A long rant about how I hate Socket AM4 but love the Zen SoCs, and how I would like Ryzen/EPYC Embedded alternatives

zir_blazer

Senior member
Jun 6, 2013
881
17
81
#1
I love the Zen architecture and think that the original Zeppelin and Raven Ridge are impressive pieces of silicon, but I have a problem with the consumer AMD ecosystem: I hate the Socket AM4 platform as a whole.
The story begins at ISSCC 2018, which is when AMD did this presentation detailing Zeppelin SoC features:
https://www.slideshare.net/AMD/isscc-2018-zeppelin-an-soc-for-multichip-architectures

...then, after seeing articles analysing the slides like this one:
https://fuse.wikichip.org/news/1064/isscc-2018-amds-zeppelin-multi-chip-routing-and-packaging/2/

...I began to believe how much Socket AM4 is limiting Zen based SoCs true potential, and how Ryzen Embedded V1000 (Raven Ridge) and EPYC Embedded 3000 (One or two Zeppelin dies, equivalent to AM4 8C Ryzen or TR3 16C ThreadRipper) are actually far superior since they expose many more features. Just check their Product Briefs:
https://www.amd.com/system/files/documents/3000-family-product-brief.pdf
https://www.amd.com/system/files/documents/v1000-family-product-brief.pdf



The problem is that Zeppelin and Raven Ridge are not merely a CPU or APU, they are actually full blown SoCs (This will stop being true in Zen 2 generation due to I/O and CPUs being separate dies, but the Processor package as a whole would still be a SoC). Zeppelin has an impressive range of integrated I/O functions, including the classic IMC, 32 PCIe Lanes that are actually multiprotocol, an USB Controller with four 3.0 Ports (Or more...), and miscellaneous stuff like Azalia/HD Audio, SPI, LPC, I2C and SMBus. Depending on the package and platform that Zeppelin is delivered in, several of those features are not used to their fullest extend, and this is the reason why I despise Socket AM4. Raven Ridge seems to have major differences (More on those later) and is perhaps even better suited for a consumer facing SoC than Zeppelin is.


The Integrated Memory Controller comes first. Zeppelin actually has two 64/72 Bits IMCs, which support the entire range of DDR4 DIMM standards: UDIMM, UDIMM w/ECC, RDIMM and RDIMM w/ECC. In Socket AM4, only UDIMM and (Somewhat oficially) UDIMM w/ECC are supported. EPYC Embedded, instead, supports them all. Interesingly, the Product Brief of EPYC Embedded mentions NVDIMM support, but this may not be a direct reference to Optane NVDIMMs in LGA 3647 Cascade Lake, which is the only one that I'm aware that can do that.
In the case of Raven Ridge, memory support is a bit... curious. In Socket AM4, standard Ryzen APUs does not support ECC whereas Ryzen Pro APUs do (Sure, go ahead and tell me where the hell can I buy one), but this has to be compared to Ryzen CPUs where even the non Pro ones supports ECC. Ryzen Embedded is even more curious, since while ECC is officially supported, AMD claims that it can use only 1 DPC (DIMM Per Channel) which means only two DDR4 Slots, whereas the standard AM4 Ryzen APUs that should be based on the same Raven Ridge die works fine with four slots/2 DPC. I think that this may be because AMD claims 3200 MHz DDR4 support on Ryzen Embedded which may not be possible with 2 DPC, but I don't know why they didn't tiered it like this:
https://www.pugetsystems.com/blog/2018/06/06/2nd-Gen-AMD-Ryzen-Supported-RAM-Speeds-1175/
I'm somewhat dissapointed about why Ryzen Embedded lacks 2 DPC support. Also, there is no mention about RDIMM.


Most of the fun in the Zen SoCs is in the multiprotocol controllers. Intel had been using these for a while in its Chipsets under the name of Flex IO, where several lanes can be configured to work with one of multiple protocols, usually either PCIe or USB, or PCIe, SATA and an 1G MAC. Here is an example:
https://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/5

Zeppelin has two 16 lane multiprotocol controllers which is where the maximum of 32 PCIe Lanes comes from, but as lanes can be configured in different ways and most times a minimum of other I/O is preferable, is almost impossible that you actually see a Zeppelin being used only for PCIe. Moreover, the controllers are not identical, they have slighty different capabilities. Reelevant to them, one controller with the full 16 lanes is lost in setups with 8 Zeppelins like those found in Dual EPYC 7000 platforms since it has to be configured for IF (Infinity Fabric). For up to 4 Zeppelin dies like a Single EPYC 7000, the multiprotocol controllers aren't touched since there are four dedicated IF controllers for die-to-die communications that covers up to 4 dies, which is why ThreadRipper with 2 Zeppelin and Single EPYC with 4 Zeppelin has a total of 64 and 128 lanes, respectively, whereas Dual EPYC remains at 128 but with each Processor providing 64.
The first 16 lane controller can be set up as either PCIe or IF (Infinity Fabric), but as the IF mode is only used in the mentioned Dual Socket 8 Zeppelin setups, for anything not Dual EPYC 7000 it is always treated as a pure PCIe Controller, and the main one at that. In PCIe mode, the PCIe lane granularity is amazing since it can scale down from one 16x port to sixteen 1x ports, with any middle ground configuration like 8x/8x, 8x/4x/4x, 8x/4x/1x/1x/1x/1x and such apparently supported (I think that there is a limit that caps out at a total of 8 PCIe Ports, not sure if per controller or between both controllers, but either way, sixteen 1x may not be possible). For comparison, Intel consumer platforms (LGA 1155/1150/1151) integrated PCIe Controller only supports 16x, 8x/8x and 8x/4x/4x modes (Also, bifurcation of the Processor PCIe Controller is ONLY supported if you paid for a high end Chipset. It falls under the "Processor feature controlled by the Chipset" category, which is a pure artificial limitation). This is important not necessarily due to the physical amount of PCIe Slots, but due to in-slot bifurcation for when a single PCIe Card has multiple individual PCIe Devices. Major examples are NVMe adapters like the AsRock ULTRA QUAD M.2 CARD, where the second M.2 slot would not be functional in an Intel consumer platform because it can't go below 8x/4x/4x, yet it should work fine with Zeppelin. If you can't bifurcate the way you need, you would have to get more expensive PCIe Cards that include a PCIe Switch like the PLX PEX series. I'm not sure if Socket AM4 has any limitation that doesn't allow this controller to bifurcate beyond 8x/8x, as I don't recall seeing 4x/4x/4x/4x to get cards like the mentioned one working in an AM4 Motherboard, but such capability is exposed in ThreadRipper.

The second multiprotocol controller is far more interesing as it supports more protocols. Besides IF and PCIe, as can be seen on Slide 13 of AMD original presentation, half of it supports SATA, configurable on a per lane basis, so possible configurations ranges from 16 PCIe Lanes and 0 SATA to 8 PCIe Lanes and 8 SATA. Additionally, it was revealed during the EPYC Embedded launch that this controller also supports pairing two lanes to provide a 10G MAC (It has to be coupled with either a 1G or 10G PHY to get full NIC functionality. This is similar to Intel Chipsets, they provide an integrated 1G MAC which has to be paired with a PHY like the Intel i219-V), and up to 4 of them are supported, thus using 8 lanes. However, due to the lack of technical information, I don't know if the 10G MAC shares the same lanes than SATA, so that you could get either 8 SATA or 4 10G MACs, or a combination like half and half, or if they are multiplexed into the other set of 8 lanes which means that you could do 0 PCIe Lanes, 4 10G MAC, and 8 SATA out of this controller. In Socket AM4, this controller isn't fully exposed, instead, only 8 lanes (Probably the SATA half) are available, and 4 of those always go connected to the Chipset. This is where Zen loses a lot of I/O potential.
Raven Ridge is actually quite different. In theory, the first controller is still 16 lanes, but it is configured as 8 PCIe and 8 IF (Unsupported arrangement in Zeppelin, since it is either all PCIe or all IF), with the IF half used to wire to the integrated GPU. These are the famous 8 lanes that you're missing, and the reason why Ryzen APUs can get only a 16x PCIe Slot working as 8x, or in 8x/8x Motherboards, the second slot doesn't work. The second controller is also a bit weird, as given Socket AM4 capabilities it should at least expose 8 lanes, yet Raven Ridge supposedly only supports 2 SATA. Due to lack of info, I'm not sure if the second controller is as capable as Zeppelin one, but in Ryzen Embedded, it is capable of providing two 10G MACs, albeit I don't know what are the possible arrangements.


Next comes the USB Controller. For some reason, compared to Intel Chipsets, USB in AMD SoCs is independent instead of being multiplexed into some of the PCIe lanes. Zeppelin has 4 USB 3.0 (5 Gb p/s) Ports, and this is perhaps its most blatant weakness as a consumer SoC since that is not enough for a full fledged desktop computer. One option was for Super I/O chips to provide a PS/2 Keyboard Port to maybe save an USB, but this is a backwards solution. In some EPYC Embedded Motherboards, like the Supermicro M11SDV-8C+-LN4F (You can check the Block Diagram in its Manual), there are onboard USB Hubs to fanout an USB 3 Port to multiple USB 2 Ports, but this arrangement is not ideal either. Wasting PCIe Lanes on separate USB Controllers is viable but at that point you may as well consider adding back the Promontory Chipset. Spending 2/4 PCIe Lanes on an Intel ThunderBolt 3 Controller with 1/2 Ports is also an option.
Raven Ridge has, again, interesing differences compared to Zeppelin. It has 4 USB 3.1 Ports capable of 10 Gb p/s instead of just 5 Gb p/s, and in addition to them, is has two more Ports, a single 5 Gb p/s, plus an USB 2.0 (Which should be XHCI based). With 6 Ports it gets a bit more viable to drive a desktop platform with no additional help.


Finally, you have miscellaneous I/O. Both Zeppelin and Raven Ridge have them in amounts that I didn't really paid attention to, but at the bare minimum, you have SPI/eSPI, which is used by the SPI Flash EEPROM with the Motherboard Firmware, Azalia (Also known as HD Audio), used by the HDA audio codec (The Realtek ALC series chips that you see integrated in every consumer Motherboard hold a monopoly over this Bus), and LPC, used by the Super I/O chip. Actually, whereas in all previous platforms the Chipset takes care of all this miscellaneous I/O, the Zen SoCs, even in Socket AM4 packaging, are wired to them directly:
https://www.kitguru.net/components/...r-direct-nvme-storage-promontory-pch-usb-3-1/
Zen has also built in I2C/SMBus, but I'm not sure about whenever they're either exposed or used in Socket AM4 platforms. Typically, SMBus is managed by the Super I/O chip, which is wired via SMBus to the DIMM Slots to read a DIMM SPD EEPROM, and it can also be optionally exposed in PCIe Slots if the Motherboard manufacturer decided to wire one or multiple slots with SMBus support (This is uncommon).



Even though as the embedded Zen SoCs versions shows us that the Zen SoCs could drive a platform entirely by themselves, AMD decided to stick with a Processor plus Chipset platform topology. Moreover, as AMD said that they were commiting to Socket AM4 for at least one more generation, this situation is not going to change for at least one or two more years, before AMD revisits if it wants a bigger Socket with more room for I/O pins or not. So, what the current Chipsets actually does for the Socket AM4 platform? Based on the fact that even miscellaneous I/O is handled by the Zen SoCs, and the few I/O that isn't is done by the Super I/O, the Zen Chipsets are downgraded to be merely just glorified PCIe Switches, SATA Controllers and USB Controllers. Actually, Promontory (Both 300 and 400 generations) completely sucks at being a PCIe Switch due to it exposing only PCIe 2.0 Lanes, and worse if you're of the PCI Passthrough crowd as these lanes offer no PCIe ACS support, being the cause why every AM4 Motherboard in Linux has all the Chipset devices grouped in a single IOMMU Group (Maybe the 500 series is competent enough in this regard). As everything that Promontory does can be done by the Zen SoCs better than it, the Zen Chipsets seems a bit redundant to me...
Now, the idea behind the Zen Chipsets is that they work as a form of I/O multiplexer so that the Processor package is not clogged with all the required pins that would be needed to provide all optional I/O itself (It could be that it was impossible to have more pins in Socket AM4 package, thus some I/O had to be sacrificed). In both Intel and AMD platforms, the Chipset takes 4 PCIe Lanes (Intel calls the Processor-to-Chipset Bus DMI 3, but it seems to be just a different protocol with physical PCIe properties), and provides a plethora of I/O pins, which are connected to physical Ports. Under normal circunstances, it is not expected for everything to be under heavy use all at once, so you shouldn't feel the massive bottleneck that the uplink between the Processor-to-Chipset really is (Well, except on Intel consumer platforms, where a single PCIe 3.0 NVMe SSD can saturate the Chipset uplink all by itself, yet Intel wants to drive two or more out of the Chipset. Socket AM4 actually has 4 lanes from the SoC used for this purpose). Point is, as the Zen SoCs already have enough I/O builtin to drive a consumer platform, unless you require more than what they provide (A problem with USB Ports only, as you have more of enough of everything else), you may not require a Chipset at all for multiplexing further I/O. Not only that, but by wiring everything directly to the SoC, you're enjoying both lower latency, as data would not have to travel though a middleman chip, dedicated bandwidth for all I/O, lower power usage as there is one less chip, and lower costs for that same reason, and no Motherboard real state for it.
Ironically, AMD missed a rather good chance to do a full demonstration of Zen SoC capabilities back at launch. I'm sure that someone should remember that AMD announced the A300/X300 Chipsets for mITX platforms, which were not going to provide any I/O as the Zen SoCs had enough to fill a mITX sized Motherboard, instead, they were to serve just as some form of platform security chips. At the end, these never appeared on the market, mITX Motherboards simply used the standard Chipset versions. In truth, these Chipsets were unnecessary to being with, albeit it seems that the Socket AM4 platform is what forces you to have a one, whereas in embedded form, Zen SoCs don't require it and can work standalone. Alas, you would have had to sacrifice socket upgradeability for an embedded Zen, but if you believe the way I do, that may not necessarily be a big deal for some kind of long term setups.

The chip which Zen can't replace is apparently the Super I/O. There are two functions that the Zen SoCs doesn't seems capable of doing: Being a PWM Fan Controller, and being an ACPI Embedded Controller. Amusingly, the previously mentioned Supermicro M11SDV-8C+-LN4F seems to not have a Super I/O at all, instead, it has an ASpeed AST2500 BMC, which is a dedicated Processor with a miniGPU intended for remote management, but apparently it can also supercede the functions of a standard Super I/O. Thus, with these two chips you're already providing the vast majority of features of a platform.
Also, for remote management, AMD supports DASH, which it claims that is similar to Intel vPro. DASH support is supposed to be embedded in the Zen SoC itself, but as a feature both it and its usage are pretty unknow and not very well covered. DASH requires Motherboard support, and very, very few do. Besides, given the fact that Server Motherboards commonly uses dedicated BMCs, it not sure what use cases vPro and DASH are supposed to be good for, but at least I'm aware that vPro works, and DASH... who knows.


So, after reading all this, you're wondering if using Zen as a SoC like I want is truly possible, or just a hypothetical scenario. If you were following the Embedded Ryzen and EPYC products, you should already know that products using them the way I'm proposing already exist: Supermicro has several Embedded EPYC 3000 based Motherboards, and there were also the UDOO BOLT and the Sapphire AMD FS-FP5V based on the Ryzen Embedded V1000. The problem with the existing products is that they're either too expensive due to being Server oriented, which is the case of Supermicro Motherboards, or too small with almost no expansion due to being HTPC oriented. I think that even if one was limited to the preexisting embedded Zen models, they could be an interesing niche in the consumer DIY Motherboard for people that know what they're getting.

For example, the Ryzen Embedded V1756B is quite similar to the Ryzen 2400G, and so is the EPYC Embedded 3251 when compared to the Ryzen 1700 (The 2700 is Zen+ based, thus not as close comparison). They also cost around the same from AMD in quantities of 1000. The worst con is that you're purchasing an inseparable Processor + Motherboard pair, which is not end user upgradeable as socketed systems are (Ironically, AMD long longevity Sockets hurts more how embedded versions could compare than say, Intel...). This may not even be an issue for people that don't upgrade the platform often, which should be most non-enthusiasts users. You're also having a non-standard heatsink mounting mechanism, since embedded Motherboards usually have a custom one due to the nonstandard height of the embedded packages, so forget about tower heatsinks and overclocking in general. The models have also lower TDP, but also lower clocks. In the pro section, you get all the features advantages of a pure Zen SoC, having no ZIF Socket or Chipset should also lower costs, and while in general there would be less I/O, the one that you get will perform equal or better.
Where does this leaves a supposed Ryzen Embedded V1756B and a coupled Motherboard at? As a more specialized 2200G/2400G + A320 based Motherboard replacement for the masses, that should be around the same price than them but with a much more interesing builtin feature set.
 
Last edited:

piesquared

Golden Member
Oct 16, 2006
1,650
107
136
#2
wow i didn't think it was possible to navigate such a narrow path to find fault with the AM4 socket.
 
Aug 25, 2001
44,098
705
126
#3
Good read. (*I read it all.)

I have had similar musings, only when I found out that Ryzen APUs could only support PCI-E x8 lanes on the primary PCI-E slot, because, presumably, they use those lanes for the video output. (I don't think that you touched on that?) If they aren't multi-plexing the pins on the SoC, but instead, the pin-outs on the socket, then that's a shame, IMHO.

Edit: Thanks for explaining it all in detail.
 
Last edited:

B-Riz

Golden Member
Feb 15, 2011
1,178
270
136
#4
LOL OMGWTFBBQ

This was a design focused on density of socket for the new reality of a cloud based / virtual data-center server world.

Make it great for servers (lower clocks, not clocking a single thread to the moon) and trickle down to something consumers can use, the modular approach.

Zen and Zen+ are making do on a mobile first process, which is great for lower power server focused chips.

Think of them as 80% good enough to get out the door while the real gem was Zen2 on a higher performance process.

In the grand scheme of things we may be the noisiest users, but we pale in comparison to the millions and millions of dollars companies spend every year on data-center and related products.
 

zir_blazer

Senior member
Jun 6, 2013
881
17
81
#5
Good read. (*I read it all.)

I have had similar musings, only when I found out that Ryzen APUs could only support PCI-E x8 lanes on the primary PCI-E slot, because, presumably, they use those lanes for the video output. (I don't think that you touched on that?) If they aren't multi-plexing the pins on the SoC, but instead, the pin-outs on the socket, then that's a shame, IMHO.
I didn't mentioned anything about the video output, but I did about the missing 8 lanes. Supposedly they're present, but internally used as IF for the IGP.

Due to the fact that AMD hasn't released any detailed data sheets about anything AM4/Ryzen related, there is no info about if they do multiplexing or anything. It may be possible that Raven Ridge in Socket AM4 reuses the pins that would be used by the 8 PCIe Lanes for the video outputs, as, after all, its either one in Ryzen APU, or the other in Ryzen CPU. Actually, I don't recall having seen any sort of detailed slides about Raven Ridge like AMD did with Zeppelin, and even in the Zeppelin slides AMD didn't mentioned about the multiplexed 10G MACs that later appeared in Ryzen/EPYC Embedded.
Also, if I recall correctly, Socket AM4 RR can have up to 3 video outputs whereas Embedded Ryzen V1000 supports 4.


This was a design focused on density of socket for the new reality of a cloud based / virtual data-center server world.

Make it great for servers (lower clocks, not clocking a single thread to the moon) and trickle down to something consumers can use, the modular approach.

Zen and Zen+ are making do on a mobile first process, which is great for lower power server focused chips.
This is not even tangentially related to what I'm talking about. The whole point about my rant, is that AMD is currently providing products that fully exposes the Zen SoCs capabilities (The Ryzen Embedded and EPYC Embedded lines), that are sold at around the same price than the socketed counterparts of similar specifications (Except consistently lower clocks across the board, but I could live with that), and that there is an interesing set of pros and cons that would make taking a Zen SoC soldered in a Motherboard a rather interesing proposition against the socketed versions due what I consider overally a superior feature sets and platform simplicity, if such a product actually existed. And I'm here to look for like-a-mindeds.
 

B-Riz

Golden Member
Feb 15, 2011
1,178
270
136
#7
This is not even tangentially related to what I'm talking about. The whole point about my rant, is that AMD is currently providing products that fully exposes the Zen SoCs capabilities (The Ryzen Embedded and EPYC Embedded lines), that are sold at around the same price than the socketed counterparts of similar specifications (Except consistently lower clocks across the board, but I could live with that), and that there is an interesing set of pros and cons that would make taking a Zen SoC soldered in a Motherboard a rather interesing proposition against the socketed versions due what I consider overally a superior feature sets and platform simplicity, if such a product actually existed. And I'm here to look for like-a-mindeds.
The embedded stuff are just targets for a market, same as the socketed stuff. Intel has just as much embedded stuff.

Like, what do you want to buy that you can't? Cause all the emebedded stuff is for industry, not the every day consumer.
 
Oct 27, 2006
19,776
281
126
#8
Threadripper doesn't have these limitations, right?

I have a C600/X79 type build with my son's 1650v2, and love the plethora of PCIe lanes. Makes me a bit wistful comparing it with my Ryzen and 8086k builds. On the Xeon I can do a buttload of network, USB 3.0, USB C (PCIe card), and SATA/HDD/nVME traffic without bogging down nearly as easily.

Of course, I'm a budget guy, and most of the time I'm not really doing all that much multitasking, so it's just fun to think about. Maybe if Zen2 has a boss 16C/32T option, used TR1s may trickle out from people looking for more power/PCIe 4/etc. I'd love to mess with one :)

Best deals I can find on monster power for a budget is D30s, I've seen dual 8-core 2.8Ghz v2 Xeon with K4000 GPU, 256GB Dual Quad Channel Ram (16 8GB DDR3 1600 RDIMMs) and 1200W PSU sell for under $600. Can barely scratch a dent in a new S2011 or TR machine for those kinds of prices. I did see one go by a local auction with dual 2690v2s, which would be dual 10C/20T/3.6Ghz/25M Cache units for 20C/40T lol. Prices for those aren't so good though.

Not cutting edge, but probably lots of fun to play with. I had a D20 for a while and it was great fun.
 
Aug 25, 2001
44,098
705
126
#9
I'm actually a little pissed, that AMD had the opportunity to offer 10G ethernet on AM4 boards, as a "standard feature" (subject to board costs / market positioning, as it still needs a PHY chip), and they didn't do so. I would have really liked that, and it would have really set their platform apart from Intel's platform, IMHO. But I kind of get it, they didn't want to upset the apple cart with their server platforms and pricing. IOW, they "saved 10GbE for enterprise".
 

B-Riz

Golden Member
Feb 15, 2011
1,178
270
136
#10
I'm actually a little pissed, that AMD had the opportunity to offer 10G ethernet on AM4 boards, as a "standard feature" (subject to board costs / market positioning, as it still needs a PHY chip), and they didn't do so. I would have really liked that, and it would have really set their platform apart from Intel's platform, IMHO. But I kind of get it, they didn't want to upset the apple cart with their server platforms and pricing. IOW, they "saved 10GbE for enterprise".
LOL

Do you have the rest of the equipment to make use of it? That seems like the only sticking point.

I ran cat 5e, small house, so, I may be able to get 10GbE speeds.

https://www.smallnetbuilder.com/lan...ions-of-a-10-gbe-network-newbie-part-1-basics
 

amd6502

Senior member
Apr 21, 2017
580
113
86
#11
To me AM4 seems like a mean lean and very powerful platform accomodating a very wide range. My only complaint is that it does not cover AM1-like spectrum very well, nor are there many/any affordable ITX boards to choose from.

IMHO we really need an AM4-L platform, which limits compatibility to 35W (or max cTDP 45W) processors; and something where normal boards come in ITX format, at a very affordable price.
 

LightningZ71

Senior member
Mar 10, 2017
250
17
86
#12
Hey, I get what you're saying, and I wish that such a fully enabled zen/zen+ product was available...

But, it's not, nor would it be. All of those SOC features do not come for free. You want to run a 10g Ethernet chipset? While you run a bunch of USB3.1 links, on top of 32 bifurcated lanes of PCIe, and a host of other SOC functions? It's going to draw a decent amount of power AND generate a bunch of heat. You lament the embedded chips being downclocked? They are giving up that performance for a reason, and it's thermal/power envelope related.

There's another reason: board complexity and price. To expose all those functions on the motherboard would have required essentially another two layers of motherboard AND a larger socket by at least 2x2 rows/columns of pins. That would have made their platform less cost competitive as the PCH chips were effectively less expensive and arguably more flexible.

There is one place where I can see things working out the best for Ryzen as a SOC with the summit/pinnacle ridge chips and that's the mobile workstation and commercial uSFF space. Such a SoC, coupled with an on board dGPU could have made for an excellent product. The final product would have been very cost and performance competitive in it's space

Going forward, things should be a bit different. With a separate IO die, they can customize that die, and those functions for many different markets, and I can see a world where they have a vast array of integrated products. I also wouldn't be shocked to see AM5 grow significantly over AM4 to unmask these features.
 

Glo.

Diamond Member
Apr 25, 2015
3,061
267
136
#13
Wow. A Product Segementation gets blame for... Product Segmentation...
 

zir_blazer

Senior member
Jun 6, 2013
881
17
81
#15
Better later than never, I guess. Still, Ryzen Embedded V1000 would have been superior to it since for as long that you're willing to sacrifice the upgradeable Processor, you could get a few more features. The A300/X300 Chipsets adds no I/O, yet you still are losing what Socket AM4 doesn't exposes from the Raven Ridge SoC. At least, A300/X300 technically frees up the 4 PCIe Lanes typically taken by Promontory. Those Chipsets were connected using a miscellaneous I/O Bus, not sure if it was LPC or SPI.


Threadripper doesn't have these limitations, right?
ThreadRipper does make full use of the two Zeppelins with 64 PCIe Lanes and 8 USB Ports.


I'm actually a little pissed, that AMD had the opportunity to offer 10G ethernet on AM4 boards, as a "standard feature" (subject to board costs / market positioning, as it still needs a PHY chip), and they didn't do so. I would have really liked that, and it would have really set their platform apart from Intel's platform, IMHO. But I kind of get it, they didn't want to upset the apple cart with their server platforms and pricing. IOW, they "saved 10GbE for enterprise".
I know that fell, bro. AMD lost the chance to single handedly push 10G equipped computers. Imagine my face when I saw high end AMD Motherboards like the AsRock Fatal1ty X370 Professional Gaming or the AsRock Fatal1ty X399 Professional Gaming having the Aquantia NICs. The Zen builtin 10G MACs could have utterly destroyed those, perhaps even for a lesser cost as a 10G PHY should be cheaper than a full NIC. Moreover, the Aquantia NICs are usually wired to the Chipset, which means that they don't get dedicated bandwidth but are instead multiplexed with the rest of the Chipset I/O. Good thing that at least on AMD consumer platforms, the Chipset uplink is not as horribly saturated as the Intel ones...

AMD didn't even saved 10G for enterprise as the standard EPYC 7000 series does not support it, either. I think that it was ServeTheHome the ones that mentioned that they asked AMD why they didn't enabled the 10G MACs on EPYC 7000 and AMD said that it was because there was low demand for them, since most enterprise vendors like to include NICs that are already part of their standard Hardware ecosystem. Did AMD had anything to lose had it left them enabled and just waited for some Motherboard manufacturer to pick it up? The only reason I can think about why AMD may not have done so is if it wasn't planning to commit to support it long term, as if it was planning to remove it in later models, whereas with nonupgradeable embedded products this is not a problem.
This is not a price/feature segmentation issue, either. AMD said that the 8C/16T EPYC Embedded 3251 has a 315 U$D price (Obviously you can't use the part standalone since it has to be soldered to the Motherboard, so the price matters more for the Motherboard manufacturers that for you), around the same than the equivalent Socket AM4 Ryzen 1700/2700. Assuming that there is some price parity with standard consumer parts, the EPYC Embedded 3101, which is 4C/4T, should cost for sure less than 150 U$D, as it seems that it still has "up to 4" 10G MACs since I didn't saw any info contradicting that. Only the lowest Ryzen Embedded V1000 model, the V1202B, is limited to two 1G MACs instead of two 10G ones, but I don't know how much these cost (For reference, there was like a 80 U$D price delta between the 2C/4T V1202B equipped UDOO Bolt V3 and the 4C/8T V1605B UDOO Bolt V8). Giving similar prices than the consumer parts we're used to, AMD is basically giving you half a 10G NIC for free.
Want to know what is even more weird? That even though the 10G MACs are almost universally available in the embedded Zen products (Except the V1202B), almost no one uses it. The first product that put them to use was AMD own Wallaby reference platform...

The AMD Wallaby platform is the first that we have seen in over a year of EPYC and over a year and a half of Zen that we have seen the AMD 10GbE NIC.
Now, I understand that standalone 10G NICs chips usually provide far more features than AMD basic connectivity, like Intel with its 10G NICs that support SR-IOV and I think that some offloading mechanism, besides Intel renowned Driver support and quality. However, assuming that the 10G MACs can also work in 1G mode like it seems to be the case of the V1202B, is there any reason why would they still choose a standalone Realtek PCIe 1G NIC chip over a 1G PHY chip for low end models, which is the case in most embedded Motherboards? Is like if everyone was actively avoiding the AMD MACs for some reason. To be honest, I don't get it. Some people even speculated that the AMD MACs had to be broken, flawed or something, as it is impressive how much embedded Motherboard manufacturers had been ignoring them.


To me AM4 seems like a mean lean and very powerful platform accomodating a very wide range. My only complaint is that it does not cover AM1-like spectrum very well, nor are there many/any affordable ITX boards to choose from.

IMHO we really need an AM4-L platform, which limits compatibility to 35W (or max cTDP 45W) processors; and something where normal boards come in ITX format, at a very affordable price.
The AM1 platform should have been embedded from Day One instead of socketed. Or did someone got any sort of mileage by upgrading the Processor at all?
Heck, when I see AM1 Motherboards like the Gigabyte GA-AM1M-S2H, that were mATX sized but powered by an Atom class Processor, I want to cry. Now we have a reverse scenario: EPYC Embedded is too big for mITX form factors as there is no physical room to expose everything it has, yet no one does a mATX sized one.

You don't need AMD to create another Socket specification for Motherboard manufacturers to unilaterally derate what a specific Motherboard supports, with undersized VRMs or such. Check the AsRock Deskmini A300 posted above, which is limited to 65W TDP APUs.
Alas, is there a real reason why it has to be socketed? There are already many mITX sized products based on Ryzen Embedded V1605B, which is 4C/8T with a 15W TDP, and provides every feature that was cut from Socket AM4 (Whenever these features gets used or not depend on the Motherboard itself). Check the UDOO Bolt or any from this list. These embedded products "fixes" the Socket AM4 shortcomings that I'm ranting about here. Do these dissapoint you? I don't see any reason why you can't have something like that but with higher end consumer availability.


Hey, I get what you're saying, and I wish that such a fully enabled zen/zen+ product was available...

But, it's not, nor would it be. All of those SOC features do not come for free. You want to run a 10g Ethernet chipset? While you run a bunch of USB3.1 links, on top of 32 bifurcated lanes of PCIe, and a host of other SOC functions? It's going to draw a decent amount of power AND generate a bunch of heat. You lament the embedded chips being downclocked? They are giving up that performance for a reason, and it's thermal/power envelope related.

There's another reason: board complexity and price. To expose all those functions on the motherboard would have required essentially another two layers of motherboard AND a larger socket by at least 2x2 rows/columns of pins. That would have made their platform less cost competitive as the PCH chips were effectively less expensive and arguably more flexible.

There is one place where I can see things working out the best for Ryzen as a SOC with the summit/pinnacle ridge chips and that's the mobile workstation and commercial uSFF space. Such a SoC, coupled with an on board dGPU could have made for an excellent product. The final product would have been very cost and performance competitive in it's space

Going forward, things should be a bit different. With a separate IO die, they can customize that die, and those functions for many different markets, and I can see a world where they have a vast array of integrated products. I also wouldn't be shocked to see AM5 grow significantly over AM4 to unmask these features.
This is a long one.

The thing is, the products ARE available, at least from AMD side. Ryzen/EPYC Embedded got everything I want, which is exposure of the full SoC capabilities. The problem is that you need to get them in a Motherboard that has the usual consumer feature set, and that is also easily available for end consumers in the retail DIY market. That is what doesn't exists yet. I can go and purchase a Ryzen 2200G and an A320 Motherboard everywhere, but the vast majority of the mentioned Ryzen Embedded Motherboards require industrial channels (There is no individual unit MSRP, I have to request a quote). I'm convinced that a budget Motherboard based on Ryzen Embedded available in retail channels would completely crush the 2200G/A320 combination. OEMs would love it, too.

Yes, using all the SoC features would consume more power, but they are not the cause of the lower clocks. The embedded products are aimed for a market with low power consumption in mind, so they have lower TDP than desktop counterparts, and in consequence, lower clocks. I suppose that if they were rated for the same TDP, base clocks should be slighty lower if everything of the SoC was being used at the same time, but turbo the same than desktop assuming that the other features were not being used.
Also, overall platform power consumption should be significantly lower, as Promontory is made using a 55nm process compared to 12/14nm in Raven Ridge and Zeppelin. If X370 has 6.8W TDP (Mentioned as Full Load) and X470 4.8W peak, I doubt that simultaneously using all what the SoCs already has lowers base clocks by more than 200 MHz. Perhaps it is even better than that, as Promontory has far more I/O while Zen SoC would expose only 8 more lanes @ 14nm power consumption, so maybe it actually has a negligible impact.

Do you have a source that going embedded is going to make the Motherboards more expensive? As far that I know is the other way around. They are actually saving the ZIF Socket cost, which is supposed to be an expensive component of a Motherboard BoM. EPYC Embedded uses the SP4 package, which I couldn't find exact measures but seems around the size of AM4. It is used by both the single die and dual die versions. The dual die exposes roughly twice the I/O (64 lanes vs 32) with a package size similar to AM4. BGA is THAT more dense.

If you didn't saw the Supermicro EPYC Embedded Motherboards, check them. They are precisely the sort of mITX powerhouses + possible discrete GPU that you are describing.

The thing about going embedded is that the designer can define the features that they want per generation and make full use of them. Long term Socket support is troublesome if you have customizable I/O, because there is little point in doing so if Motherboards using a specific Socket already tells you what you can expose and in what pins would be exposed, because if you change the definition or protocol of the pins, the Processor will either not be compatible, or some features will not work, or require a new Motherboard to be exposed. I think that embedded is far superior for platforms that will not be upgraded during their generation lifetime. This is usually the case of budget users, they purchase a low end computer and will stick with it for years. Ryzen Embedded has more advantages than disadvantages for these users.
 
Last edited:

LightningZ71

Senior member
Mar 10, 2017
250
17
86
#16
I speculate that enabling those processor features also has a licensing cost associated with it. It MAY be free for them to put the circuits on each processor, but they have to actually pay for those IP licenses when they are actually connected to a motherboard. Just a thought. I have actually run into that in other IC products in the past.
 

amd6502

Senior member
Apr 21, 2017
580
113
86
#17
The economics of fabrication makes it such that server products are co-produced with either mobile and/or desktop. And that is why it happens that there are dedicated server logic areas on the die that get gated off in consumer products. It is not only product segmentation. Some of these server features may be under IP license. Secondly, the AM4 platform kind of needs to stay lean and mean so that it is not overly complex. This could cause unforseen issues and bugs, and cost both AMD and the board makers a lot of staff and designers. It is mean and lean so it can be affordable to the masses. There is TR4 platform for those that want more server like features, including doubled memory bandwidth.

The AM1 platform should have been embedded from Day One instead of socketed. Or did someone got any sort of mileage by upgrading the Processor at all?
People could have and should have gotten much more mileage from the AM1 platform, and they would have, if support for that platform hadn't ended so prematurely. There were products that could have gotten ported to AM1, such as puma APUs and Stoney CPUs.

The Deskmini A300 is a pretty good suggestion; yes it probably is the best low cost SFF option out there. I'm not a huge fan of 19V DC PSU but still it is very nice.
 
Aug 25, 2001
44,098
705
126
#18
Secondly, the AM4 platform kind of needs to stay lean and mean so that it is not overly complex. This could cause unforseen issues and bugs, and cost both AMD and the board makers a lot of staff and designers. It is mean and lean so it can be affordable to the masses. There is TR4 platform for those that want more server like features, including doubled memory bandwidth.
I see your point, that's a very good point, actually. Does the TR4 socket bond out the 10GbE NIC support, by any chance?
 

NTMBK

Diamond Member
Nov 14, 2011
8,320
315
126
#19
I'm actually a little pissed, that AMD had the opportunity to offer 10G ethernet on AM4 boards, as a "standard feature" (subject to board costs / market positioning, as it still needs a PHY chip), and they didn't do so. I would have really liked that, and it would have really set their platform apart from Intel's platform, IMHO. But I kind of get it, they didn't want to upset the apple cart with their server platforms and pricing. IOW, they "saved 10GbE for enterprise".
I mean, 99% of consumers use WiFi. They don't want to drill holes in their walls just to get slightly lower ping in Quake, or whatever. And the main use of networking in the home is to connect your PC to the Internet, and consumers sure as hell aren't going to have >1Gb/s internet connections. If you actually need 10Gb/s networking, pay for the enterprise gear.
 
Aug 25, 2001
44,098
705
126
#20
I mean, 99% of consumers use WiFi. They don't want to drill holes in their walls just to get slightly lower ping in Quake, or whatever. And the main use of networking in the home is to connect your PC to the Internet, and consumers sure as hell aren't going to have >1Gb/s internet connections. If you actually need 10Gb/s networking, pay for the enterprise gear.
Where's the fun in paying for enterprise gear? What's to say that Ryzen / AM4 couldn't have sparked a "home networking revolution", just like they did with their processors, by offering built-in 10GbE ethernet?

Anyways, some of us do have some NAS units and like to file-share and backup on the LAN. 10GbE would help that. And it's not like Ryzen (CPUs, at least), needing a dGPU rather than include an iGPU (like the APUs do), is more enthusiast-oriented, in my opinion. At least offer it with TR4 platform, the pro-sumer / HEDT platform.

Then again, hopefully sometime soon, Club3D will have their USB 3.1 Gen 1 to 2.5GbE adapters out, and I can upgrade (partially) on the cheap. Just have to look out and make sure that the switches are "multi-Gig", and not simply 10GbE, as many existing 10GbE switches, will only auto-select down to 1GbE or 100MbE, and not 5GbE and 2.5GbE.
 

NTMBK

Diamond Member
Nov 14, 2011
8,320
315
126
#21
Where's the fun in paying for enterprise gear? What's to say that Ryzen / AM4 couldn't have sparked a "home networking revolution", just like they did with their processors, by offering built-in 10GbE ethernet?

Anyways, some of us do have some NAS units and like to file-share and backup on the LAN. 10GbE would help that. And it's not like Ryzen (CPUs, at least), needing a dGPU rather than include an iGPU (like the APUs do), is more enthusiast-oriented, in my opinion. At least offer it with TR4 platform, the pro-sumer / HEDT platform.

Then again, hopefully sometime soon, Club3D will have their USB 3.1 Gen 1 to 2.5GbE adapters out, and I can upgrade (partially) on the cheap. Just have to look out and make sure that the switches are "multi-Gig", and not simply 10GbE, as many existing 10GbE switches, will only auto-select down to 1GbE or 100MbE, and not 5GbE and 2.5GbE.
"Some of us" = a hobbyist niche of a hobbyist niche that isn't worth targeting.
 

B-Riz

Golden Member
Feb 15, 2011
1,178
270
136
#22
Where's the fun in paying for enterprise gear? What's to say that Ryzen / AM4 couldn't have sparked a "home networking revolution", just like they did with their processors, by offering built-in 10GbE ethernet?

Anyways, some of us do have some NAS units and like to file-share and backup on the LAN. 10GbE would help that. And it's not like Ryzen (CPUs, at least), needing a dGPU rather than include an iGPU (like the APUs do), is more enthusiast-oriented, in my opinion. At least offer it with TR4 platform, the pro-sumer / HEDT platform.

Then again, hopefully sometime soon, Club3D will have their USB 3.1 Gen 1 to 2.5GbE adapters out, and I can upgrade (partially) on the cheap. Just have to look out and make sure that the switches are "multi-Gig", and not simply 10GbE, as many existing 10GbE switches, will only auto-select down to 1GbE or 100MbE, and not 5GbE and 2.5GbE.
Have you priced out some used 10GbE equipment to setup a home LAN? What numbers do you get? I still think it can be cost prohibitive compared to gigabit and just running a second cable to do some NIC teaming.

I don't think AMD alone would spark a home networking revolution. Intel would have to push 10GbE too.
 

zir_blazer

Senior member
Jun 6, 2013
881
17
81
#23
The economics of fabrication makes it such that server products are co-produced with either mobile and/or desktop. And that is why it happens that there are dedicated server logic areas on the die that get gated off in consumer products. It is not only product segmentation. Some of these server features may be under IP license. Secondly, the AM4 platform kind of needs to stay lean and mean so that it is not overly complex. This could cause unforseen issues and bugs, and cost both AMD and the board makers a lot of staff and designers. It is mean and lean so it can be affordable to the masses. There is TR4 platform for those that want more server like features, including doubled memory bandwidth.
The only thing which I didn't knew is that there could be selective licensing about exposing features which are already on the silicon on top of the base cost of the chip instead of being fused off on a per model basis, as usual. The closest thing which I can think about is when Intel experimented with upgradeable Processors back with Sandy Bridge. But that is a mere hypothesis, we don't know if AMD does that or not, and I will be dissapointed if that was the case as I never hear that AMD ever gouged customers like that.

TR4 doesn't have "server like features", AMD targeted it as HEDT, so it is, at most, Workstation. Otherwise, get me a TR4 Motherboard with a BMC/IPMI, which is perhaps the feature that most defines a Server oriented Motherboard. At least Socket AM4 has like two or three Motherboards with BMC (One of those is the AsRock X470D4U). In other communities (ServeTheHome) there were a lot of people dissapointed that the lack of AM4 Motherboards with IPMI means that Ryzen can't be a serious contender against Xeons E3 for Server usage. This isn't even the fault of AMD, because there is nothing stopping the Motherboard manufacturers to create such products, like AsRock did. There is nothing stopping them either from picking a product aimed for the embedded market and putting it in consumer oriented products, which is what I'm trying to push here, for users that prefers the full capabilities of the Zen SoC at the cost of not being able to upgrade the Processor.


People could have and should have gotten much more mileage from the AM1 platform, and they would have, if support for that platform hadn't ended so prematurely. There were products that could have gotten ported to AM1, such as puma APUs and Stoney CPUs.

The Deskmini A300 is a pretty good suggestion; yes it probably is the best low cost SFF option out there. I'm not a huge fan of 19V DC PSU but still it is very nice.
From my point of view, the AM1 platform was mishandled. The Jaguars would have been ideal for cheap, low power, small mITX systems, but instead, you got a truckload of mATX sized Motherboards that were far too expansible for such low processing level. Motherboard manufacturers tried to scale up a Netbook APU to desktop instead of creating a subdesktop segment. Again, you have the Motherboards manufacturer to put the blame on. That scenario was similar to what usually happens to AMD Notebooks: The APUs are decent, the OEM makes them look bad by using a single DIMM instead of two for Dual Channel, and so on.

I insist: A Ryzen Embedded V1000 would have been superior to the AM4 based Deskmini A300. The Chipset adds nothing, and being AM4 based means less exposed features. mITX makes a very good scenario for a consumer oriented Ryzen Embedded. And I expect the end user price to be around the same.


Have you priced out some used 10GbE equipment to setup a home LAN? What numbers do you get? I still think it can be cost prohibitive compared to gigabit and just running a second cable to do some NIC teaming.

I don't think AMD alone would spark a home networking revolution. Intel would have to push 10GbE too.
I'm aware that 10G Routers and Switches are rather expensive. Alas, if you're talking about teaming up dual NICs, assuming that you have an AMD based NAS and main computer with dual 10G NICs, wouldn't it be more convenient to have a crossover cable running between them with the other one going to the standard 1G Switch? While it is not scalable, you could get 10G speeds that way on at least a single computer link for a budget price. At least it could kickstart demand for 10G consumer gear, otherwise you're on a chicked-and-egg scenario where there isn't volume production of consumer 10G routing gear that would eventually drive the price down because there is no installed 10G NIC userbase...
 

NTMBK

Diamond Member
Nov 14, 2011
8,320
315
126
#24
The only thing which I didn't knew is that there could be selective licensing about exposing features which are already on the silicon on top of the base cost of the chip instead of being fused off on a per model basis, as usual. The closest thing which I can think about is when Intel experimented with upgradeable Processors back with Sandy Bridge. But that is a mere hypothesis, we don't know if AMD does that or not, and I will be dissapointed if that was the case as I never hear that AMD ever gouged customers like that.
I think that what amd6502 was suggesting was actually the opposite- that AMD is avoiding paying the companies who supplies the IP integrated into their SoC. Which is entirely feasible! AMD didn't go out and develop an in-house 10GbE controller. They licensed the design from an external company to put onto their die. Presumably AMD has to pay a per-chip licensing fee for this IP, but if the silicon is fused off and unusable then perhaps they don't need to pay the license. Meaning that AMD can reduce their costs, and either pass the savings onto consumer customers or just improve their own profit margins.

You should also bear in mind that OEMs love sockets. A socket means that they can take a single motherboard design and easily offer a range of models just by slotting in a different CPU and different capacity DIMMs. They only need to stock and maintain a single motherboard model, and it significantly simplifies things for them.
 

Tuna-Fish

Senior member
Mar 4, 2011
971
222
136
#25
I'm actually a little pissed, that AMD had the opportunity to offer 10G ethernet on AM4 boards, as a "standard feature" (subject to board costs / market positioning, as it still needs a PHY chip), and they didn't do so. I would have really liked that, and it would have really set their platform apart from Intel's platform, IMHO. But I kind of get it, they didn't want to upset the apple cart with their server platforms and pricing. IOW, they "saved 10GbE for enterprise".
The cost would have been substantial. There is still a significant (>$10 per port) patent license cost associated with 10GbE. Presumably, they have a licensing contract that allows them to not pay it for parts that have it disabled.

The "home networking revolution" will come in 2023 (IIRC), which is when the relevant patents expire and 10GbE becomes cheap.
 

ASK THE COMMUNITY

TRENDING THREADS