'll wait until Ryzen 4th gen comes out since that will be on new silicon (tho still 7nm), leaving a bit of room for extra production of their APUs for DIY since this is still Zen 2 chips just with the 4000 naming.

Not sure from where, over here the shelves are full with laptops.

But feels like this is a short-sighted move. DIY market might statistically only account for 20% of the sales of these APUs, but that's historical data.

These new Ryzen 4000 generation should be a significant improvement over the previous generation, and better than what Intel can put out (especially on the graphics front). Just look at the rave reviews regarding the Ryzen 4000 laptops.

The DIY market is where a lot of innovation and word-of-mouth brand awareness is built up. Granted it'll be released to DIY market at some point, but it seem like a delayed opportunity. If laptop market is again of any indication, OEMs can give these the "budget unit" treatment and refrain from putting them in their flagship products.

Disclaimer: I've been anticipating this release, especially the 4700G chip, so this is very underwhelming.

Some will be available for individual purchases in some markets[0] if you absolutely must have them:

Launches 8/8 11:00(JST)

Ryzen 7 PRO 4750G ¥39,980($413 w/tax)

Ryzen 5 PRO 4650G ¥26,980($279, same)

Ryzen 3 PRO 4350G ¥19,980($207, same)

JPY prices are MSRP before tax(10%) [0]: https://www.gdm.or.jp/pressrelease/2020/0721/356407

> The AMD Accelerated Processing Unit (APU), formerly known as Fusion, is the marketing term for a series of 64-bit microprocessors from Advanced Micro Devices (AMD), designed to act as a central processing unit (CPU) and graphics processing unit (GPU) on a single die

https://en.wikipedia.org/wiki/AMD_Accelerated_Processing_Uni...

EU ( https://www.centralpoint.nl/ prices )

- AMD Ryzen 3 PRO 4350G - 4.0GHz (Tray) - 151,- EUR

- AMD Ryzen 3 PRO 4350G - 4.0GHz (MPK) - 194,- EUR

- AMD Ryzen 5 PRO 4650G - 4.2GHz (Tray) - 214,- EUR

- AMD Ryzen 5 PRO 4650G - 4.2GHz (MPK) - 276,- EUR

- AMD Ryzen 7 PRO 4750G - 4.4GHz (Tray) - 313,- EUR [1]

- AMD Ryzen 7 PRO 4750G - 4.4GHz (MPK) - 405,- EUR [1]

-----------

[1]

https://www.centralpoint.nl/zoeken/amd%2520ryzen%25207%2520p...

It's a bit funny that one of the problems these APUs have is that the gpu power is linked to the cpu power. The 4700G has more Vega CUs than the 4300G, which really is a pity for gamers that are interested in an APU to get a cheap gaming capable system, and unnecessary for those that want a strong processor with just some integrated graphics and would be just fine with the weaker graphics module. It would be nice if you could separate the processor and the graphics part and just mix them as needed... ;)

Anyone know a good workstation story for AMD? I'd really like to have a development workstation with an EPYC CPU (like the 7302P) and a motherboard that actually supports ECC, DDR4-3200, and PCIe 4.0. All I can find are not-yet-launched products like the Supermicro H12 mobo. Whitebox system integrators are shipping older Zen-compatible mobos that don't have PCIe 4 and 3200MHz memory, and there don't seem to be any OEMs in the game (Dell, HP, Lenovo are all Intel-only).

System 76?

From a few days ago: https://www.anandtech.com/show/15910/amd-announces-ryzen-thr...

tl;dr: Lenovo ThinkStation P620 coming in September

If you've got the setup for rackmount, the Dell R7515 has PCIe4, but not 3200MHz memory support I think.

Supermicro's H12 range was launched a year ago.

It was "launched" into the "Coming Soon" channel. That's why all the usual supermicro integrators are still shipping H11 boards with DDR4-2933 and PCIe 3.

No, our company's first H12 server batch with DDR4-3200 and second-generation EPYCs arrived last fall.

Do you need RDIMM/LRDIMM or would ECC UDIMM suffice? If the latter, have you considered a Ryzen 3950X or a Threadripper 3960X? There are plenty of X570/TRX40 motherboards that properly support ECC UDIMM (but there are also some that don't).

If you need RDIMM/LRDIMM the only current AMD alternative to EPYC is the recently announced Threadripper Pro. Launch partner is Lenovo with the ThinkStation P620 where you can get 12-64 cores with up to 1 TB of RAM. Expected to be available in September.

I don't need a ton of memory so whatever DIMMs are electrically compatible with the CPU will do the job (I imagine that 8 DIMMs of 8GB each is common?). I want it as a development platform for this PCIe 4 peripheral, so that's a hard requirement, and I would like if it resembles a production server as much as possible, so 8 channels at 3200MHz is desirable.

PCIe 4 and 3200MHz ECC are readily available on the Ryzen and Threadripper platforms. But if 8 memory channels is a hard requirement, then AMD's answer for you is the above mentioned Threadripper Pro in the ThinkStation P620.

If you can live with 4 memory channels, consumer Threadripper opens up your options enormously.

TRX40 with a Threadripper 3960X supports PCIe 4.0 (64 CPU lanes) and 8x ECC UDIMM. This would be a quad channel configuration (2 DIMMs per channel), not 8 channels and DDR4-3200 ECC UDIMM may be hard to find at retail. Threadripper is based on the same silicon as EPYC, though, so it may be close enough for your purpose. ASRock TRX40 Taichi, ASUS Prime TRX40-Pro or maybe ASRock Rack TRX40D8 could be a suitable motherboard.

Threadripper Pro will be even closer to EPYC with 8 memory channels, 128 PCIe 4.0 lanes and RDIMM/LRDIMM support. However, you'd have to wait until September and the Lenovo ThinkStation P620 may be the only choice at time of launch.

I've been using a previous gen system for almost 3 years: EPYC 7401P with a Gigabyte MZ31-AR0, 8x16GB of ECC, and a bunch of PCIe flash and some SATA rust. I love it; it's my favorite workstation in a long time.

For my new machine next year. I'm currently leaning towards ASRock Rack's ROMED8-2T, which is available for order but is not "in stock" at most places (e.g. provantage). It's PCIe 4.0, runs DDR4 at 3200 MHz, and has a reasonably long RAM QVL. I'll go with whatever the best value is for a CPU when I pull the trigger on the upgrade.

One of the 3 computers I work on a daily basis is a Threadripper. It never crossed my mind to look for an Epic CPU just because I have no special needs not met by the TR. For the build I just ordered parts from Amazon and assembled instead of getting a prebuilt configuration.

AMD states that the initial release of Ryzen 4000G hardware is for OEMs like Dell and HP only for their pre-built systems.

They did something similar with the 4xxxHS series for laptops - Asus has exclusivity over those for the time being.

Anyone have an idea if these could possibly have enough power for a media PC? I'm looking to handle at least 2 1080p transcodes via Jellyfin/Plex. My current plan was to get an old PC and toss an old graphics card in it, but these could be a great alternative.

Any recent intel cpu 8th gen and you with quicksync is going to be your best bet. An i5-8600 can do 15-20 transcodes on its own using the quick sync part of the CPU.

Plex has puzzled me for a while - why such limited hardware support for HW transcoding? Is this some sort of paid partnership they have with Intel/nVidia?

I don’t believe so. It’s just that Intel and Nvidia have put HW accelerated encoding into their consumer products.

You’re free to use the CPU to transcode but you may only get a couple of streams. The quicksync on chip can do 15+ on the lower end but sacrifices encoding quality for speed. Though recent quick sync and nvenc quality has significantly improved.

Edit: Intel has added the dedicated encoder chip inside most consumer (almost no xeons) where as far as I know AMD is not doing the same thing (not counting GPU encoding).

AMD APUs/GPUs have HW accelerated encoding/decoding as well, at least in Jellyfin. I'm not sure what you're getting at here.

AMD's video encoding HW is comparatively poor in performance/quality. I.e., if you're bandwidth-constrained (wireless or internet uplink from home), you're likely better off with x264 or maybe SVT-AV1 on a CPU instead.

Nvidia's H.265 encoder in Turing (the 2000-series of GPUs, many RTX branded) delivers superb quality for >=1080p30 real-time encoding compared to the alternatives (x264, QuickSync, AMD GPU) (x265 isn't really suitable for realtime-encoding).

The only reason I haven't yet planned a concrete AMD workstation is the lack of proper rr[0] support[1] in their CPUs, and for the GPU the rather poor hardware encoding[2] that makes it infeasible to seriously stream out 2160p60 without problematic visual artifacts.

[0]: https://rr-project.org/

[1]: https://github.com/mozilla/rr/issues/2034#issuecomment-58361...

[2]: I have not found serious competitive benchmarks for the recent AMD hardware encoders. I would appreciate such.

Can you point to some independent comparisons that provide support for the "superb quality" statement for RTX GPUs vs x264?

The reason I'm asking is that's not what I expect from my experience. The main issue here is the CPU necessary to get equivalent quality x264 encoding? I would expect x264 medium profile to be better quality than any (consumer) hardware solution and I can do x264 1080@30 medium encoding in real time using ~70% of my old 5820k CPU. A modern Ryzen (3600 or higher) would use even less CPU so really not an issue, for a single stream (but it seems Nvidia has artificial limits to the number of concurrent streams anyway).

Oh no. The RTX GPU would encode to H.265, I'm sorry for the bad wording. If you stick to H.264, x264 is gonna win against most stuff, except for maybe some exotic high-end broadcasting hardware/FPGA stuff.

I'm a bit confused by what you're saying. Are you saying that it's worth buying an Nvidia GPU because of the h265 performance? h264 is what most streaming devices have the capability to hardware decode, and generally is the default target of transcoding. It's pretty rare to have a device that would force your files to be encoded in h265 as opposed to h264. For me the most common use-case for transcoding is decoding an h265 file and re-encoding it as h264. For my devices that can stream either format, I usually end up in direct play mode rather than doing any sort of video transcoding

I was referring consumer AMD CPU’s which don’t have the comparable video encoder/decoder core like what Intel quicksync offers. Building a box can be tough because almost all Xeons don’t have quicksync, so trying to get ECC and quicksync is a pain.

If it's a desktop CPU then there will be a graphics card, which will have the hardware, unless you run truly headless.

If it's a desktop APU then it has the hardware.

I imagine Plex uses FFMpeg, Gsteamer, or some other video library underneath. If I remember correctly they started as a fork from XBMC (now Kodi).

For nVidia that’s an artificial limitation in their GPU drivers. They’re basically saying “if you want to encode more than 2 video streams at the same time, you must pay us much more for the same hardware in a different package”.

As far as I’m aware, neither AMD nor Intel have such limits in their hardware encoders or GPU drivers. The only limit is performance of their hardware but 1080p doesn’t have too many pixels, pretty sure they both can handle at least a dozen of streams.

It’s possible, but you’d be better off spending the $50 on a true GPU in almost all situations and using it in lieu of the onboard graphics. Just from a lifespan perspective on the CPU, better to offload that work and not pin the CPU for hours while you’re watching movies :)

They have full hardware transcoding support, it wouldn't even pin the CPU to do 2 simultaneous 4k streams, let alone 1080p

You're probably right. I could pick up an old PC on ebay and just toss a cheapo card in and that would probably cover my use case just fine. Thanks!

The CPU lifespan should not be depleted by using the built-in GPU or running 100% for extended periods.

For modern electronics it technically will be, but the only relevant lifespan is related to the electromigration of the bond wires. That occurs over ~100y of on time.

Keeping it hot basically won't matter unless it already had defects.

Nobody i have ever met has ever hit the lifespan limits of their CPU before that CPU was hopelessly outdated. CPUs just do not break, motherboard / RAM / powersupply etc. Will always die before

These APUs have 6/8 Vega CUs onboard, so they're going to be way more capable (AFAICT) than comparable Intel integrated GPUs, and in fact are considerably more capable than a GT730, which is about all you're going to get for that $50.

AFAIK Jellyfin fully supports both AMD AMF and VA-API for HW accelerated transcoding, so you could use either Windows or Linux and have full support for the HW encoding/decoding chips built into these.

I can't imagine you'd have any issue. My jellyfin box has an fx6300 with an even older quadro gpu (600 I think), and it doesn't seem to have any issue running multiple streams at once. I'm honestly not even sure if it's using hardware encoding on the gpu, or just cpu, but in Ubuntu 18.06 it worked out of the box with the open source drivers so I haven't messed with it.

Eagerly awaiting a mini PC with these chips. Hope DELL/HP come up with an option soon.

I'm waiting for the ASUS PN50 line up to be available worldwide, even though they use AMD 4000 laptop APUs.

I agree - the GE-parts look very promising and since these are desktop CPUs, we could finally see a NUC-like solution with a reasonable upgrade path.

Intel's Compute Elements-idea [1] doesn't really resonate with me as upgrades seem to be very expensive (judging from the non-existent pricing information).

[1] https://www.intel.com/content/www/us/en/products/boards-kits...

I jus got a second hand 6 core (6T) HP elitedesk mini for cheaper than I'd like to admit. Something similar with these APUs is enough to make me drool!

You can already build a computer that is basically the size of a miniITX motherboard by buying a $50 case, an external AC/DC power supply of 12v with 10-20 amps and a low profile noctua heat pipe cooler.

They mean you can't buy without a computer.

Actually from the article "Just to be clear, AMD specified OEM and not system integrators (SIs). On our call, AMD clarified that the market for its APUs is skewed very heavily towards the big mass-market prebuilt customers like HP and Dell, rather than custom home builds. The numbers quoted were around 80% of all APU sales end up in these systems, and by working with OEMs only, AMD can also help manage stock levels of the Renoir silicon coming out of the fabs between desktops and notebooks."

I could be way off, but I imagine OEM's buy a lot more parts in general than all system integrators combined. If this is the case, it could certainly explain the 80% number but still wouldn't really allow you to draw the conclusion that you should work with OEM's only.

They specifically refer to APUs which IIRC means they are referencing only models that include GPU capability.

Perhaps OEMs tend to favour CPU-only Ryzen models paired with other GPUs, except at the low/budget end of the market which I assume the 4000 series is not aimed at, in which case the 80% value is perfectly plausible.

Working only with OEMs makes the company vulnerable to fewer clients.

You get the feeling that OEMs aren't really playing fair with AMD. Lenovo is practically hiding their AMD options, pricing them too high without the usual discounts that are always there for the Intel models, then there's inadequate base configs (128gb ssd really for a 1500$ laptop?!), Crappy display options etc.

Just read Asus shutting off the vents on AMD laptop yet identical Intel one doesn't have that problem somehow.

I bought an XPS 13 5 or 6 years ago. It is working perfectly after all these years of intense usage. (I wish actually it broke so I had an excuse to buy the newer version.)

Unfortunately, XPS only carry intel. What would be a valid alternative to XPS with a Ryzen CPU?

I think the thinkpad series may have something with a Ryzen CPU. Maybe just hold off for a while?

ThinkPad X13 AMD

I just looked to price one on their website, and it starts with some absurdly low specs including a 128GB SSD and a horrible 1366x768 TN screen for $1140. This is not acceptable in 2020.

Indeed. For that price I still choose XPS

Thats downright disrespectfull. Some poor sucker bought that!

I don't know if AMD intended to mock Intel when they named new Athlon Gold and Silver which are on lower end, meanwhile Intel named Xeon Gold & Silver on the higher end.

I think it's competing with the Pentium Gold and Silver branding which are more or less in the same budget market.

Yeah, I forgot Intel also has Pentium Gold & Silver, IIRC Intel just rebranded Celeron?

Celeron still exists somehow. It "reigns" in the ~200 integrated pc area and cheap laptops.

Yep - my little two bay, pre-built NAS has a Celeron in it. I suspect this is the market they're targeting.

I don't think so, low end Intel above Atom but before i3 is still named Celeron and Pentium, and the Pentium Gold is a well known "brand" in that area.

Maybe and ironic given the history. Intel got upset about clone x86 CPU's using comparable numbering systems. Reason being 286,386,486 are just numbers and they couldn't trademark numbers. Hence the brand name pentium was born as a way to stop the alternative x86 cpu's capitalising upon what Intel deemed a brand and at the time, it was.

Gold and Silver are a bit like numbers, can't be trademark protected and with that, history repeats in many ways and with all that in mind, this is funny for me.

Since the GPU uses the system memory, is there any advantage to using these APUs for machine learning?

GPU RAM is typically under 32GB (more commonly under 11GB) and quite expensive - for the price of a V100 you could buy 1TB of system DDR4 RAM.

I'm guessing there are disadvantages in memory bandwidth, number of GPU cores, overall FLOPS, but was curious if anyone knew how these pros/cons balance out.

The big issue I'd expect you to hit is that you're doing ML without CUDA, and there's just inherent problems to working in an environment where your hardware isn't the first-class citizen.

ROCm/HIP/HCC is a broken ecosystem as of today. The 5700XT launched last year and still doesn't have support. I have some faith that they're going to get it straightened out in the next year or so, but today you're just asking for trouble if you're trying to use pytorch or tensorflow without being on team green.

These machines are limited to 64GB RAM AFAIK, or maybe 32GB.

The memory controller should support 4 DIMMs of 32 GB each, but in practice the motherboard implementation may limit to 4 x 16GB. How much can be allocated to the GPU part, no idea.

I was an ML engineer in a previous life.

For most ML and GPGPU workloads the bottleneck is memory bandwidth. It’s normal to hit 2TB/s because of all the caches, memory banks and the wide memory bus. I wouldn’t expect the APU to perform nearly as well as even an entry level GPU (say 1660) but I would be curious to see.

I’m very much interested in building a low power home server with one of these chips, but I’m wondering how I support a large disk array?

I’m not interested in an existing NAS setup. I’d like to do it myself.

What would you do normally? How do these change anything?

To support a large disk array you just need a few things. - Lots of drive connectors. - That can be an 8 port SATA motherboard. - A SATA HBA card. I've seen some with up to 24 ports, or some of these support port expanders which you would generally plug one cable into a hot-swap box of 2 to 4 drives. - A SAS HBA card. More expensive but you can use SAS drives with enterprise features. If you're okay with used drives you can sometimes find deals on 10 drives at a time at 25% of the new drive cost. - Lots of power connectors and a good PSU. If you have 8 SATA drives you need 8 SATA drive power connectors. Or you use hot-swap boxes and plug each one into one power connector, but watch the totals and remember drive spin-up can use 4-5 times as much power than normal. Numbers I've seen are 8W per active drive and 35W (maximum) for drive startup.

But at any rate for large disk arrays the CPU can be ignored because almost anything will work. Hard disks are slow enough that you need like, 20 of them before the CPU speed starts to matter. With SSDs you can max out the controllers and PCIe lanes and even RAM bandwidth all too easily.

I recently built a Celeron J4105 system for a NAS/server/HTPC, in a compact Fractal Node 304 case.

The ASRock J4105-ITX has four SATA 3 ports and a PCIe slot, to which I added a 2-port SATA 3.0 controller. Total six disks, which fits in the very compact case.

I run a 120GB SSD for the system, and a random assortment of 2-3TB disks for storage, in a Btrfs RAID1 pool.

All in, about 2/3rds the price I would have paid for a 4-disk Synology NAS.

You can get micro-ATX boards with 6-10 SATA 3 ports, and PCIe board with basically as many SATA ports as you're willing to pay for.

There is nothing different versus a regular Ryzen 3x00 CPU. If you want many drives and good performance you need to use add-in SATA or SAS controllers instead of the on-board SATA controller that is fairly slow; it is good for one disk at a time, but slow for multiple disk access like a raid-z3 array w. ZFS. In this case the limit is what controller you use (one can find sh LSI SAS controllers at good price and good compatibility; see the Freenas.org forums for more info) and how many.

For the "how many" part the number of available PCIe lanes depends mostly on the chipset, the CPUs have 24 in total - 4 to the chipset, 20 available to use. This is where the difference between B450 and X570, for example, can make a difference. The graphics part of the APU is not relevant for a NAS, but it can be if you also do transcoding or any GPU assisted work.

The standard controllers included on my ASUS Pro x370 board show you don't need a add-in controller. It's got 8 ports on the board, and when I do btrfs scrubs the 6 HDDs get 760 MiB/s and the 2 SATA SSDs get 1.05 GiB/s.

The scrubs start at the same time each week so it's pretty obvious the SATA controllers on that motherboard can handle at least 2 GiB/s.

I've honestly never seen a motherboard with SATA-3 6 Gbps controllers so bad that they couldn't handle a full 500 MB/s on every port.

If there is a bottleneck it would be the connection to the board chipset from the CPU. Ryzen and its board support doesn't seem to have any problems here. Especially not if the board and CPU are able to use PCIe 4.0.

My B450 board with Ryzen 2700 has only 4 usable SATA ports (if a M.2 drive is connected, the SATA ports 5 and 6 are disabled) and the 4 disk scrub in FreeNAS is fairly slow compared to the sum of the speed of each drive. Heavy FreeNAS users confirmed many times on the forum that the performance of a dedicated controller is visibly (not only benchmark measurable) better than the onboard SATA. I did not test, personally, just pointed to the FreeNAS forums.

Lame. I thought one of these was finally going to replace my 4770K. Guess I'll wait a few more years.

Reply on news.ycombinator.com