Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

Page 280 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Sort by date Sort by votes
yottabit

yottabit

Golden Member
Jun 5, 2008
1,374
238
116
mikegg said:
Wake me up when AMD can make an M1 competitor. Yes, I said M1. A SoC that can go into something as thin and small as a Macbook Air, fanless, 20 hours of battery life, extremely snappy, and performs the same plugged in or on battery.

In 2024, 3+ years after the M1, AMD is still no where close.
Click to expand...
I don’t think that’s really true. The Lenovo p14s G2 AMD and p16s G2 AMD are good examples. Depending on configuration, 10 to 16 hours of battery life watching videos. The power limit is only slightly reduced when on battery.

Full load will drain the battery faster, yeah, but it also will get way more work done with 8 P cores. Integrated graphics are on the level of a 2050

I’d say they are quite close already. A lot of the problem though is laptop OEMs but they seem to be catching on to the fact that people want Macbook level battery life on PCs too. Another problem is Windows vs OS X. Did you ever load Windows on an Intel Mac via bootcamp and compare battery life, even with all the right drivers? Not great

Edit: Meant to link reviews

These are Zen 4 notebooks so if Zen 5 offers any further efficiency improvements I don’t think it will be as hard as you think to close the gap
 
  • Like
Reactions: Tlh97, dr1337 and lightmanek
Upvote 0 Downvote
DisEnchantment

DisEnchantment

Golden Member
Mar 3, 2017
1,659
6,101
136
yuri69 said:
This confirms MLID has been supplied by real slides for some time.

The 4 decode with a big uop cache carried on?
Click to expand...
Seems the slides are legit indeed.

Basically whatever Mike Clark's remarked about Zen's future are materializing, of course since he is the CPU Chief.
  • Bigger CCXs.
  • Wider Frontend
  • Xilinx IP integration
But the decode is still not clear, now from the patches still at four.
The patches are probably not final and they even mentioned they would update it going forward.
Zen 4 scheduler model did not change since 2022 for instance.

Mike did make the remark below when asked about wider decode.
I think going beyond four with the number of transistors and the smarts we have in our branch predictor, and the ability to feed it worked fine. But we are going to go wider, you're going to see us go wider, and to be efficient, we'll have the transistors around the front end of the machine to make it the right architectural decision.
Click to expand...

I am interested in the decode width and how they implemented it, lots of patents point towards Tremont style dual decode.
Whatever it may be, they would have modelled it well enough and if they added lots of execution ports they would need to increase the decode width/op cache.
 
  • Like
Reactions: Tlh97, Elfear, Vattila and 4 others
Upvote 0 Downvote
I

itsmydamnation

Platinum Member
Feb 6, 2011
2,848
3,389
136
DisEnchantment said:
Seems the slides are legit indeed.

Basically whatever Mike Clark's remarked about Zen's future are materializing, of course since he is the CPU Chief.
  • Bigger CCXs.
  • Wider Frontend
  • Xilinx IP integration
But the decode is still not clear, now from the patches still at four.
The patches are probably not final and they even mentioned they would update it going forward.
Zen 4 scheduler model did not change since 2022 for instance.

Mike did make the remark below when asked about wider decode.


I am interested in the decode width and how they implemented it, lots of patents point towards Tremont style dual decode.
Whatever it may be, they would have modelled it well enough and if they added lots of execution ports they would need to increase the decode width/op cache.
Click to expand...
i assume if its dual decode , and each "stream" is max 4 wide then compilers don't need to care.

If i remember correctly , AMD have had 32byte cycle of fetch but have only been able to really hit ~18bytes of decoded instructions. So i wonder if L1i stays same-ish with the same fetch but getting much closer to 32bytes of instructions.
 
  • Like
Reactions: Tlh97, HutchinsonJC, moinmoin and 1 other person
Upvote -1 Downvote
I

itsmydamnation

Platinum Member
Feb 6, 2011
2,848
3,389
136
FlameTail said:
How great is Mike Papermaster?

Is he on the same league as the likes of Jim Keller?
Click to expand...

Mike Clarke ,

he was/is an architect from pre Zen for AMD , and was one of the people who was in Jim kellers inner circle while the convinced the rest of AMD they could actually build Zen1. He now does the CPU road map and doesnt focus on specific cores now.
 
  • Like
Reactions: Tlh97, Vattila, RnR_au and 3 others
Upvote 0 Downvote
M

mikegg

Golden Member
Jan 30, 2010
1,815
445
136
yottabit said:
I don’t think that’s really true. The Lenovo p14s G2 AMD and p16s G2 AMD are good examples. Depending on configuration, 10 to 16 hours of battery life watching videos. The power limit is only slightly reduced when on battery.

Full load will drain the battery faster, yeah, but it also will get way more work done with 8 P cores. Integrated graphics are on the level of a 2050

I’d say they are quite close already. A lot of the problem though is laptop OEMs but they seem to be catching on to the fact that people want Macbook level battery life on PCs too. Another problem is Windows vs OS X. Did you ever load Windows on an Intel Mac via bootcamp and compare battery life, even with all the right drivers? Not great

Edit: Meant to link reviews

These are Zen 4 notebooks so if Zen 5 offers any further efficiency improvements I don’t think it will be as hard as you think to close the gap
Click to expand...
1. This laptop isn't fanless.

2. The CPU goes up to 51 watts. On the M1 Air, the CPU only goes up to 10w.

1707648014597.png

3. Even though the 7840U uses 51watts, it still loses in GB6 ST and barely wins in GB6 MT - while using up to 5x more power.

4. This thing is $2,200 compared to $750 for M1 Air.

Like I said, AMD still can't match the M1 3.5 years ago. I doubt the even Zen5 can match the M1 in per/watt. AMD is about 3-5x behind in efficiency. Zen5 isn't going to make much of a difference. AMD needs to go back to the drawing board if it wants to compete against Apple, and soon Qualcomm as well.
 
Upvote 0 Downvote
DisEnchantment

DisEnchantment

Golden Member
Mar 3, 2017
1,659
6,101
136
Bigos said:
  • cost of a divide/mod for
    • QI: 12 -> 10
    • HI: 13 -> 11
    • SI: 13 (no change)
    • DI: 18 -> 16
    • Other: 18 -> 16
  • cost of DIVSS instruction: 13 -> 10
  • cost of DIVSD instruction: 13 (no change)
  • cost of SQRTSS instruction: 15 -> 14
  • cost of SQRTSD instruction: 21 -> 20
Click to expand...
One thing which I am wondering is that znver5_cost table is aligned with the instruction construction in znver5.md in terms default latency for 8b/16b/32b/64b divisions. Comparing the znver4.md to znver5.md makes it seems as if there are regressions here but not in the scheduling costs table.

1707648235163.png
1707648287272.png

Any thoughts?

Update:

Actual latencies from https://www.amd.com/content/dam/amd...l-docs/software-optimization-guides/57647.zip
So the default latencies in the znver4.md are correct.

1707650118950.png

It looks to me there is 1 cycle regression in all sub 64bit div, but 64b div improved by 2 cycles.
 
Last edited:
  • Like
Reactions: Tlh97, Elfear, Vattila and 5 others
Upvote 0 Downvote
FlameTail

FlameTail

Platinum Member
Dec 15, 2021
2,930
1,657
106
itsmydamnation said:
Mike Clarke ,

he was/is an architect from pre Zen for AMD , and was one of the people who was in Jim kellers inner circle while the convinced the rest of AMD they could actually build Zen1. He now does the CPU road map and doesnt focus on specific cores now.
Click to expand...
Ah. The two Mikes are confusing. One is a Clark and the other is a Papermaster! :p

Edit: It's *Mark Oapermaster.
 
Last edited:
Upvote 0 Downvote
yottabit

yottabit

Golden Member
Jun 5, 2008
1,374
238
116
mikegg said:
1. This laptop isn't fanless.

2. The CPU goes up to 51 watts. On the M1 Air, the CPU only goes up to 10w.

View attachment 93644

3. Even though the 7840U uses 51watts, it still loses in GB6 ST and barely wins in GB6 MT - while using up to 5x more power.

4. This thing is $2,200 compared to $750 for M1 Air.

Like I said, AMD still can't match the M1 3.5 years ago. I doubt the even Zen5 can match the M1 in per/watt. AMD is about 3-5x behind in efficiency. Zen5 isn't going to make much of a difference. AMD needs to go back to the drawing board if it wants to compete against Apple, and soon Qualcomm as well.
Click to expand...
Ok, but I don’t think users care about any of that. It might not be fanless but it’s quiet. Geekbench 6 sucks and doesn't represent true MT performance. Nobody cares how many peak watts it uses (and that may not be the flex you think it is since the M1 only has 4 P cores). I’m sure it doesn’t use 51 W in ST, or the battery life would be garbage. And once you get past “all day” / 8 hour light usage battery life I don’t think many people care whether its 10 or 20 hours.

I agree they still have a ways to go in PC market though, I’m only arguing the rift isn’t as big as you think. At least in terms of “practical things a user cares about” and not specs on paper. When the M1 came out I thought it was revolutionary enough to consider buying a Mac, nowadays with even Zen4 I can *mostly* get that balance from a PC.

Arguably the much bigger rift would be the Macbook Pro with M- Pro/Max GPU, AND 20 hour battery life. You definitely won’t find dGPU Windows laptops like that. I’m hopeful Strix Halo will close the gap somewhat.

Edit: anyway, it was probably a bad comparison. I just happened to be shopping those Lenovo laptops recently. The 7840hs would be better to compare against the M1 Pro/Max for its MT and GPU. Notebookcheck has some comparisons here: https://www.notebookcheck.net/R7-7840HS-vs-M2-Pro-vs-M1-Max_14948_14973_13843.247596.0.html

I’ll concede AMD doesn’t seem to care as much about SKUs for the 10-15 W market (at least as far as I can tell, maybe someone can correct me)
 
Last edited:
  • Like
Reactions: Tlh97 and lightmanek
Upvote 0 Downvote
uzzi38

uzzi38

Platinum Member
Oct 16, 2019
2,690
6,348
146
S'renne said:
Uhh....the M1 air macs sold like hot cakes on release and still do now everywhere
Click to expand...
As much as I would LOVE for you to be right, I'm afraid it's true that OEMs actually don't really want M1 style products targeting lower TDPs, and LNL is a perfect example of that. Early on in the process OEM's complained that Intel wasn't letting them set power limits high enough, which is why max configurable TDP on LNL went up.
 
  • Like
Reactions: Tlh97, Elfear, Exist50 and 4 others
Upvote 0 Downvote
S'renne

S'renne

Member
Oct 30, 2022
136
99
61
uzzi38 said:
As much as I would LOVE for you to be right, I'm afraid it's true that OEMs actually don't really want M1 style products targeting lower TDPs, and LNL is a perfect example of that. Early on in the process OEM's complained that Intel wasn't letting them set power limits high enough, which is why max configurable TDP on LNL went up.
Click to expand...
More HP probook styled laptops incoming, I got one with an i3 ice lake, it literally thermal throttles and I had to prop it out, clean vents and everything
 
Upvote 0 Downvote
You must log in or register to reply here.

TRENDING THREADS

COMPANY
RESOURCES
FOLLOW
Top Bottom