Wednesday, July 25th 2018

Rumor: AMD's Zen 2, 7 nm Chips to Feature 10-15% IPC Uplift, Revised 8-core per CCX Design

A post via Chiphell makes some substantial claims on AMD's upcoming Zen 2 microarchitecture, built on the 7 nm process. AMD has definitely won the core-count war once again (albeit with a much more decisive blow to Intel's dominance than with Bulldozer), but the IPC battle has been an uphill one against Intel's slow, but sure, improvement in that area over the years. AMD did say, at the time they introduced the Zen architecture, that they had a solid understanding on Zen's choke points and its improveable bits and pieces - and took it to heart to deliver just that.
A 10 to 15% improvement should bring AMD close to Intel's Kaby Lake solutions in pure per-core, single-thread performance - with the possibility for higher performance in multithreaded workloads even at the same clocks and core counts, since AMD's SMT implementation and per-core communication seems to best that of even Intel's Coffee Lake chips. Add to that the sure clock increases AMD will bring to their Zen 2 parts - you can bet the power consumption and top frequency benefits for the 7 nm manufacturing process (some estimated 1.6X logic density, ~20% speed improvement, and ~40% power reduction compared to TSMC's 10 nm process) won't go to waste when they can be used for an even greater performance uplift.
Add to that the possibility -also posted via the Chiphell thread) of a per-CCX core-count increase of 8 cores per CCX (with 16-core parts being available for the consumer products) and a purely hypothetical claim for the performance crown now seems not only plausible, but likely. Sources: Chiphell, TSMC
Add your own comment

108 Comments on Rumor: AMD's Zen 2, 7 nm Chips to Feature 10-15% IPC Uplift, Revised 8-core per CCX Design

#76
ratirt
Prima.Vera said:
Wait. What? You mean having an 8c/16t CPU with 5Ghz it's not even trying? You're joking right? :laugh::laugh:
I wan to see a simmilar AMD CPU going on stock air cooler on the same freq without any o.c.
Air cooled you say? I'm sure those pipes there were blowing air. Plus the 1600Watt PS that could run almost my entire house devices. What a bull crap on a barn floor dude. Start from 8 minute and go from there. <div class="youtube-embed" data-id="-Zp_iXYl4IQ"><img src="https://i.ytimg.com/vi/-Zp_iXYl4IQ/hqdefault.jpg" /><div class="youtube-play"></div><a href="https://www.youtube.com/watch?v=-Zp_iXYl4IQ" target="_blank" class="youtube-title"></a></div>
Posted on Reply
#77
Xajel
Regardless if these leaks are true or exaggerated. AMD is on the right track. not like Intel with their new 8C/8C without hyperthreading.. again if the leaks are true.
Posted on Reply
#78
cucker tarlson
That 8c/8t is the best thing that could happen for gaming. Look how stellar 8600k performs against 8700k, now add +25% more cores and +33% more cache.
Posted on Reply
#79
Vayra86
Prima.Vera said:
Wait. What? You mean having an 8c/16t CPU with 5Ghz it's not even trying? You're joking right? :laugh::laugh:
I wan to see a simmilar AMD CPU going on stock air cooler on the same freq without any o.c.
Hold on now.

- Intel K CPUs are not supplied with stock coolers at all, and we know why: its pointless
- Since Kaby Lake Intel is just providing its CPUs with factory OC's. There is a reason the TDP jumped up to 95W and they go way beyond that if you so much as look at the BIOS.
- Intel relies heavily on binning, another aspect AMD has them beat from a design/yield perspective.
Posted on Reply
#80
RejZoR
Vya Domus said:
10% IPC increase isn't the same thing as a generic 10% increase in single thread performance and the fact that the article sort of refers to them as being interchangeable metrics annoys me.

Zen's IPC is already on par with something like Coffee Lake , minus AVX.
Minus AVX ? All ZEN based CPU's support both, AVX and AVX2. Unless you meant something else with that...
Posted on Reply
#81
Vya Domus
RejZoR said:
Minus AVX ? All ZEN based CPU's support both, AVX and AVX2. Unless you meant something else with that...
Zen can do 2x 128 bit FMAs per clock cycle whereas Skylake and upwards can do 2x 256 bit.
Posted on Reply
#82
RejZoR
But what has FMA to do with AVX? Isn't FMA a separate instruction set for a different thing?
Posted on Reply
#83
Vya Domus
RejZoR said:
But what has FMA to do with AVX? Isn't FMA a separate instruction set for a different thing?
FMA is a type of operation not an instruction set , well there is an instruction set named FMA but really it's the same thing.

What I mentioned is the hardware support for instructions like AVX. AVX is an instruction set that can allow for 256 bit wide hardware registers or 2x 128 bit ones. AMD chose the latter , therefore instructions such as AVX are carried out at half the throughput that Skylake can do. In practice it's never really half , the gap is much smaller but it's there.
Posted on Reply
#84
Captain_Tom
For anyone interested, new leaked roadmap from Intel confirms no 10nm till 2020 at the earliest lol:

https://wccftech.com/intel-xeon-scalable-family-roadmap-revealed-points-out-cascade-lake-sp-in-q4-2018-cooper-lake-sp-in-q4-2019-ice-lake-sp-in-1h-2020/


I really can't decide how much I want Intel to be skullcrushed. I want them to really learn their lesson so they never let themselves stagnate like this again, but at the same time it looks like they might be in almost as bad a position as AMD was by 2021 - and I do not want AMD to simply become the new dominating force without competition. At this point it looks like AMD may have 7nm+ out before Intel launches any 10nm product...

Who knows, maybe Intel will become so desperate they actually launch Graphene 1nm chips in 2025 or something...
Posted on Reply
#85
trparky
ToxicTaZ said:
1.45v
Most overclockers recommend never to go past 1.35v on the vCore.
Posted on Reply
#86
efikkan
HTC said:

While the IPC uplift that the rumor claims would be great, due to how Zen architecture works, i'd rather they work on having the ability to use rated RAM speed for high speed RAM, such as the one from the example in the post above from @ToxicTaZ . This alone would greatly increase Zen chips performance across the board. Managing to add like 3% - 5% IPC would be the cherry on top.
Why? Faster memory only helps workloads which are bottlenecked by memory bandwidth. And if you really need more bandwidth, go with HEDT and get twice the bandwidth.

cucker tarlson said:
That 8c/8t is the best thing that could happen for gaming. Look how stellar 8600k performs against 8700k, now add +25% more cores and +33% more cache.
If you mean 8 vs 6 cores, that's 33% more cores.
Just FYI, even though L3 cache is shared, it's still tied to a core. This means L3 cache may only benefit other cores if two or more cores accesses the same data within a very short timespan, it's not like one core can "borrow" the L3 from another.

But yes, it will be a nice product.

Vya Domus said:
Zen can do 2x 128 bit FMAs per clock cycle whereas Skylake and upwards can do 2x 256 bit.
And Skylake-X/-SP can do 2× 512-bit FMA.

RejZoR said:
But what has FMA to do with AVX? Isn't FMA a separate instruction set for a different thing?
FMA3 is an extension to AVX introduced by AMD(!). It performs operations like a × b + c in a single operation, which speeds up many types of calculations like matrix multiplication. But it saves more than you might immediately think, since the single operation saves a whole cycle through the pipeline.
Posted on Reply
#87
HTC
efikkan said:
Why? Faster memory only helps workloads which are bottlenecked by memory bandwidth. And if you really need more bandwidth, go with HEDT and get twice the bandwidth.
Again, due to how Zen architecture works, the faster the memory, the faster infinity fabric operates, which translates into performance increase. This is why faster memory is recommended, when going for a Ryzen system. I'm not only referring to the normal timings: the sub-timings are also important in a Ryzen system.

Unfortunately, even though RAM with 3600+ speed is available and has been for quite some time, Ryzen has been unable to take advantage of this and you'll be hard pressed to find Ryzen systems working with RAM @ speeds 3600+. They exist (i've seen 4000 myself), but they are very uncommon because tinkering with the sub-timings is required in order to get it to work: something you usually don't need in an Intel system because you're normally able to have the rated XMP speeds simply by selecting them in the BIOS.
Posted on Reply
#88
Captain_Tom
HTC said:
Again, due to how Zen architecture works, the faster the memory, the faster infinity fabric operates, which translates into performance increase. This is why faster memory is recommended, when going for a Ryzen system. I'm not only referring to the normal timings: the sub-timings are also important in a Ryzen system.

Unfortunately, even though RAM with 3600+ speed is available and has been for quite some time, Ryzen has been unable to take advantage of this and you'll be hard pressed to find Ryzen systems working with RAM @ speeds 3600+. They exist (i've seen 4000 myself), but they are very uncommon because tinkering with the sub-timings is required in order to get it to work: something you usually don't need in an Intel system because you're normally able to have the rated XMP speeds simply by selecting them in the BIOS.
Yeah the price fixing on DDR couldn't have come at a worse time for AMD. Having said that, I still think it's ridiculous that PC Hardware journalism continues to ignore the fact that 3600MHz+ RAM makes Ryzen better than Intel at gaming (Even if Intel is @5GHz+).


This is an incredible point people are missing: If you care about Price/Perf (at any resolution), Ryzen 2 is the obviously better choice. If you care about just having the highest performance possible, then you should ALSO choose Ryzen with fast memory.
Posted on Reply
#89
GoldenX
I just overclock my ram, it works up to 3200, but that could be my first gen Ryzen's limit.
Posted on Reply
#90
Vayra86
Captain_Tom said:
Yeah the price fixing on DDR couldn't have come at a worse time for AMD. Having said that, I still think it's ridiculous that PC Hardware journalism continues to ignore the fact that 3600MHz+ RAM makes Ryzen better than Intel at gaming (Even if Intel is @5GHz+).


This is an incredible point people are missing: If you care about Price/Perf (at any resolution), Ryzen 2 is the obviously better choice. If you care about just having the highest performance possible, then you should ALSO choose Ryzen with fast memory.
Except for the tiny fact that the cost/performance perspective is lost on the Ryzen option altogether then :) Just as much as it is on Intel with a nice OC, and probably even more so.

The tweaking required to get Ryzen to optimal performance is still too much for most people. So Intel's winning because it offers almost the same but in a much easier way, and often cheaper even on its most expensive mainstream CPUs. Fiddling with timings is advanced user niche territory. An OC on Intel is really not - as much, at least.
Posted on Reply
#91
Captain_Tom
Vayra86 said:
Except for the tiny fact that the cost/performance perspective is lost on the Ryzen option altogether then :) Just as much as it is on Intel with a nice OC, and probably even more so.

The tweaking required to get Ryzen to optimal performance is still too much for most people. So Intel's winning because it offers almost the same but in a much easier way, and often cheaper even on its most expensive mainstream CPUs. Fiddling with timings is advanced user niche territory. An OC on Intel is really not - as much, at least.
1. If you want price/perf - nothing is going to beat a Ryzen build the saves money where possible.

2. If you read what I said, anyone getting an 8700K doesn't care about Price/perf (if they did, they wouldn't even glance at Intel). So the "Extra cost ruins the value" BS is really not true for these people.

3. In terms of Enthusiast Price/Perf, you are still wrong. Ryzen is cheaper AND better performing. Here are 2 different ram kits:

https://www.newegg.com/Product/Product.aspx?Item=N82E16820313771&cm_re=DDR4-_-20-313-771-_-Product

https://www.newegg.com/Product/Product.aspx?Item=N82E16820313618&cm_re=DDR4-_-20-313-618-_-Product

It only costs $70 to go from 2400 to [B]3733 memory! That's less than the price difference required to buy the 8700K (+$50), the AIO Liquid Cooler (+$75), and the more expensive Intel motherboard (+$25). You are actually saving a lot of money (and energy usage) by going with the 2700X + Fast RAM instead of the 8700K + Peasant RAM. Also, it is going to beat the 8700K system at gaming and allow upgrades to 7nm Zen next year.[/B]
Posted on Reply
#92
HTC
Captain_Tom said:
1. If you want price/perf - nothing is going to beat a Ryzen build the saves money where possible.

2. If you read what I said, anyone getting an 8700K doesn't care about Price/perf (if they did, they wouldn't even glance at Intel). So the "Extra cost ruins the value" BS is really not true for these people.

3. In terms of Enthusiast Price/Perf, you are still wrong. Ryzen is cheaper AND better performing. Here are 2 different ram kits:

https://www.newegg.com/Product/Product.aspx?Item=N82E16820313771&cm_re=DDR4-_-20-313-771-_-Product

https://www.newegg.com/Product/Product.aspx?Item=N82E16820313618&cm_re=DDR4-_-20-313-618-_-Product

It only costs $70 to go from 2400 to [B]3733 memory! That's less than the price difference required to buy the 8700K (+$50), the AIO Liquid Cooler (+$75), and the more expensive Intel motherboard (+$25). You are actually saving a lot of money (and energy usage) by going with the 2700X + Fast RAM instead of the 8700K + Peasant RAM. Also, it is going to beat the 8700K system at gaming and allow upgrades to 7nm Zen next year.[/B]
Dude: let's not kid ourselves.

Sure: 2700X catches up to Intel (and passes, on some instances) when overclocked with RAM using good timings in gaming but, as soon as you overclock the 8700K, it's game over, simply because the 8700K overclocks allot more, and even further if you end up de-lidding. I think it's more "common" to find 8700K running @ 5.2 GHz then it is to find a 2700X @ 4.4 GHz and that clock speed difference is simply too massive atm for AMD.

You'll need to spend some cash to achieve a good OC on 2700X, both on RAM and in cooling, which can be as low as you said but can also be quite significantly higher, depending on where you live. The same is true for the 8700K but you're forgetting that those that buy the 8700K also tend to pair it with good RAM which, although doesn't give a gain as big as it does in a Ryzen system, still gives it. This will ofc make the 8700K system even more expensive but also even better performer.
Posted on Reply
#93
Captain_Tom
HTC said:
Dude: let's not kid ourselves.

Sure: 2700X catches up to Intel (and passes, on some instances) when overclocked with RAM using good timings in gaming but, as soon as you overclock the 8700K, it's game over, simply because the 8700K overclocks allot more
No, it doesn't. Actually look at the attachment (and other information found on the internet). That is a $170 R5 2600 @4.2GHz matching a $350 i7 lol. A 4.4GHz 2700X would easily win (and still cost less).

Heck in that benchmark they used ram slower than I found for cheap!
Posted on Reply
#94
HTC
Captain_Tom said:
No, it doesn't. Actually look at the attachment (and other information found on the internet). That is a $170 R5 2600 @4.2GHz matching a $350 i7 lol. A 4.4GHz 2700X would easily win (and still cost less).

Heck in that benchmark they used ram slower than I found for cheap!
While having quite lower min FPS. Dunno about you but to me, the min FPS is the more important of the two.

If you "pull all the stops" to make the Ryzen faster, it will surpass the 8700K but, if you do the same to the 8700K, it pulls even further ahead and that's the problem here because, and unless you have bad luck with the 8700K (silicon lottery), it will overclock past the 2700X every single time.

This WILL change when AMD increases the IPC but that depends on how much it increases and even then, it could still lose to Intel, depending on how much it OCs (or doesn't).
Posted on Reply
#95
ToxicTaZ
Captain_Tom said:
For anyone interested, new leaked roadmap from Intel confirms no 10nm till 2020 at the earliest lol:

https://wccftech.com/intel-xeon-scalable-family-roadmap-revealed-points-out-cascade-lake-sp-in-q4-2018-cooper-lake-sp-in-q4-2019-ice-lake-sp-in-1h-2020/


I really can't decide how much I want Intel to be skullcrushed. I want them to really learn their lesson so they never let themselves stagnate like this again, but at the same time it looks like they might be in almost as bad a position as AMD was by 2021 - and I do not want AMD to simply become the new dominating force without competition. At this point it looks like AMD may have 7nm+ out before Intel launches any 10nm product...

Who knows, maybe Intel will become so desperate they actually launch Graphene 1nm chips in 2025 or something...
Very funny

Unfortunately for your point you missed the point. Intel is still releasing 10nm Icelake S for mainstream next fall. You're article is only Server/Work Station are not getting 10nm in 2019.

So all Intel fans don't have to worry about 10nm next year for Dual channel launch.

I Reilly hope Zen 2 IPC is big improvement and get a little better closer to Intel IPC performance and trying harder and harder to get all around better performance!

It only makes, Intel do something otherwise they sit back and milk it.

Intel is 100 times bigger than AMD so if if there wasn't a law in place of buying out your compilation in place, Intel would throw pocket change at AMD.... Done

So Zen 2 vs Icelake next year going to be amazing.
Posted on Reply
#96
springs113
Just to add some data, I used to run my 2700x @4.3 with 3600mhz memory. I have since scaled it back to 4.15 for 24/7 use cases with the same memory speeds(16 cas). All I did was set it n forget it no real extra tweaking.
Posted on Reply
#97
HTC
springs113 said:
Just to add some data, I used to run my 2700x @4.3 with 3600mhz memory. I have since scaled it back to 4.15 for 24/7 use cases with the same memory speeds(16 cas). All I did was set it n forget it no real extra tweaking.
If you manage to fine tune your sub-timings @ that speed, you'll probably be getting better performance @ 4.15 then you did @ 4.3: it depends on the tuning.

EDIT

Look @ the following, and pay attention to only the graphs with RAM @ 3200 MHz:



As you can see, there's quite a difference between them, with just messing on the timings. Only the blue one to the right has a different CL timing but all the others have CL14.
Posted on Reply
#98
Captain_Tom
ToxicTaZ said:
Very funny

Unfortunately for your point you missed the point. Intel is still releasing 10nm Icelake S for mainstream next fall. You're article is only Server/Work Station are not getting 10nm in 2019.

So all Intel fans don't have to worry about 10nm next year for Dual channel launch.

I Reilly hope Zen 2 IPC is big improvement and get a little better closer to Intel IPC performance and trying harder and harder to get all around better performance!

It only makes, Intel do something otherwise they sit back and milk it.

Intel is 100 times bigger than AMD so if if there wasn't a law in place of buying out your compilation in place, Intel would throw pocket change at AMD.... Done

So Zen 2 vs Icelake next year going to be amazing.
"Next fall". You mean after 16-core 5GHz Ryzen 3 is out? Who cares at that point?!

And come on, everyone with a brain knows it's like a 66%+ chance that it slips to 2020.
Posted on Reply
#99
Vayra86
Captain_Tom said:
1. If you want price/perf - nothing is going to beat a Ryzen build the saves money where possible.

2. If you read what I said, anyone getting an 8700K doesn't care about Price/perf (if they did, they wouldn't even glance at Intel). So the "Extra cost ruins the value" BS is really not true for these people.

3. In terms of Enthusiast Price/Perf, you are still wrong. Ryzen is cheaper AND better performing. Here are 2 different ram kits:

https://www.newegg.com/Product/Product.aspx?Item=N82E16820313771&cm_re=DDR4-_-20-313-771-_-Product

https://www.newegg.com/Product/Product.aspx?Item=N82E16820313618&cm_re=DDR4-_-20-313-618-_-Product

It only costs $70 to go from 2400 to [B]3733 memory! That's less than the price difference required to buy the 8700K (+$50), the AIO Liquid Cooler (+$75), and the more expensive Intel motherboard (+$25). You are actually saving a lot of money (and energy usage) by going with the 2700X + Fast RAM instead of the 8700K + Peasant RAM. Also, it is going to beat the 8700K system at gaming and allow upgrades to 7nm Zen next year.[/B]
You realize there is also an 8600K that has 100% similar gaming performance and will cut the cost down slightly below the Ryzen with fast memory option? AND clocks higher on average than the 8700K at less heat = lower cost of cooling?

Quit acting like you've found the holy grail or something. The landscape hasn't changed - Intel is the top option for high refresh rate gaming - cost and performance wise with the 8600K and 8700K respectively. For everything else, Ryzen 2 is the go-to option. And that's still quite a feat.
Posted on Reply
#100
Caring1
cucker tarlson said:
That 8c/8t is the best thing that could happen for gaming. Look how stellar 8600k performs against 8700k, now add +25% more cores and +33% more cache.
In essence I agree, but they should have left the option for the end user to turn hyperthreading off.
Posted on Reply
Add your own comment