Speculation: Ryzen 3000 series

[Arkaign]

Athlon didn't really fail as a foundation, AMD just didn't have the resources to continue what it excelled at: hand optimized die circuits. The K10 (ending with Phenom II) still had superior IPC than the Construction/Cat cores that succeeded it, but the latters' die circuits were completely done by computers. The Construction/Cat cores just were bad follow ons to the K10 line, but AMD learned well with the Zen cores, and the switch to computer generated die circuits gives them the flexibility to do the semi-custom business as well as the near yearly cadence on Zen iterations.


Is anybody really expecting Icelake desktop parts with 8+ cores? Intel's 10nm still seems too much of a mess to allow for any bigger dies or higher frequencies.

10nm Intel is about 7nm TSMC in practical terms, it's just a question of if Intel can get their spit together lol.

However, a huge negative for Intel is that even on equal footing (10nm vs 7nm) they are putting iGPU on die, an absolutely enormous waste of transistors. Hell even today they could cut the iGPU and fit 50-100% more cores on 14nm. On die GPU is something I've absolutely always hated. They're never been good, and a chipset resident version with a bit of dedicated vram would have always been superior. I guess it was just for mobile, but they should have left it there. Notice AMD didn't make the mistake of cramming a GPU onto every Ryzen lol.

 

[inf64]

let me paint you the realistic picture/scenario: when 10nm will have good enough yields to mass produce a 8C processor, it will have to compete with Zen 3 either right away or likely the next quarter

I agree, that was exact thing I said multiple times past week. AMD expects Zen3 to intersect Icelake in 2020. Initially Zen2 was scheduled to go against Icelake this year but due to intel slipping up it's Zen3 vs Sunny Cove in 2020 (better outcome for AMD of course).

 

[piesquared]

10nm Intel is about 7nm TSMC in practical terms, it's just a question of if Intel can get their spit together lol.

However, a huge negative for Intel is that even on equal footing (10nm vs 7nm) they are putting iGPU on die, an absolutely enormous waste of transistors. Hell even today they could cut the iGPU and fit 50-100% more cores on 14nm. On die GPU is something I've absolutely always hated. They're never been good, and a chipset resident version with a bit of dedicated vram would have always been superior. I guess it was just for mobile, but they should have left it there. Notice AMD didn't make the mistake of cramming a GPU onto every Ryzen lol.

I think that after what we have witnessed, it simply isn't true to say that intel's 10nm is equal to TSMC's 7nm. 10nm is broken and 7nm isn't, so whatever intel was shooting for didn't work properly in the real world and they haven't given any information on what the transistor will look like if and when they make the modifications necessary to make it work. So i think it's more accurate to say that 10nm is about 7nm in theoretical or in paper terms, while in practical terms it's just not.

 

[Markfw]

I think that after what we have witnessed, it simply isn't true to say that intel's 10nm is equal to TSMC's 7nm. 10nm is broken and 7nm isn't, so whatever intel was shooting for didn't work properly in the real world and they haven't given any information on what the transistor will look like if and when they make the modifications necessary to make it work. So i think it's more accurate to say that 10nm is about 7nm in theoretical or in paper terms, while in practical terms it's just not.

I think thats what he was saying.

 

[DrMrLordX]

Are any of you going to use/enable the PBO on your 3900x (or any Ryzen cpu for that matter) when you get one? I'm not familiar with overclocking on AMD and how it works.

I'll experiment with it vs. static overclocks. Mine's going to be under water. Assuming I don't bung things up, it should do pretty well. Eventually I expect to have the thing pushing 250W or more, which isn't that much when you consider how much an 1800x can put out on its own.

Could be wrong about the 3700x and it may require such a cooler too but something tells me for stock the supplied stock cooler is going to be fine.

Might want to head over to the builder's thread. The stock cooler for the 3700x should be the same one for the 3600 - matched to TDP of 65W. If you want more out of the chip, you're going to need a bigger HSF. Fortunately, there are some options out there that are reasonably-priced and perform quite well.

Is anybody really expecting Icelake desktop parts with 8+ cores?

No. Yields on the 4c parts are likely abysmal. And Intel is going to be burning a lot of wafers on whatever 26c Skylake-SPs they can crank out, for whoever is goofy enough to want one of those. Desktop IceLake-S may never happen. It's more likely that Intel will release Willow or Golden Cove on 7nm in 2022.

 

[Arkaign]

I think thats what he was saying.

Thanks, lol yeah that was exactly my point. They're apparently starting 10nm in a short time, but it's just low power laptop stuff, and YET AGAIN they're dedicating tons of space (more than half the die) on iGPU. Once again, a slightly less crappy one, but still nothing to write home about.

10nm on boring SKUs 2019

Now they can say 10nm is shipping, but it says something to me about volume and capability that they're doing this garbo first. And so so many transistors on an IGP when launching 6C and 8C 10nm products would be monumentally more interesting. Just move the IGP to a chipset option with say 2GB-4GB embedded DDR4 for products that would be relevant to it.

I think Ryzen 3000 7nm APU and CPU+dGPU will simply be a better option for laptops, which is momentous. Intel topping out at 4C/8T with anything remotely close to reality is basically giving up on anything past Macbook Air style fluff and gimpy cheapo walmart laptops. I mean given the choice between a top 4C/8T Ice Lake laptop and a Ryzen with dGPU, I think that choice is easy. Or even one of the Ryzen 3000 APU models, though to be honest IGP isn't my idea of a good time.

 

[jpiniero]

Matisse not having an IGP really limits the appeal to OEMs. Once they seperate things into CPU and GPU chiplets I think you will see AMD's desktops mostly include the IGP.

 

[DrMrLordX]

Matisse not having an IGP really limits the appeal to OEMs. Once they seperate things into CPU and GPU chiplets I think you will see AMD's desktops mostly include the IGP.

I guess? Nothing like that shows up on their roadmaps. For all we know, Renoir may be monolithic in the same vein as Picasso. It's a big question mark right now.

 

[amd6502]

Matisse not having an IGP really limits the appeal to OEMs. Once they seperate things into CPU and GPU chiplets I think you will see AMD's desktops mostly include the IGP.

Which is why they should have included Vega 3 on the IO hub.

I think they should update matisse ASAP with such an IO hub. (and single chiplet capable would be just fine). the matisse++ derivatives could be used for laptop market too.

 

[Yeroon]

I don't think the missing small igpu is a big issue for the oem market.
I'm going to say igpu oem systems would be at the lower end of the price spectrum. Adding a vega 3 to the i/o die would increase costs to go after a lower ASP market - doesn't make a lot of sense considering amd is trying to raise their ASP.
They already have the monolithic die on 12nm for the lower end of the igpu market. Sure its only 4c/8t, but needing more cores while not being able to spend more for a discrete and a chiplet cpu must be a small market outside of laptops.
Theres a good chance the 7nm monolithic die will increase the core count. So they will get to the higher end laptop market where that might get some higher ASPs for their effort, just not immediately.
Lastly, OEMs have been screwing up APU systems with lower tier discrete gpus than even the igpu for reasons that can only be explained by ignorance or incompetence, which means its probably not as big a deal as it initially seems anyways. If they cared about being cost effective, most haven't shown it.
Not that ignorance is a good reason, but in the short term, its probably not a market worth the effort when they can cover it with their next monolithic die. AMD is just too small to tackle every market at the moment.

 

[Arkaign]

I think AMD could really do well to make a laptop chipset on 5xx with a light GPU in the chipset with integrated vram. It could be relatively tiny and power efficient just to make for h265/4k capability for general use, and because it wouldn't be heating up the same die and sharing memory bandwidth, it could easily be better than Intel IGP as well. It would free up the potential for power efficient 6-8 core laptops for them, and be paired with Navi dGPU in higher end configs.

 

[JDG1980]

Which is why they should have included Vega 3 on the IO hub.

"Vega 3" wouldn't be that much smaller than the "Vega 11" on Raven Ridge. The CUs themselves are not that big; most of the die space on a small GPU is taken up by the I/O and fixed function blocks. Of course, an iGPU has an advantage in this regard, since it can share the memory controller with the CPU.

Polaris 10 is a 123mm^2 chip with 16 CUs. Polaris 12 is a 103mm^2 chip with 10 CUs. Since both have the same architecture, same bus width, and same video codec hardware, we can thus calculate that each extra CU only uses about 3.3 square millimeters of die space on GloFo 14nm/12nm. It might be a bit more for Vega, but probably not much. Going down below 8 CUs cuts out a lot of functionality for minimal area savings, so don't expect to see any iGPUs smaller than that.

 

[amd6502]

"Vega 3" wouldn't be that much smaller than the "Vega 11" on Raven Ridge. The CUs themselves are not that big; most of the die space on a small GPU is taken up by the I/O and fixed function blocks. Of course, an iGPU has an advantage in this regard, since it can share the memory controller with the CPU.

Polaris 10 is a 123mm^2 chip with 16 CUs. Polaris 12 is a 103mm^2 chip with 10 CUs. Since both have the same architecture, same bus width, and same video codec hardware, we can thus calculate that each extra CU only uses about 3.3 square millimeters of die space on GloFo 14nm/12nm. It might be a bit more for Vega, but probably not much. Going down below 8 CUs cuts out a lot of functionality for minimal area savings, so don't expect to see any iGPUs smaller than that.

The CU's are small but add up. The memory controller is already on the IO hub.

I at one point had an estimate. The difference between 3CU and 5 or 6CU would be very small. The difference between 3CU and 11CU would be considerable. [guesstimating ~30mm2 with a 4mm2 CU]

Gotta drawn the line somewhere. And for both Stoney and Raven 2 it was drawn at 3CU, because it's more than enough for movies and basics. Offer 8CU or higher and you're competing with dGPU sales while paying for transistors you don't get revenue back for.

With the faster 12nm transistors 3CU should max out single channel low cost DDR4. 8CU would max out most people's dual channel memory. 6CU could be a happy middle ground and would only add ~10% extra (over 3CU) to an IO hub of ~ 120mm2..

 

[Zstream]

The CU's are small but add up. The memory controller is already on the IO hub.

I at one point had an estimate. The difference between 3CU and 5 or 6CU would be very small. The difference between 3CU and 11CU would be considerable.

Gotta drawn the line somewhere. And for both Soney and Raven 2 it was drawn at 3CU, because it's more than enough for movies and basics. Offer 8CU or higher and you're competing with dGPU sales while paying for transistors you don't get revenue back for.

quicksync is a great tool for Intel because it can encrypt and decrypt rapidly for software targeting that use case. What does AMD have as an equivalent?

 

[piesquared]

I think thats what he was saying.

oh

 

[Topweasel]

Matisse not having an IGP really limits the appeal to OEMs. Once they seperate things into CPU and GPU chiplets I think you will see AMD's desktops mostly include the IGP.

I still don't see this I really don't. I order these business machines for a living. Outside barebones classroom PC's that can get away with a 4c iGPU. All my general use desktops have GPU's. The cheap as heck $30 GPU has almost no impact on buying decisions. What's going to hurt AMD is that OEM's might wait out A520 and B550 before offering Zen 2 systems for business users. People act like AMD screwed up by not adding a GPU in, but its only been 8 months since Intel launched their 8c CPU and its $150 more. That covers a lot of cheap GPUs.

 

[Thunder 57]

quicksync is a great tool for Intel because it can encrypt and decrypt rapidly for software targeting that use case. What does AMD have as an equivalent?

I think you mean encode/decode , I wouldn't mind something similar, but perhaps with a bit more functionality. Have it be capable of encoding x264 and x265, for example. I don't think it's likely we'll see that because they already have the VCE on Radeon. It's just a matter of the CPU's having an iGPU for that to work.

 

[Tup3x]

I think you mean encode/decode , I wouldn't mind something similar, but perhaps with a bit more functionality. Have it be capable of encoding x264 and x265, for example. I don't think it's likely we'll see that because they already have the VCE on Radeon. It's just a matter of the CPU's having an iGPU for that to work.

x264 and x265 are software encoders, not formats. Anyway, AMD is quite a bit behind NVIDIA when it comes to hardware encoding and decoding video. Not sure where Intel is at. I doubt that they will ever include any fixed function hardware for non-APUs though.

 

[beginner99]

Notice AMD didn't make the mistake of cramming a GPU onto every Ryzen lol.

Matisse not having an IGP really limits the appeal to OEMs. Once they seperate things into CPU and GPU chiplets I think you will see AMD's desktops mostly include the IGP.

Exactly. The GPU is for OEM and that is the main difference between AMD and Intel. Intel is targetting OEM while AMD is targeting server and DYI. AMD is in the big picture barley existing in the laptop space which is huge.

AMD was clever go after highest margin (server) as simple and quickly as possible and reuse it for DYI desktop. Deal with OEM / laptop later on after they actually have brand value (negotiation power).

 

[Thunder 57]

x264 and x265 are software encoders, not formats. Anyway, AMD is quite a bit behind NVIDIA when it comes to hardware encoding and decoding video. Not sure where Intel is at. I doubt that they will ever include any fixed function hardware for non-APUs though.

Intel has Quicksync, which is a fixed function hardware encoder for x264. Being specialized hardware it only handles x264. I think the iGPU assists with the decode part of transcoding and then that gets sent to the fixed function hardware to encode it. Could be wrong. Apparently I got that backwards. The decode is all fixed function, and the encode uses the EU's on the iGPU.

Anyway, it is very fast and produces pretty decent results for a quick transcode. Sure, software produces a better quality result, but for quick and dirty Quicksync is great. When it first came out Anand had a good piece on it, but unfortunately there doesn't seem to be much newer information on it, other than it has gotten better and more versatile.

 

[lobz]

quicksync is a great tool for Intel because it can encrypt and decrypt rapidly for software targeting that use case. What does AMD have as an equivalent?

as I learned from some guys here in the past days: only sour grapes

 

[zrav]

AMD has VCE.

Intel has Quicksync, which is a fixed function hardware encoder for x264. Being specialized hardware it only handles x264.

The format is h.264. x264 is a software encoder implementation of it. Quicksync, NVENC and VCE are hardware implementations of a couple of codecs, one of which is h.264.

 

[dnavas]

Quicksync, NVENC and VCE are hardware implementations of a couple of codecs, one of which is h.264.

AMD has VCE.

Which, last I saw a comparison, was the worst of the three. All three are also only as good as the support in software. Both Edius and Premier have support for Quicksync, and at least according to HardwareCanucks, it's about twice as fast

. iirc, Premier also has or had support for cuda as well, but the last numbers I saw on that weren't anything to write home about. Edius uses the QS hardware for decode acceleration as well, and that makes it a lot easier to have a real-time 4k60p timeline during editing, which my 2950 struggles with (after a few edits). In my opinion, that's more valuable than encode acceleration, because I'm happy to let my 4k encodes run overnight.... Even better would be if Neat could be hardware accelerated -- denoising is currently an absolute hog.

All three hardware solutions suffer from being restricted to 4:2:0 video, and so I've frequently dropped the idea of "someone" leap-frogging this stuff and, say, replacing a few of those chiplets with some dedicated hardware for 8k-sized, 4:2:2 and 4:4:4 hardware decode/encode. I think I was also pretty careful to mention that they need to engage software vendors, otherwise the hardware is going to wind up being pointless. I've also spent some time in the Edius forums asking for them to get nvenc working (nvenc has actually seen updates, I haven't heard hide nor hair of VCE :shrug: -- I think AMD has other struggles with their gpus).

Thus far, only Intel has gotten traction. :sigh:

 

[amrnuke]

I kept hearing originally that today would be the review embargo lift for Zen 2. However, it seems like there would have already been some reviews posted. In the last week, rumors have shifted to the embargo being lifted on release day.

Do we have any official date/time? Clearly, it seems to not be today, as they would have almost certainly been released by this time of day.

 

[Shivansps]

"Vega 3" wouldn't be that much smaller than the "Vega 11" on Raven Ridge. The CUs themselves are not that big; most of the die space on a small GPU is taken up by the I/O and fixed function blocks. Of course, an iGPU has an advantage in this regard, since it can share the memory controller with the CPU.

Polaris 10 is a 123mm^2 chip with 16 CUs. Polaris 12 is a 103mm^2 chip with 10 CUs. Since both have the same architecture, same bus width, and same video codec hardware, we can thus calculate that each extra CU only uses about 3.3 square millimeters of die space on GloFo 14nm/12nm. It might be a bit more for Vega, but probably not much. Going down below 8 CUs cuts out a lot of functionality for minimal area savings, so don't expect to see any iGPUs smaller than that.

I think the reason is a I/O limit, rather than space to provide just a basic GPU... also the encoder block was not needed.

I think the reason is the same one of why Ryzen APU are limited to 8x mas PCIe... i belive they are using the pci-e lanes to provide video output.