Mini-news: AMD Zen 2 architecture said to have 16% higher IPC than the original 1st-gen Zen CPUs

[porina]

6 minutes ago, Taf the Ghost said:

AVX512 benchmarking is still in its early days, and there are some wonky results. Mostly it seems like the lower core count Xeons stack up really well against much larger core count CPUs. On 6 channels, 12 cores might be the maximum that is useful. Or maybe even less. As we're lamenting, it's DDR5 tech. Maybe even DDR6.

More ram channels would help. I think even mainstream could use more than 2 now, especially as AMD kicked off the core wars.

 

I'm following the developments of AVX-512 implementation into Prime95. Suffice to say, they're ram bandwidth limited before, so AVX-512 isn't looking likely at giving a boost in those situations. However, for smaller FFT sizes that can run within cache, that could be really sweet. The math code in Prime95 is also used in other prime number finding software, and while it doesn't help the GIMPS project directly, others could see a nice boost from its implementation even without the extra ram bandwidth.

 

Think Intel were looking at rolling AVX-512 into consumer CPUs too, although I'm assuming this depends on their 10nm deployment so I'm not holding my breath.

 

6 minutes ago, Taf the Ghost said:

at 7nm for AMD, I would expect at least to match Intel for AVX2 units. (Probably 1 dedicate and 2 fuseable 128bit AVX units.) 

If you're right on that, I look forward to it.

[Mira Yurizaki]

11 minutes ago, Okjoek said:

Do you think a better VRAM solution could ever happen for the APUs? I'm kinda curious what difference it would make in games for even the current APUs to have access to some kind of GDDR or HBM whether it be built into the APU package or as a module on the motherboard. Or at the very least maybe Zen 3 will use a new DDR5 system RAM with some theoretical better bandwidth.

Including VRAM in the APU would drive the cost up, which to me doesn't make sense at that point given what market APUs are targeting. An add-on module wouldn't really take off either because it's a specialty item and the cost would be shifted to motherboard manufacturers to support it.

 

Maybe there could be an option for a "high-end" APU, but at that point, you may as well have purchased a cheap video card and a standalone CPU.

[DrMacintosh]

3 hours ago, Trixanity said:

There is a good chance that AMD will split the consumer and server chips this time.

No they won’t. The same architecture scaled all the way up and all the way down is exactly why AMD is doing great in the current market. 

[cj09beira]

1 hour ago, M.Yurizaki said:

Including VRAM in the APU would drive the cost up, which to me doesn't make sense at that point given what market APUs are targeting. An add-on module wouldn't really take off either because it's a specialty item and the cost would be shifted to motherboard manufacturers to support it.

 

Maybe there could be an option for a "high-end" APU, but at that point, you may as well have purchased a cheap video card and a standalone CPU.

thats the thing though, with having them fused prices could be a lot cheaper, you dont need 2 coolers 2 vrms and 2 motherboards/pcbs, it would be a more limited solution in terms of upgrading it but it might be a good option for quite a lot of people, the way i see it is as a apu with a low cost hbm module next to it (the low cost hbm modules have few pins so they end up not needing an interposer)

[GoldenLag]

"mini News"........ Zen 2 architeccture

 

 

 

*instantly 3 pages

[Trixanity]

1 hour ago, DrMacintosh said:

No they won’t. The same architecture scaled all the way up and all the way down is exactly why AMD is doing great in the current market. 

Scalable, yes, but only to a point. You're limited by the very same scaling. Consumer and server markets continue to diverge more and more. Eventually they'll split it. It may be next year, the year after or in five years. It's hard to extract more performance without adding more units within a core, adding more cores or adding more specialized units. There's also stuff like CCX design which is what their scalability hinges on (along with the interconnect) where it could quickly become a problem for consumer workloads but not server workloads. There's also the rumored paradigm shift of chiplets with a massive LLC chiplet surrounded by CCXs. Whether that actually becomes a thing remains to be seen but if AMD only consider the production costs they'll run into performance problems eventually. They're already facing hard-to-solve problems with current generation chips. It'll grow bigger the more you want it to scale unless AMD comes up with a solution. Monolithic is better in most regards for a reason.

 

TL;DR eventually it doesn't make sense to make the same chip across the board but that doesn't mean there won't be any scaling. Making three chips instead of two isn't the end of the world economically. Especially with the growth AMD is aiming for.

[GoldenLag]

2 minutes ago, Trixanity said:

Scalable, yes, but only to a point.

what point? essentially each chips has its own memmory controller that wnats 2 sticks of memmory, the more chiplets the more memmory. essentially AMD can scale forever. you do run into some issues with memmory intensive tasks. if that was what you were reffereing to. 

 

4 minutes ago, Trixanity said:

Consumer and server markets continue to diverge more and more. Eventually they'll split it

eventually, yes, but as far as we know. AMD will be producing the same base CCX design for all plattforms, the difference then being the infrastructure on the substrate.

 

5 minutes ago, Trixanity said:

There's also stuff like CCX design which is what their scalability hinges on (along with the interconnect) where it' could quickly become a problem for consumer workloads but not server workloads.

hence low consumer CCX counts. its the same base tech, hence why servers are getting the best chips while us consumers are getting the handmedowns

6 minutes ago, Trixanity said:

They're already facing hard-to-solve problems with current generation chips.

enlighten me on these problems, i do not have the greatest insight on what AMD engineers are currently struggling with. with the exception of latency and bandwidht on the interconnect.

7 minutes ago, Trixanity said:

Monolithic is better in most regards for a reason.

monolithic is great as long as you dont need high yields on very large chips. which is what chiplets are good for. 

[yian88]

Good its really good if its average IPC, if its IPC in Cinebench or some synthetic weird test then it means nothing, either way they probably tested an engineering sample.

We will have to wait for final optimized silicon to know, as always.

I recently bought a ryzen 5 2600 and im not going to upgrade 2019, but i will upgrade 2020/2021 to Ryzen 4000, built on Zen2+ or Zen3  7nm/7nm+ with 8+ cores, for now these 6c/12T is more than enough.

 

I hope by 2020 they can reach 4.8ghz on most chips with Zen2+/Zen3, i dont really like these low frequencies of 4 -4.2ghz max at high voltages, 1.35-1.45 volts its waay too much.

[Mira Yurizaki]

Just now, GoldenLag said:

what point? essentially each chips has its own memmory controller that wnats 2 sticks of memmory, the more chiplets the more memmory. essentially AMD can scale forever. you do run into some issues with memmory intensive tasks. if that was what you were reffereing to. 

Given AMD's current way of designing the packages, it isn't scalable. Every Zeppelin die is connected to each other to limit the number of hops required to 1 (see: https://en.wikichip.org/wiki/File:amd_epyc_interconnect.png). You can't really do this forever:

 

And currently there's a problem on Threadripper 2 in that two of the Zeppelin dies aren't connected to memory. This makes the NUMA problem that the first Threadripper had even worse.

 

23 minutes ago, cj09beira said:

thats the thing though, with having them fused prices could be a lot cheaper, you dont need 2 coolers 2 vrms and 2 motherboards/pcbs, it would be a more limited solution in terms of upgrading it but it might be a good option for quite a lot of people, the way i see it is as a apu with a low cost hbm module next to it (the low cost hbm modules have few pins so they end up not needing an interposer)

I forgot the other reason why an external module wouldn't really work: you need to tie the pins back to the CPU. Socket AM4 already looks like a full socket pinout wise. So unless you want to do another system platform to accommodate external modules, an external solution won't work.

 

Also I don't have high confidence in the reliability of availability of HBM based packages.

[GoldenLag]

1 minute ago, M.Yurizaki said:

And currently there's a problem on Threadripper 2 in that two of the Zeppelin dies aren't connected to memory. This makes the NUMA problem that the first Threadripper had even worse.

yeah, threadripper 2 essentially has 16 cores + 16 Compute cores who do not have direct access to memmory. certainly hinders performance greatly. 

3 minutes ago, M.Yurizaki said:

ven AMD's current way of designing the packages, it isn't scalable. Every Zeppelin die is connected to each other to limit the number of hops required to 1 (see: https://en.wikichip.org/wiki/File:amd_epyc_interconnect.png). You can't really do this forever:

so essentially AMD can maybe have a couple of leach dies on their current dies, but cant expand to say a 4x4 square of dies. hence the rumoured 9 die config with the 8 CPU dies (7nm?) and 1 cache/interconnect die (12?/14?nm)

 

makes sence, thanks. 

[cj09beira]

13 minutes ago, M.Yurizaki said:

Given AMD's current way of designing the packages, it isn't scalable. Every Zeppelin die is connected to each other to limit the number of hops required to 1 (see: https://en.wikichip.org/wiki/File:amd_epyc_interconnect.png). You can't really do this forever:

 

And currently there's a problem on Threadripper 2 in that two of the Zeppelin dies aren't connected to memory. This makes the NUMA problem that the first Threadripper had even worse.

 

I forgot the other reason why an external module wouldn't really work: you need to tie the pins back to the CPU. Socket AM4 already looks like a full socket pinout wise. So unless you want to do another system platform to accommodate external modules, an external solution won't work.

 

Also I don't have high confidence in the reliability of availability of HBM based packages.

the only thing limiting hbm supply is lack volume, a volume product is exactly what hbm needs 

[TetraSky]

Can't wait to finally upgrade! as long as it is priced adequately. 

[Linus_torvalds]

So they improved zen by 16%. Considering intel was improving by 3-5% each year after ivy bridge they jumped from haswell to Kaby lake (Yeah Kaby lake because skylake and Kaby lake are THE SAME!)

[D13H4RD]

Was expecting a 10-15% IPC jump, so if true, it's shaping up to be a very solid improvement. 

 

One thing though. What about clockspeeds? Arguably the one reason why Intel CPUs perform the way they do in gaming is their ability to achieve really high stable clocks. 

[Trixanity]

2 hours ago, GoldenLag said:

what point? essentially each chips has its own memmory controller that wnats 2 sticks of memmory, the more chiplets the more memmory. essentially AMD can scale forever. you do run into some issues with memmory intensive tasks. if that was what you were reffereing to. 

 

eventually, yes, but as far as we know. AMD will be producing the same base CCX design for all plattforms, the difference then being the infrastructure on the substrate.

 

hence low consumer CCX counts. its the same base tech, hence why servers are getting the best chips while us consumers are getting the handmedowns

enlighten me on these problems, i do not have the greatest insight on what AMD engineers are currently struggling with. with the exception of latency and bandwidht on the interconnect.

monolithic is great as long as you dont need high yields on very large chips. which is what chiplets are good for. 

Latency, latency and latency. 

 

How many RAM slots do you expect to feed each chip if we go beyond 4 chips? Okay, so we double the size of the CCX instead. Now we might get added core-to-core latency and it now becomes unpredictable meaning you'll get poor performance in many workloads. We already had enough trouble with CCX-to-CCX latency.

 

We don't know if AMD wants to maintain a common CCX design or if they want to differentiate based on market. One size fits all is good until it isn't.

 

Latency and bandwidth are two very big issues (core-to-core, CCX-to-CCX, cache and memory). Cache structure is as well. AVX is as well (why add tons of silicon to a consumer chip that won't use it?). Injecting some steroids into FP performance with the space afforded by 7nm isn't that farfetched either.

 

I think I already explained the pitfalls of monolithic.

[Morgan MLGman]

4 hours ago, Okjoek said:

Do you think a better VRAM solution could ever happen for the APUs? I'm kinda curious what difference it would make in games for even the current APUs to have access to some kind of GDDR or HBM whether it be built into the APU package or as a module on the motherboard. Or at the very least maybe Zen 3 will use a new DDR5 system RAM with some theoretical better bandwidth.

 

Well, it kinda already happened with the Kaby Lake-G CPUs that have a Vega GPU on it (with some Polaris similarities), it has HBM (aka the only way to currently give iGPUs dedicaded VRAM due to its small footprint). And it really improves performance.

4 hours ago, Okjoek said:

If I had to upgrade my system right now though I'd try to keep my RX 460 because it's still awesome and has 4GB of VRAM

 

I can't imagine how are you satisfied with the RX 460 for gaming, I recently got a laptop with an RX 560X 4GB in it, which is exactly the desktop RX 560 that boosts to 1275MHz (like the desktop reference design) and it's not too fast IMO Maybe I'm kind of biased because I have a 1080 in my main rig, but I can't personally imagine having lower performance than an RX 580 for a primary gaming PC

3 hours ago, GoldenLag said:

"mini News"........ Zen 2 architeccture

 

 

 

*instantly 3 pages

Well, it's mini-news because it's nothing confirmed, just something I found online that I decided was worth discussing

 

[Stefan Payne]

7 hours ago, Morgan MLGman said:

My opinion: If this information proves to be true, AMD's next-gen consumer CPUs based on the Zen 2 architecture may be even more competitive than the 1st or 2nd generation, if this IPC gain is paired with lower power consumption due to the 7nm technological process used, and potentially higher clock speeds achieved, they might equal with Intel in terms of single-threaded performance. Which is the only thing Ryzen lacks when compared to higher-end Intel chips.

Your thoughts on this rumour?

The Real Problem for Intel will be in the Server/Workstation Market.

If it is true, it means that its ~13% higher IPC in this case + another ~22% percent higher clockrates. Without increasing the power consumption. 

 

That is a very bad situation for the blue side, especially since there are news about new AMD Super Computers, for example the Haas F1 team got a new cray with AMD EPYC Processors.

 

And with Rome we don't even know if the power consumption is the same or if it even could be lowered!

 

With all that going on right now, there are very bad times coming towards the blue team - something comparable to the Prescott era of CPUs...

And then there is also the Intel CPU Shortage right now as well...

 

And I doubt Intel will survive another situation like in the late 90s/early 2000s, when they paid off or even blackmailed their "partners" to not use AMD Products - wich lead to some Boards sold in a white box!!
Most here might not remember that there was once the ASUS K7M for the AMD Plattform that came without any ASUS Branding in a literally white Box.

[Stefan Payne]

4 hours ago, DrMacintosh said:

No they won’t. The same architecture scaled all the way up and all the way down is exactly why AMD is doing great in the current market. 

I kinda agree with you here.

From what we know right now is that AMD might throw out all the I/O Stuff and transfer it into a seperate "I/O Chip" that essentially does the PCIe, S-ATA and Memory Stuff and also connects to a couple of Core Chiplets that only have the CPU Cores and IF to connect to this System Controller and other dies.

 

The reason for that is that 7nm is freaking expensive and the I/O stuff doesn't shrink at all with the 7nm process so it makes sense to not integrate that into the CPU... 

 

And for that, we might have 3 dies:

a) the CPU Core die, that stays the same between Ryzen and EPYC

b) the I/O Core where there will be 2 different ones: One big one for the TR4 plattform with 128 PCIe Lanes, eight 64bit DDR4 Memory Channels, 8 S-ATA, USB and other stuff and the small Ryzen one with only 32 PCIe lanes, 2 S-ATA and so on.
That will be connected via IF to the Cores.

 

So at least the main cores they can use for both, the "uncores" probably not.

3 hours ago, Trixanity said:

TL;DR eventually it doesn't make sense to make the same chip across the board but that doesn't mean there won't be any scaling. Making three chips instead of two isn't the end of the world economically. Especially with the growth AMD is aiming for.

Yeah, but you'd want to share as much between the plattforms as possible.

Because if someone messed something up and you don't have any "Server Dies", you can put the "Ryzen Dies" onto the right package, set the right information and you're good to go.

And vice versa, if you made too many server chips, you can still relabel them and sell it as a Threadripper.

Or, if you didn't put it on a package, normal Ryzen...

2 hours ago, M.Yurizaki said:

Given AMD's current way of designing the packages, it isn't scalable. Every Zeppelin die is connected to each other to limit the number of hops required to 1 (see: https://en.wikichip.org/wiki/File:amd_epyc_interconnect.png). You can't really do this forever:

True but there are also rumors about a 10 Core Ryzen 2800x floating around.

So they might eventually increase the Size of a CCX...

 

2 hours ago, M.Yurizaki said:

And currently there's a problem on Threadripper 2 in that two of the Zeppelin dies aren't connected to memory. This makes the NUMA problem that the first Threadripper had even worse.

Thing is:
That's old. Very old. So old that Windows VISTA at the latest should have been patched for a Memoryless CPU Core and know about it.

People seem to have forgotten the MSI K8M Master lineup...

That was back in 2003, when there was no Socket 939 and all there was was Socket 940 wich required (IIRC) Registred ECC memory and the Single Channel S754...

 

2 hours ago, M.Yurizaki said:

I forgot the other reason why an external module wouldn't really work: you need to tie the pins back to the CPU. Socket AM4 already looks like a full socket pinout wise. So unless you want to do another system platform to accommodate external modules, an external solution won't work.

 

Also I don't have high confidence in the reliability of availability of HBM based packages.

Only if you do it off package...

[pas008]

51 minutes ago, Stefan Payne said:

The Real Problem for Intel will be in the Server/Workstation Market.

If it is true, it means that its ~13% higher IPC in this case + another ~22% percent higher clockrates. Without increasing the power consumption. 

 

That is a very bad situation for the blue side, especially since there are news about new AMD Super Computers, for example the Haas F1 team got a new cray with AMD EPYC Processors.

 

And with Rome we don't even know if the power consumption is the same or if it even could be lowered!

 

With all that going on right now, there are very bad times coming towards the blue team - something comparable to the Prescott era of CPUs...

And then there is also the Intel CPU Shortage right now as well...

 

And I doubt Intel will survive another situation like in the late 90s/early 2000s, when they paid off or even blackmailed their "partners" to not use AMD Products - wich lead to some Boards sold in a white box!!
Most here might not remember that there was once the ASUS K7M for the AMD Plattform that came without any ASUS Branding in a literally white Box.

this is where amd can hold mainstream for awhile

 

and on intel surviving dont forget intel can do mesh emib(might not be able to get same power consumption though lol) plus they have many other areas besides cpus they have been dabbling everywhere

think it would take along time for them to die

[leadeater]

12 hours ago, porina said:

Good question. AMD chose not to compete with Intel in that area with Ryzen to date, but Intel are not resting and are pushing AVX-512 which could put them up to 4x IPC of Ryzen. If AMD were to match Intel consumer CPU performance, I'd expect to see closer to 100% improvement (depending on workload), not 16%.

 

If I had to guess, AVX is not where I'd look at. Maybe more towards further memory controller, cache and internal connectivity optimisations. Scientific tasks can depend more on those factors than consumer workloads.

I've actually seen a few tests where AVX-512 was slower than AVX2 on the Skylake-SP CPU tested, came down to being a very memory bandwidth demanding workload so only took a few cores doing AVX-512 to completely use all the bandwidth there was leaving a lot of cores unable to do much. DDR5 really needs to come quickly now.

[Okjoek]

5 hours ago, Morgan MLGman said:

I can't imagine how are you satisfied with the RX 460 for gaming, I recently got a laptop with an RX 560X 4GB in it, which is exactly the desktop RX 560 that boosts to 1275MHz (like the desktop reference design) and it's not too fast IMO Maybe I'm kind of biased because I have a 1080 in my main rig, but I can't personally imagine having lower performance than an RX 580 for a primary gaming PC

It's all about application.

 

Rocket League runs fine,

Minecraft bedrock edition runs fine,

Planetside 2 runs fine,

Old School Runescape runs fine,

World of Tanks runs fine,

Stardew Valley runs fine,

Divinity Original Sin 2 runs fine,

 

 

Planetside 2 is one of my day-to-day favorites and one of the most demanding on this list. I run on custom, mostly low settings, but that would be the case no matter what GPU I'm using because it's a PVP shooter and the visuals be damned if the gameplay is good. I've under-clocked my RX 460 to 650 MHz and I was still able to have an enjoyable Planetside 2 experience. (It's a very CPU-bound game).

 

That's not to say if there was some groundbreaking new game title that my system couldn't run I would be unwilling to upgrade.

[ravenshrike]

3 minutes ago, Okjoek said:

but that would be the case no matter what GPU I'm using because it's a PVP shooter and the visuals be damned if the gameplay is good.

But what about the fire reflected off shiny surfaces?

[Okjoek]

1 minute ago, ravenshrike said:

But what about the fire reflected off shiny surfaces?

What? I don't understand the question.

 

[Taf the Ghost]

11 hours ago, porina said:

More ram channels would help. I think even mainstream could use more than 2 now, especially as AMD kicked off the core wars.

 

I'm following the developments of AVX-512 implementation into Prime95. Suffice to say, they're ram bandwidth limited before, so AVX-512 isn't looking likely at giving a boost in those situations. However, for smaller FFT sizes that can run within cache, that could be really sweet. The math code in Prime95 is also used in other prime number finding software, and while it doesn't help the GIMPS project directly, others could see a nice boost from its implementation even without the extra ram bandwidth.

 

Think Intel were looking at rolling AVX-512 into consumer CPUs too, although I'm assuming this depends on their 10nm deployment so I'm not holding my breath.

 

If you're right on that, I look forward to it.

Cannonlake, the doomed 10nm Skylake shrink, does have a single AVX512 unit per core. So it was supposed to come to desktop. We'll see it with Icelake in 2020, but the AVX512 unit actually isn't the full instruction set just yet. That appears to be either Tigerlake or the one after that. Another reason AMD isn't also going to be adding it too soon.

[Taf the Ghost]

10 hours ago, cj09beira said:

thats the thing though, with having them fused prices could be a lot cheaper, you dont need 2 coolers 2 vrms and 2 motherboards/pcbs, it would be a more limited solution in terms of upgrading it but it might be a good option for quite a lot of people, the way i see it is as a apu with a low cost hbm module next to it (the low cost hbm modules have few pins so they end up not needing an interposer)

APUs with "easy" to attach HBM would probably be for Apple.