Does an 8/8 Zen "Summit Ridge" SKU make good sense?

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Does an 8/8 Zen "Summit Ridge" SKU make good sense?


  • Total voters
    52
  • Poll closed .
Oct 19, 2006
194
1
81
Like a few other people have said, It would cost AMD more to create a new die that lacked the SMT, than it would to simply salvage defective dies and disable the SMT. The only way we would see an 8/8 layout is if there are enough defective dies to produce such a product or the 8/16 is so expensive and there is enough demand + profit margin to sell working 8/16 dies as 8/8. I wouldn't count on the latter.
 

HurleyBird

Platinum Member
Apr 22, 2003
2,684
1,268
136
It really depends on yields, how much if any additional dies can be salvaged by disabling SMT versus other methods of cutting, binning, and overall competitiveness.

My assumption is that disabling SMT isn't likely to save many dies compared with more obvious things like cache and cores, and that unlike Intel AMD is going to need to do their absolute best to compete. In other words, unless Zen is way more competitive than anyone expects, I doubt they have the luxury of disabling performance features to aid in product segmentation. I think it's more likely that they'll need to put their best foot forward against every product segment versus Intel, which would mean as little artificial gimping as possible.
 

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
Cool, that's one game. One game does not a trend make, though I would love to see games take advantage of more cores as it gets harder to wring more ST perf out of these chips.

It's also only the 6700. The 6700K would be much faster.
 

StinkyPinky

Diamond Member
Jul 6, 2002
6,766
784
126
What am i missing here? Can't you just disable the threads in the bios if you don't want them?
 

ClockHound

Golden Member
Nov 27, 2007
1,108
214
106
What am i missing here? Can't you just disable the threads in the bios if you don't want them?

Not just threads, but cores too. You want power savings from your 18C/36T Xeon? Turn all the cores OFF. Maybe leave one running...for emergency gaming purposes.
 

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
Not just threads, but cores too. You want power savings from your 18C/36T Xeon? Turn all the cores OFF. Maybe leave one running...for emergency gaming purposes.

Can you do this? Some of the work computers are 8C 16T Xeons, whereas I'd ideally like a 4T machine with much higher effective clocks.

Obviously, SMT is disabled, but could I totally disable Core1, 2, 5 and 6 leaving Cores0, 3, 4 and 7 active? [which would hopefully get a good thermal spread...]
 

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
I haven't read all of the discussion above, but to the original question, an 8/8 SKU makes little sense to me. Why?

(note: all of this assumes AMDs SMT solution to be comparable in performance to Intel's, which I see no reason for it not to be)

1) It'll most likely be more expensive than a 6/12 SKU (no bad dies allowed!), while performing worse in most workloads
2) It will most likely clock lower than a 6/12 SKU. Again, worse performance.
3) Why would AMD waste fully functioning dies on non-SMT SKUs?
3) What would consumers gain from it? From looking at Intel's chips, a high-clocked 4/8 chip is better for most workloads than a lower-clocked 8/8 chip.


For most users, I'd say unlocked, 95+W (non-power limited) 4/8 or 6/12 Zen SKUs make the most sense. There's a reason why Intel has no X99 i5s - they would generally be beaten in every task by a high-end Z170 i7. Going non-SMT only makes sense in lower ranges, else you start competing with your own midrange chips.
 
  • Like
Reactions: BeepBeep2

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
Can you do this? Some of the work computers are 8C 16T Xeons, whereas I'd ideally like a 4T machine with much higher effective clocks.

Obviously, SMT is disabled, but could I totally disable Core1, 2, 5 and 6 leaving Cores0, 3, 4 and 7 active? [which would hopefully get a good thermal spread...]
Some BIOSes allow for choosing the number of active cores. They don't let you choose specific ones, just the amount active. And regarding thermal spread: the cores on 4-core i7s are grouped just as close together as any other die, and they don't have issues with thermal spread while clocking very high ...
 

jpiniero

Lifer
Oct 1, 2010
14,599
5,218
136
I haven't read all of the discussion above, but to the original question, an 8/8 SKU makes little sense to me. Why?

It would be faster, maybe not by much but it would be. I do think it's going to be just 8C16T, 6C12T and 4C8T, but only because of simplicity, with maybe the best of the 8C16T dies having a higher priced model.
 

DrMrLordX

Lifer
Apr 27, 2000
21,633
10,845
136
Wow .. 5960x vs 6700 haswell vs. skylake, 3.0 vs 3.4 ... thats an amazing gain ~40% (mins)?
Wonder how the hell they achieved that level of parallelism in a game engine..

Don't forget that the 5960x also has more L3 than the 6700.
 

ClockHound

Golden Member
Nov 27, 2007
1,108
214
106
Can you do this? Some of the work computers are 8C 16T Xeons, whereas I'd ideally like a 4T machine with much higher effective clocks.

Obviously, SMT is disabled, but could I totally disable Core1, 2, 5 and 6 leaving Cores0, 3, 4 and 7 active? [which would hopefully get a good thermal spread...]

The bios on my Xeon systems allow disabling cores. But, I choose not too. Because...moar cores! ;-)

With your 8 core Xeon unless its unlocked won't allow overclocking, so that takes the sport out it.
 

dogen1

Senior member
Oct 14, 2014
739
40
91
No. AMD will want to be as competitive as possible, and I doubt disabling SMT will improve yields much at all.
 

jpiniero

Lifer
Oct 1, 2010
14,599
5,218
136
No. AMD will want to be as competitive as possible, and I doubt disabling SMT will improve yields much at all.

I could see a 8C8T model for die that don't have any busted CPU cores but maybe some L3 is busted or the frequency/voltage doesn't meet the standard for the top model.
 

dogen1

Senior member
Oct 14, 2014
739
40
91
I could see a 8C8T model for die that don't have any busted CPU cores but maybe some L3 is busted or the frequency/voltage doesn't meet the standard for the top model.

True, that might be a possibility.
 

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
It would be faster, maybe not by much but it would be. I do think it's going to be just 8C16T, 6C12T and 4C8T, but only because of simplicity, with maybe the best of the 8C16T dies having a higher priced model.
Faster? Faster than what? A 4C8T chip? Sure. Bur sure as hell not faster than a 6C12T chip. It might be clocked slightly higher than the 8C16T versions, but so what? It would still be very, very expensive for what you get (as there is no chance this SKU would be cheaper than any 6C12T SKU, let alone 4C8T). And you would more than likely be able to reach better clocks with a 6C12T chip anyway. No thanks.
 

jpiniero

Lifer
Oct 1, 2010
14,599
5,218
136
It would still be very, very expensive for what you get (as there is no chance this SKU would be cheaper than any 6C12T SKU, let alone 4C8T). And you would more than likely be able to reach better clocks with a 6C12T chip anyway. No thanks.

Well, yeah. The point of it is to maximize AMD's revenue.
 

DrMrLordX

Lifer
Apr 27, 2000
21,633
10,845
136
It would be awesome if we could isolate that cause this is, IMO, unprecedented parallelism in a game, it warrants further investigation..

It would shed some light on the situation. Disabled all the L3 on both CPUs (or part of the L3 on the 5960x) would help . . .
 

coercitiv

Diamond Member
Jan 24, 2014
6,203
11,909
136
1) It'll most likely be more expensive than a 6/12 SKU (no bad dies allowed!), while performing worse in most workloads
2) It will most likely clock lower than a 6/12 SKU. Again, worse performance.
Let's see how this "worse performance" looks like depending on CPU load, considering SMT would account for about +25% IPC increase and taking 6c/6t as the base for relative performance increase:
  • Until 6 threads both 8c/8t and 6c/12t will perform at almost identical performance levels with a 6c/6t, including power usage.
  • At 7 threads the 8c/8t will offer 16% more performance and the 6c/12t will bring +4% over 6c/6t.
  • At 8 threads the 8c/8t will be at +33% , the 6c/12t will offer +8%.
  • At 10 threads the 6c/12t will jump at 16% relative advantage, still considerably behind the 33% advantage of the "inferior" 8c/8t.
  • It takes 12 threads for the 6c/12t to hit it's max throughtput of +25%, and with the help of somewhat smaller power usage at full load come out even or slightly on top of the 8c/8t.
While some may argue SMT offers a bit more power efficiency than extra raw cores - leading to higher clocks , this will only happen in scenarios where thread count well past 8, more likely 10-12.

Moral of the story? SMT needs higher thread count to truly shine. Physical cores kick in faster.

3) Why would AMD waste fully functioning dies on non-SMT SKUs?
Ask Intel, they've been doing this kind of wasting for many years now.

3) What would consumers gain from it? From looking at Intel's chips, a high-clocked 4/8 chip is better for most workloads than a lower-clocked 8/8 chip.
If that were true, then consumer would have nothing to gain from a 6c/12t chip either. Everything above 4c/8t simply doesn't make any sense, high clocked Intel 4/8 chips also beat lower clocked 6/12 chips.
 
Last edited:

whm1974

Diamond Member
Jul 24, 2016
9,460
1,570
96
Let's see how this "worse performance" looks like depending on CPU load, considering SMT would account for about +25% IPC increase and taking 6c/6t as the base for relative performance increase:
  • Until 6 threads both 8c/8t and 6c/12t will perform at almost identical performance levels with a 6c/6t, including power usage.
  • At 7 threads the 8c/8t will offer 16% more performance and the 6c/12t will bring +4% over 6c/6t.
  • At 8 threads the 8c/8t will be at +33% , the 6c/12t will offer +8%.
  • At 10 threads the 6c/12t will jump at 16% relative advantage, still considerably behind the 33% advantage of the "inferior" 8c/8t.
  • It takes 12 threads for the 6c/12t to hit it's max throughtput of +25%, and with the help of somewhat smaller power usage at full load come out even or slightly on top of the 8c/8t.
While some may argue SMT offers a bit more power efficiency than extra raw cores - leading to higher clocks , this will only happen in scenarios where thread count well past 8, more likely 10-12.

Moral of the story? SMT needs higher thread count to truly shine. Physical cores kick in faster.


Ask Intel, they've been doing this kind of wasting for many years now.


If that were true, then consumer would have nothing to gain from a 6c/12t chip either. Everything above 4c/8t simply doesn't make any sense, high clocked Intel 4/8 chips also beat lower clocked 6/12 chips.
Depends on the workload and kind of software you are using.
 

coercitiv

Diamond Member
Jan 24, 2014
6,203
11,909
136
Depends on the workload and kind of software you are using.
Of course it depends on the workload/software, but don't you think it's strange to claim 6c/12t is superior to 8c/8t considering it takes 11+ threads to match throughtput in a SMT friendly environment? Is that not a specific scenario?

I believe I was quite honest in the comparison, considering I could have invoked software where SMT yields less than 25% IPC boost (for some it's even negative, including some games) and where raw cores would have a significant advantage all over the load spectrum. But that would be counterproductive.

It's not as wasteful, from the point of view of a typical gaming enthusiast, to have an 8/8 than a 4/4.
Is it? How exactly are you measuring the waste? Keep in mind you loose the exact same relative performance. If X% is the boost SMT gives you, then both 4/8 and 8/16 will have exactly X% advantage over 4/4 and 8/8 respectively.

If we take SMT at +25% again and look at relative perf increase over 4/4 in games with 8+ threads:
4/4 is base
4/8 is 25% faster
6/6 is 50% faster
6/12 is 75% faster at 8t, 90% faster at 12t
8/8 is 100% faster
8/16 is 100% faster at 8t, 150% faster at 16t

6/12 may make more sense financially, but 8/8 also makes sense if AMD really wants (and can afford) to take on Intel's mainstream 4/8 chips, especially for gaming enthusiasts and games designed with consoles in mind. I have a hard time imagining them both on the SKU list, except maybe with some "tweaks" to make segmentation possible.

For consumers, 8/8 chips would be a win, not a loss.

PS: look at that 6/6 vs 6/12 and think 7 threads from console games.
 
Last edited:

whm1974

Diamond Member
Jul 24, 2016
9,460
1,570
96
A 8c/8t Zen could a real big win for consumers depending on the pricing. If the price is reasonable then I expect that it will sell well.
 

AtenRa

Lifer
Feb 2, 2009
14,001
3,357
136
After 4C8T SKUs the price is too high and its better to play with two models per segment, one clocked lower and one clocked higher which is also unlocked at a little higher price.

Example,

4C8T = $300
4C 8T unlocked = $350

and then

6C 12T = $400
6C 12T Unlocked = 450

8C16T = $500
8C16T Unlocked = $600

So you dont really need an 8C8T
 

whm1974

Diamond Member
Jul 24, 2016
9,460
1,570
96
I think AMD will have to sell Zen CPUs a little bit cheaper then that, AtenRa.