News Zen 2 live Blog

https://www.anandtech.com/show/13547/amd-next-horizon-live-blog-starts-9am-pt-5pm-utc

We have Zen 2 with 7nm and GPUs on 7nm today! Post your reactions to the AMD event in San Francisco below!

 

[el01]

"Better Infinity Fabric" is being hinted at.
Zen 2 will hit market in 2019.
1.25x performance at same power!

Based on TSMC 7nm node.

"The competition can't compete"

 

[NateDawg72]

Interesting stuff for Epyc but I don't think it tells us much about consumer zen2. Unlike last time where everything except APUs used the same 8 core die, it looks like the server and consumer chips are diverging. I can't imagine AMD will use the chiplets with separate IO for consumers except maybe for the next generation of Threadripper. Since they are already separate dies, I'm wondering if Epyc 7nm cores will be a bit different from the Ryzen 7nm cores similar to how Intel went with Skylake-X.

There were lots of good bits mentioned though, and I'm hopeful on the overall improvements.

 

[el01]

Interesting stuff for Epyc but I don't think it tells us much about consumer zen2. Unlike last time where everything except APUs used the same 8 core die, it looks like the server and consumer chips are diverging. I can't imagine AMD will use the chiplets with separate IO for consumers except maybe for the next generation of Threadripper. Since they are already separate dies, I'm wondering if Epyc 7nm cores will be a bit different from the Ryzen 7nm cores similar to how Intel went with Skylake-X.

There were lots of good bits mentioned though, and I'm hopeful on the overall improvements.

Maybe we can get some moar cores with moar performance on desktop?

(and maybe a Ryzen 3 2200G-Ultra-E for as little power usage as possible)

 

[Kandirma]

7nm offers 2x density, 0.5x power (at same perf), 1.25x perf (at same power)

So, would that mean we can get a chip that is 8c/16t 5GHz at 95w stock? (1.25x a 2700x would be 4.6GHz with boost to 5.4GHz ) I have a really hard time believing that, but if so that'd certainly more than give intel a run for their money, that'd be handily beating a 9900k with better thermal performance if it performed the same...especially if it hits market in 2019 as suggested...

 

[el01]

So, would that mean we can get a chip that is 8c/16t 5GHz at 95w stock? (1.25x a 2700x would be 4.6GHz with boost to 5.4GHz ) I have a really hard time believing that, but if so that'd certainly more than give intel a run for their money, that'd be handily beating a 9900k with better thermal performance if it performed the same...especially if it hits market in 2019 as suggested...

I think that's for servers, but who knows? Maybe it carries over?

 

[Kandirma]

I think that's for servers, but who knows? Maybe it carries over?

I suspect you're right for this presentation- but [other than cost] I don't see why the physics that lead to performance/thermal gains would be any different for desktop chips.

My guess is we will hear more about it at CES2019 which is right around the corner, so should be interesting to see how this translates to consumer stuff.

 

[el01]

I suspect you're right for this presentation- but [other than cost] I don't see why the physics that lead to performance/thermal gains would be any different for desktop chips.

My guess is we will hear more about it at CES2019 which is right around the corner, so should be interesting to see how this translates to consumer stuff.

I mean, product segment differentiation?

Another random thought: What if AMD decided to make a 2x 7nm Vega 64 (thanks to lowered power consumption)? Imagine how great that might be... (not in terms of cost)

 

[NateDawg72]

Regarding the 1.25x number, AMD is basically quoting TSMCs numbers. I wouldn't read too much into it. I really want zen2 to be awesome, but if I consider it realistically then I think there is zero chance it can do 5ghz on safe voltages.

Just speculating for the fun of it, I'm expecting ~4.5ghz and 5-10% improvement in IPC

Based on what they showed with epyc I'm expecting mainstream will remain 8 cores. I don't see 16 cores happening on AM4. I'm doubtful about 12 as a possibility

 

[VegetableStu]

Just speculating for the fun of it, I'm expecting ~4.5ghz and 5-10% improvement in IPC

that's all I ask for, LOL. my current 6800K turbos at about 3.6GHz (well, stock settings ,_,) so anything above 4 is already welcome, LOL

 

[NateDawg72]

that's all I ask for, LOL. my current 6800K turbos at about 3.6GHz (well, stock settings ,_,) so anything above 4 is already welcome, LOL

Yes it sounds like great improvement to me too

I'm just concerned because I'm seeing optimism go beyond reason in some parts of the internet (especially Amd subreddit, no surprise, but other places too). The hype train memes are starting and people are saying 4.7+ghz and 15% IPC uplift people will end up disappointed in an excellent product if they expect that

 

[VegetableStu]

The hype train memes are starting and people are saying 4.7+ghz and 15% IPC uplift people will end up disappointed in an excellent product if they expect that

 

[el01]

Kill it with fire! Kill it with water! Kill it with voltage! Just kill it!

 

[Kandirma]

i fI consider it realistically then I think there is zero chance it can do 5ghz on safe voltages.

Just speculating for the fun of it, I'm expecting ~4.5ghz and 5-10% improvement in IPC

Like I said in my first post, I have a really hard time believing 25% gain at same power. I honestly don't even expect to see a jump to 4.5GHz stock right off the bat, maybe boost. I think the most we'll see from first gen is - hopefully - lower TDP.

I don't think it is impossible to do 5GHz on safe voltages in the future - BUT I also don't expect the first chip using a new process to be substantially better. The jump from final generation 22nm chips to first gen 14nm chips was slight, and actually maybe even a step backwards*. However we're still on 14nm today, and the jump from EOL 22nm to present-day 14nm is ~25% less power for ~2-15% more performance - or maybe even slightly more depending which chip you think is most comparable. That said the % drop from 14 -> 10 is much smaller than 22 ->14

Of course, that is all Intel, let's look at AMD's jump from 22nm to 14.......unfortunately I can't because AMD skipped straight from 32nm to 14nm, no wonder they couldn't compete for so long! Unfortunately I don't think I have any way to even try to compare 32 ->14 to even 14 -> 10.

If we see similar numbers to the 22->14 jump intel did but over 1-2 years instead of 5 I'd be extremely excited, that'd make a R7 2700x 'equivalent' go from Clock: 3.7 Boost:4.3 TDP: 105W - > Clock 3.8 Boost: 4.9-5.3 TDP:~80W.

I'm not going to get hyped by any means, but I'm cautiously optimistic that this could seriously help AMD do more than play catchup.


*I'm also not comparing like-priced chips here by any means, but the top of the line enthusiast chip at the respective times to the best of my knowledge.

 

[NateDawg72]

Yeah I think that's a good assessment

I have a hard time believing TSMC can actually pull off >5Ghz so that's why I'm skeptical on clockspeeds. Maybe on a "7+" and/or when they start using EUV but I don't know. I think Intel has focused harder than any other foundry on making processes that can handle high clock speeds, and my gut feeling is that it is harder than Intel makes it look. Maybe because Intel was forced to spend several more years at it with how delayed 10nm became lol

I think the power efficiency and density improvements of TSMC 7nm are going to be pretty great though!

(oh man, I shouldn't have checked in at the AMD subreddit. So many people there think AMD will do MCM for the full lineup down to ryzen, chiplets everywhere, separate IO dies for ryzen/apus I can only imagine terrible things will happen to latency)

 

[br3nd0]

Keywords "cautiously optimistic" I like that. I was close to jumping on the AMD bandwagon for Zen+ but decided to wait for Zen 2. As an SFF enthusiast, anything 4ghz+ is more than OK for me

 

[el01]

Yeah I think that's a good assessment

I have a hard time believing TSMC can actually pull off >5Ghz so that's why I'm skeptical on clockspeeds. Maybe on a "7+" and/or when they start using EUV but I don't know. I think Intel has focused harder than any other foundry on making processes that can handle high clock speeds, and my gut feeling is that it is harder than Intel makes it look. Maybe because Intel was forced to spend several more years at it with how delayed 10nm became lol

I think the power efficiency and density improvements of TSMC 7nm are going to be pretty great though!

(oh man, I shouldn't have checked in at the AMD subreddit. So many people there think AMD will do MCM for the full lineup down to ryzen, chiplets everywhere, separate IO dies for ryzen/apus I can only imagine terrible things will happen to latency)

Silicon glue and pipeline tape should do the job, if you know what I mean

I just came up with something stupid: Since Vega 7nm is two times as smol, what if we have 2 Vega 64 dies on the normal package size for current Vega 64?

 

[NateDawg72]

I just came up with something stupid: Since Vega 7nm is two times as smol, what if we have 2 Vega 64 dies on the normal package size for current Vega 64?

It would be really interesting to see if they could design it to work as if it was a single chip rather than 2 chips in crossfire/SLI. If it worked just like crossfire or the old 7990 style cards then it would probably not be any good

Edit: not be any good for gaming I mean

 

[Stevo_]

So, would that mean we can get a chip that is 8c/16t 5GHz at 95w stock? (1.25x a 2700x would be 4.6GHz with boost to 5.4GHz ) I have a really hard time believing that, but if so that'd certainly more than give intel a run for their money, that'd be handily beating a 9900k with better thermal performance if it performed the same...especially if it hits market in 2019 as suggested...

Sounds like they are using TSMC 7nm raw marketing numbers. I only use the TSMC foundry standard cell libraries but usually the speed increase they cite in node vs node is at some constant lower core VDD that the something like 28nm(also TSMC) we came from would tank timing-wise. If they were to quote both libraries at their individual max VDD(which is different, less as you go down in nodes) it's not so much. I doubt any large company is not using in-house custom libraries as the foundry ones kinda suck. The other reality is that at 7nm the routing rules(including power grid), high routing layer impedance, and mandatory use of vias is brutal so the actual attainable speed increase from the technology is not as impressive at least that I've seen so far versus a gate vs gate comparison. Leakage and dynamic power has been significantly less but some of that is due to the lower VDD as well and not the technology itself.

I've seen some slides on 7nm+ where it was simpler to use especially for layout but potentially not as fast, again using foundry IP. Very prelim though. There's a move to go from or even skip TSVs to a more die on die approach as well(even more 3D) which appears to be the chiplet approach as they infer in the article on scaling, the IO cells stay relatively the same size so might as well hang at 14nm there and use it as a base.

 

[el01]

It would be really interesting to see if they could design it to work as if it was a single chip rather than 2 chips in crossfire/SLI. If it worked just like crossfire or the old 7990 style cards then it would probably not be any good

Edit: not be any good for gaming I mean

Perhaps they could use "infinity fabric"?