AMD 12 core 24 thread Zen2 engineering sample benchmark leaked

[GoldenLag]

Just now, Taf the Ghost said:

It *is* a PCIe generation change, so some hiccups were always on the table.

AMD seems to have had a couple of hiccups. 

 

AdoredTV mentioned a error during the taping process of Navi. a late issue they discovered. hopefully they dont rush it. 

[Taf the Ghost]

1 minute ago, GoldenLag said:

AMD seems to have had a couple of hiccups. 

 

AdoredTV mentioned a error during the taping process of Navi. a late issue they discovered. hopefully they dont rush it. 

It's an AMD GPU. RTG has had significant problems since AMD bought them. AMD had to respin Zen a few extra times as well, though AMD was so comfortable with the chips after the first respins that they practically launched Ryzen 1st gen with ES models. That's why the motherboard manufacturers were scrambling. 

 

Navi seems to have been ahead of schedule then hit a late snag, which is why the Radeon 7 was on & off the schedule in the rumor mill all year.

[CarlBar]

Regarding DDR5, again AMD says they need a new socket, AMD also says that isn't happening till Zen 4. Though they have also indicated it may not be DDR5, (what with persistent low latency memory like Optane DIMM's and their competitors coming out, HBM, and it's competitor,s and DDR5 on the field no one's quite sure what the memory landscape is going to look like in another couple of generations).

[Taf the Ghost]

2 hours ago, CarlBar said:

Regarding DDR5, again AMD says they need a new socket, AMD also says that isn't happening till Zen 4. Though they have also indicated it may not be DDR5, (what with persistent low latency memory like Optane DIMM's and their competitors coming out, HBM, and it's competitor,s and DDR5 on the field no one's quite sure what the memory landscape is going to look like in another couple of generations).

Everyone will need new sockets for DDR5, but I'm not sure AMD has said that it won't happen until Zen 4. Since the IMC is on the I/O die, AMD can move to a new memory standard whenever it best suits them.  However, the industry is going to slow-roll out to DDR5. Mostly because they haven't maxed out DDR4 yet.

[Belgarathian]

2 hours ago, CarlBar said:

Regarding DDR5, again AMD says they need a new socket, AMD also says that isn't happening till Zen 4. Though they have also indicated it may not be DDR5, (what with persistent low latency memory like Optane DIMM's and their competitors coming out, HBM, and it's competitor,s and DDR5 on the field no one's quite sure what the memory landscape is going to look like in another couple of generations).

Somewhere in AMD they'll have an engineering sample of a SoC using Zen + Vega on HBM2.

[CarlBar]

5 hours ago, Taf the Ghost said:

Everyone will need new sockets for DDR5, but I'm not sure AMD has said that it won't happen until Zen 4. Since the IMC is on the I/O die, AMD can move to a new memory standard whenever it best suits them.  However, the industry is going to slow-roll out to DDR5. Mostly because they haven't maxed out DDR4 yet.

 

They've specifically said they won't go to a new socket till Zen 4. I'd have to dig but i think it was during their Rome reveal event.

[Taf the Ghost]

10 minutes ago, CarlBar said:

 

They've specifically said they won't go to a new socket till Zen 4. I'd have to dig but i think it was during their Rome reveal event.

Naples, Rome and Milan (Zen 1-3) will be on the same platform for Server. Genoa (Zen4 server) will be on a new platform, but that should have PCIe 5.0 as well.

 

On Desktop, AMD will "support AM4 with new processors" through 2020. That doesn't actually mean 2020's Ryzen 4th Gen is explicitly on AM4 though. AMD has been hedging on what they'll do with the desktop socket, as DDR5 isn't actually finalized. (There are working DDR5 chips, IP and controllers though, so they'll announce a final standard around launch.) Right now it looks like Zen2 + Refresh will cover 2019/2020, then likely Zen3 in 2021 will be the last of the DDR4 models. It's reading between the lines of a random Jim Anderson aside, before he left to take a CEO position at another company, that they had their Roadmap out to 2022. That's likely to be Zen3 Refresh and DDR5 on AM5.

 

Now, AMD could move off AM4 in 2021 because they need a physically larger space for the dies, but I'd be rather surprised if AM4 dies before DDR5 lands. And that's a bit of a ways off. It doesn't look like Intel, in their Cooper Lake/Icelake generation, is going for DDR5, so Servers won't move there in mass until at least the 2021/2022 range.

[Kukielka]

2 hours ago, Taf the Ghost said:

Naples, Rome and Milan (Zen 1-3) will be on the same platform for Server. Genoa (Zen4 server) will be on a new platform, but that should have PCIe 5.0 as well.

 

On Desktop, AMD will "support AM4 with new processors" through 2020. That doesn't actually mean 2020's Ryzen 4th Gen is explicitly on AM4 though. AMD has been hedging on what they'll do with the desktop socket, as DDR5 isn't actually finalized. (There are working DDR5 chips, IP and controllers though, so they'll announce a final standard around launch.) Right now it looks like Zen2 + Refresh will cover 2019/2020, then likely Zen3 in 2021 will be the last of the DDR4 models. It's reading between the lines of a random Jim Anderson aside, before he left to take a CEO position at another company, that they had their Roadmap out to 2022. That's likely to be Zen3 Refresh and DDR5 on AM5.

 

Now, AMD could move off AM4 in 2021 because they need a physically larger space for the dies, but I'd be rather surprised if AM4 dies before DDR5 lands. And that's a bit of a ways off. It doesn't look like Intel, in their Cooper Lake/Icelake generation, is going for DDR5, so Servers won't move there in mass until at least the 2021/2022 range.

I kinda expect the I/O Die to shrink until then since it's still at 14nm, making space for a third possible chiplet.

This would mean that they would not require more space until Zen4.

(As long as 3 chiplets + I/O is enough for AMD 

[Taf the Ghost]

6 minutes ago, Kukielka said:

I kinda expect the I/O Die to shrink until then since it's still at 14nm, making space for a third possible chiplet.

This would mean that they would not require more space until Zen4.

(As long as 3 chiplets + I/O is enough for AMD

AMD has space to make the I/O die quite a bit bigger, actually, and AMD's contract to produce a significant amount of wafers (over 1 Billion USD, I believe) with GloFo should end in 2020. The 2021 product cycle will be Zen3 for AMD, with the I/O die more than likely moving to a different node. Given that the I/O die has different requirements, we might see AMD use a mobile process for it. Samsung's 8nm, for instance, might make the most sense for that type of product. We won't know for a couple of years.

[leadeater]

12 minutes ago, Kukielka said:

I kinda expect the I/O Die to shrink until then since it's still at 14nm, making space for a third possible chiplet.

This would mean that they would not require more space until Zen4.

(As long as 3 chiplets + I/O is enough for AMD

The I/O die may not, signal logic doesn't benefit as much by node shrinks and can also regress, GloFlo has an optimized 14nm (12nm) which wasn't used for the I/O die.

[Stefan Payne]

24 minutes ago, Kukielka said:

I kinda expect the I/O Die to shrink until then since it's still at 14nm, making space for a third possible chiplet.

I don't because you can't shrink it much because of the current/signal strenth needed to drive the signals.

THAT is why AMD choose the Chiplet design and went for 14nm, because its widely available, rather cheap and offer no significant disadvantage.

 

Quote

This would mean that they would not require more space until Zen4.

(As long as 3 chiplets + I/O is enough for AMD

Yes but due to physical limitations, the I/O Die might be on 14nm forever as there is no way to shrink the PHY of the CHip noteworthy. THe logic might be shrinkable but most of the I/O is PHYs -> PCIe, S-ATA USB.

You can already see that happening with the 28nm -> 14nm Shrink of some chips, that there are areas on it that were hardly any smaller due to this circumstance...

[Taf the Ghost]

@Kukielka

 

Since I mentioned the WSA, AMD has amended it again.

 

Quote

AMD on Tuesday said it had amended its wafer supply agreement with GlobalFoundries. Under the terms of the new deal the two companies agreed about prices and volumes of wafers that AMD will purchase from the U.S.-based foundry till at least 2021. Meanwhile, since the original WSA is continuing through March 1, 2024, the companies will have to continue talks.

I'm going to go with "expect I/O dies on GloFo 14nm through 2021", haha. @leadeater , I hadn't actually seen that the WSA lasts until 2024. Which, amazingly enough, is exactly the same period AMD will be producing Polaris GPUs until. (AMD appears to be selling a lot of them into the embedded device market right now.)

[Belgarathian]

Lisa has basically confirmed that there will be up to 16c/32t Ryzen processors in this interview:

 

https://www.anandtech.com/show/13909/ces-2019-amd-ceo-dr-lisa-su

 

Quote

Q: AMD was showing quoted an 8 core 16 thread set, is that generally an indication of where Ryzen 3rd gen is going to be at?

 

LS: If you look at the evolution of Ryzen, we've always had an advantage in core count and so in this particular case we wanted to show sort of a head-to-head comparison of eight cores and sixteen threads [against the same]. Some people may have noticed on the package that there's some extra room...

 

[Stefan Payne]

Yeah, 16c/32t will come to AM4, the  only question is: When.

THey might release it later this year for example and start with up to 12 cores in Summer...

[Kukielka]

@Taf the Ghost

Damn you're spot on then! Well I'll be darned!

I guess we will indeed need some more space on that cpu

[Stefan Payne]

10 hours ago, Taf the Ghost said:

I'm going to go with "expect I/O dies on GloFo 14nm through 2021", haha. @leadeater 

They probably got a good deal on it.

But just look at the Memory Controller of for example Tahiti, wich was optimized for Speeds more than Hawaii or Pitcairn and compare that with Polaris. How much smaller did it really get? Or just look at a Die shot of Fiji and compare that with VEGA 20..

 

So that isn't really an issue, on the contrary...

 

10 hours ago, Taf the Ghost said:

, I hadn't actually seen that the WSA lasts until 2024. Which, amazingly enough, is exactly the same period AMD will be producing Polaris GPUs until. (AMD appears to be selling a lot of them into the embedded device market right now.)

I don't know if they'd sell it that long. And the Console Chips all come from TSMC (with probably some weird contract that allow Sony/M$ to produce/order it in AMD's name and also pay for them and only give AMD a royality fee, that's how I'd think that might work good thing for AMD is that it's an enormously low risk product with some revenue).

 

But that they're going to the Embedded market makes sense, especially since AMD supports 3rd Party Driver development. That might be a Bonus to that market. So if they need something special in the GPU Driver, they can do it themselves. ANd that might also be the way AMD handles it. Just the Hardware and software as is, if they need software additions, they can do whatever they want with the code).

 

But I'd suspect that they use the smaller Polaris Chip (ie Radeon 560) for this and not the big Polaris 10/20/30 or so...

[Taf the Ghost]

@Stefan Payne because much of both Intel's & AMD's portfolios aren't public facing, people don't realize half of the stuff they produce for. Polaris parts are going to be showing up in Medical Devices for years to come. Once they bump the RX550 GPU to handle HDMI 2.1, AMD is going to be selling those into devices for ages. 

[leadeater]

58 minutes ago, Taf the Ghost said:

@Stefan Payne because much of both Intel's & AMD's portfolios aren't public facing, people don't realize half of the stuff they produce for. Polaris parts are going to be showing up in Medical Devices for years to come. Once they bump the RX550 GPU to handle HDMI 2.1, AMD is going to be selling those into devices for ages. 

And some of which will be sold under Matrox since they use AMD GPUs on their graphics cards and they really only sell to medical, production audio visual and grid wall displays/monitor arrays.

[cj09beira]

1 hour ago, Stefan Payne said:

But I'd suspect that they use the smaller Polaris Chip (ie Radeon 560) for this and not the big Polaris 10/20/30 or so...

they do have quite a bit of usage for the hiher performance gpus, in things like high resolution casino machines

[cj09beira]

5 minutes ago, leadeater said:

And some of which will be sold under Matrox since they use AMD GPUs on their graphics cards and they really only sell to medical, production audio visual and grid wall displays/monitor arrays.

thats quite interesting didn't know there was another company selling amd gpus as graphics cards 

 

8 minutes ago, leadeater said:

And some of which will be sold under Matrox since they use AMD GPUs on their graphics cards and they really only sell to medical, production audio visual and grid wall displays/monitor arrays.

this might be part of the reason why amd doesn't have as much color compression as nvidia, the medical guys would probably not buy them if it was the case

[leadeater]

3 minutes ago, cj09beira said:

this might be part of the reason why amd doesn't have as much color compression as nvidia, the medical guys would probably not buy them if it was the case

Colour compression and memory compression are different, there's more Nvidia GPUs in the medical field than AMD. Edit: Colour compression would come under the bigger title of memory compression.

[Taf the Ghost]

6 minutes ago, leadeater said:

And some of which will be sold under Matrox since they use AMD GPUs on their graphics cards and they really only sell to medical, production audio visual and grid wall displays/monitor arrays.

I was going to say, "I believe AMD only recently stopped selling Geode devices, a design from the early 2000s."

 

https://en.wikipedia.org/wiki/Geode_(processor)

 

Turns out that isn't even correct. They've added *another* extension to the product.

 

https://www.pressebox.com/pressrelease/congatec-ag/congatec-partners-with-AMD-for-long-time-support-of-AMD-Geode-processors/boxid/888930

 

Quote

congatec – a leading technology company for embedded computer modules, single board computers and embedded design and manufacturing services – and AMD have teamed together to provide extended life cycle support for one of the world's longest-serving x86 processors. The result is that AMD Geode™ processor boards from congatec will have planned availability until the end of 2021.

 

[Taf the Ghost]

2 minutes ago, leadeater said:

Colour compression and memory compression are different, there's more Nvidia GPUs in the medical field than AMD.

You would also just turn that off for a Medical Device. They aren't pushing high memory bandwidth situations anyway.

[Taf the Ghost]

This reminds me of a TSMC earnings report where 3% of their revenue is still from Fabs at 250 nm and over. The 250nm Node was from 1997. Granted, I think TI has some nodes from the 1980s that are still active, possibly even the 1970s. Some things just don't scale.

[cj09beira]

@Taf the Ghost @leadeater

what do you guys think of that leak about arcturus being on the new xbox,

i dont trust it too much, i dont think amd would have anything based on it working or even specs defined