New Zen microarchitecture details

Page 212 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

lolfail9001

Golden Member
Sep 9, 2016
1,056
353
96
Benchmark ran with the same 6C/12T ES, 3.3/3.7Ghz with 3.4Ghz ACT meaning the MT test was ran at 3.4Ghz
Actually, CPU-Z screen from yesterday and recent Passmark reaffirms that ACT is 3.45Ghz on this sample. Or is your idea that 6 means 0.15Ghz turbo, 5 means 0.05 and 1 means 0.1?
 

inf64

Diamond Member
Mar 11, 2011
3,685
3,957
136
Actually, CPU-Z screen from yesterday and recent Passmark reaffirms that ACT is 3.45Ghz on this sample.
Yep, divide my estimate by 1.0147, makes world of difference ;). Thank you for the correction though.

edit:
lolfail said:
Or is your idea that 6 means 0.15Ghz turbo, 5 means 0.05 and 1 means 0.1?
That is not my idea. 3301 in the OPN and Sample basically ran at 3.4Ghz(34). What could it be, very difficult question.
 
  • Like
Reactions: Drazick

inf64

Diamond Member
Mar 11, 2011
3,685
3,957
136
Yeah, it does not really affect results, but you nailed the meaning of the last number, it looks like.
I didn't nail it myself, dunno if I would figure it out eventually (I suppose not, even with the cpuz screens from baidu). It doesn't matter in the end, the mystery is solved.
 
  • Like
Reactions: Drazick

CatMerc

Golden Member
Jul 16, 2016
1,114
1,149
136
I didn't nail it myself, dunno if I would figure it out eventually (I suppose not, even with the cpuz screens from baidu). It doesn't matter in the end, the mystery is solved.
I see. it's just an incredibly odd way to denote all core turbo.

I'm still not sure this is actually the meaning.
 

Dresdenboy

Golden Member
Jul 28, 2003
1,730
554
136
citavia.blog.de
You did good ;). Delta is always a big n1ce round number.
Nice avatar btw :)
Thx. I'm working on a Gear S2/S3 watchface in the spare time of my spare time. :D The avatar is a byproduct. And there is no hidden message...

BTW I was wrong about the "5", because this was in the third spot, not the fourth. Even if I've won the game, I'm still not fully convinced. ;)
 
  • Like
Reactions: inf64

Greyguy1948

Member
Nov 29, 2008
156
16
91
Very much about overclocking in most fora. To me it looks like way to shorten the life of the CPU. One field I believe Ryzen must be much better in is cache design. Look at this comparison:
http://users.atw.hu/instlatx64/AuthenticAMD0600F12_K15_Zambezi8C_NewMemLat.txt

Best Page_Random Tests
M32768 Mem 62.955 nsec 227.406 clks (MemSize: 32768 KiB Stride: 128 Movs: 64 Reps: 262144 B:2097152 RT: 10.562 s) ------
M32768 L1D 1.111 nsec 4.014 clks (MemSize: 16 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.001 s) ------
M32768 L2 10.775 nsec 38.921 clks (MemSize: 1984 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.010 s) ------
M32768 L3 29.591 nsec 106.887 clks (MemSize: 10240 KiB Stride: 64 Movs: 64 Reps: 163840 B:2097152 RT: 10.239 s) ------
Best Linear_Forward Tests
M32769 Mem 18.001 nsec 65.022 clks (MemSize: 32768 KiB Stride: 128 Movs: 64 Reps: 262144 B:2097152 RT: 10.268 s) ------
M32769 L1D 1.109 nsec 4.006 clks (MemSize: 16 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.002 s) ------
M32769 L2 5.637 nsec 20.363 clks (MemSize: 1984 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.002 s) ------
M32769 L3 9.665 nsec 34.910 clks (MemSize: 10240 KiB Stride: 64 Movs: 64 Reps: 163840 B:2097152 RT: 10.033 s) ------
Best Linear_Backward Tests
M32770 Mem 21.540 nsec 77.807 clks (MemSize: 32768 KiB Stride: 128 Movs: 64 Reps: 262144 B:2097152 RT: 10.119 s) ------
M32770 L1D 1.109 nsec 4.006 clks (MemSize: 16 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.002 s) ------
M32770 L2 5.641 nsec 20.376 clks (MemSize: 1984 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.008 s) ------
M32770 L3 10.762 nsec 38.872 clks (MemSize: 10240 KiB Stride: 64 Movs: 64 Reps: 163840 B:2097152 RT: 10.043 s) ------
Best Random_Full Tests
M32771 Mem 74.718 nsec 269.894 clks (MemSize: 32768 KiB Stride: 128 Movs: 64 Reps: 262144 B:2097152 RT: 10.028 s) ------
M32771 L1D 1.255 nsec 4.532 clks (MemSize: 16 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.001 s) ------
M32771 L2 10.794 nsec 38.990 clks (MemSize: 1984 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.005 s) ------
M32771 L3 38.710 nsec 139.829 clks (MemSize: 10240 KiB Stride: 64 Movs: 64 Reps: 163840 B:2097152 RT: 10.148 s) ------

http://users.atw.hu/instlatx64/GenuineIntel00306C3_Haswell_NewMemLat.txt

Best Page_Random Tests
M32768 Mem 56.081 nsec 190.621 clks (MemSize: 32768 KiB Stride: 128 Movs: 64 Reps: 262144 B:2097152 RT: 10.350 s) ------
M32768 L1D 1.186 nsec 4.030 clks (MemSize: 32 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.001 s) ------
M32768 L2 3.549 nsec 12.062 clks (MemSize: 256 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.002 s) ------
M32768 L3 12.809 nsec 43.538 clks (MemSize: 4096 KiB Stride: 64 Movs: 64 Reps: 65536 B:2097152 RT: 10.047 s) ------
Best Linear_Forward Tests
M32769 Mem 15.423 nsec 52.425 clks (MemSize: 32768 KiB Stride: 128 Movs: 64 Reps: 262144 B:2097152 RT: 10.092 s) ------
M32769 L1D 1.214 nsec 4.125 clks (MemSize: 32 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.001 s) ------
M32769 L2 3.520 nsec 11.965 clks (MemSize: 256 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.003 s) ------
M32769 L3 3.638 nsec 12.366 clks (MemSize: 4096 KiB Stride: 64 Movs: 64 Reps: 65536 B:2097152 RT: 10.010 s) ------
Best Linear_Backward Tests
M32770 Mem 19.127 nsec 65.013 clks (MemSize: 32768 KiB Stride: 128 Movs: 64 Reps: 262144 B:2097152 RT: 10.269 s) ------
M32770 L1D 1.212 nsec 4.118 clks (MemSize: 32 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.001 s) ------
M32770 L2 3.530 nsec 11.998 clks (MemSize: 256 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.001 s) ------
M32770 L3 3.637 nsec 12.361 clks (MemSize: 4096 KiB Stride: 64 Movs: 64 Reps: 65536 B:2097152 RT: 10.006 s) ------
Best Random_Full Tests
M32771 Mem 60.721 nsec 206.393 clks (MemSize: 32768 KiB Stride: 128 Movs: 64 Reps: 262144 B:2097152 RT: 10.187 s) ------
M32771 L1D 1.214 nsec 4.125 clks (MemSize: 32 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.001 s) ------
M32771 L2 3.548 nsec 12.060 clks (MemSize: 256 KiB Stride: 64 Movs: 64 Reps: 32768 B:2097152 RT: 10.000 s) ------
M32771 L3 12.912 nsec 43.889 clks (MemSize: 4096 KiB Stride: 64 Movs: 64 Reps: 65536 B:2097152 RT: 10.019 s) ------
 

Greyguy1948

Member
Nov 29, 2008
156
16
91
Too much latency info I guess. Too make it short:
Best Random Full Test
FX-8150 L3=39 ns
FX-8150 Mem=75 ns
Haswell L3=13 ns
Haswell Mem=61 ns
 
  • Like
Reactions: Crumpet

richierich1212

Platinum Member
Jul 5, 2002
2,741
360
126
Can anyone decode this CPU?

RzfegV2.jpg
RzfegV2.jpg
 

Doom2pro

Senior member
Apr 2, 2016
587
619
106
1700x qual sample. The same first Passmark leak was about.

ZD3406BAM88F4 is 3.4 base, 8c/16t, 95W - HS65(A)

1700x is: RYZEN 7 1700X YD170XBCM88AE - 95W SR3+ - HS81(F) - 8C/16T
1800 PRO is: RYZEN 7 PRO 1800 YD180BBAM88AE - 95W - HS65(A) - 8C/16T

Emphasis on BAM... I think this Qualification Sample is closer to 1800 PRO than any other SKU/OPN, the only difference is the base clock.
 
Last edited:

lolfail9001

Golden Member
Sep 9, 2016
1,056
353
96
ZD3406BAM88F4 is 3.4 base, 8c/16t, 95W - HS65(A)

1700x is: RYZEN 7 1700X YD170XBCM88AE - 95W SR3+ - HS81(F) - 8C/16T
1800 PRO is: RYZEN 7 PRO 1800 YD180BBAM88AE - 95W - HS65(A) - 8C/16T

Emphasis on BAM... I think this Qualification Sample is closer to 1800 PRO than any other SKU/OPN, the only difference is the base clock.
Good catch.
 

Saylick

Diamond Member
Sep 10, 2012
3,084
6,184
136
Interesting to see the aspect ratio of the die be so rectangular... By my math, it's roughly 2.25:1. If the die size is 200 mm^2, then it's dimensions are roughly 9.4 mm x 21.3 mm.

I wonder if they could have eeked out more yields per wafer if they rearranged the CCXs to have the long sides facing each other instead of having the short sides facing each other, which results in something like 1.75:1 (10.6 mm x 17.8 mm). One thinks that a more square-shaped die has better packing on a single wafer than a die with a 2.25:1 aspect ratio...

Hmmm, nevermind, the online calculators actually suggest very similar dies/wafer between those two hypothetical die sizes. Even a hypothetical 5:1 die doesn't drastically reduce the total dies/wafer. Not bad, AMD, not bad.

2.25:1
T5B1sKX.png


vs.

1.75:1
qjAdH80.png


vs.

1.00:1

90nbxeO.png


vs.

5.00:1

0hkb9Sk.png
 

Doom2pro

Senior member
Apr 2, 2016
587
619
106
Interesting to see the aspect ratio of the die be so rectangular... By my math, it's roughly 2.25:1. If the die size is 200 mm^2, then it's dimensions are roughly 9.4 mm x 21.3 mm.

I wonder if they could have eeked out more yields per wafer if they rearranged the CCXs to have the long sides facing each other instead of having the short sides facing each other, which results in something like 1.75:1 (10.6 mm x 17.8 mm). One thinks that a more square-shaped die has better packing on a single wafer than a die with a 2.25:1 aspect ratio...

Hmmm, nevermind, the online calculators actually suggest very similar dies/wafer between those two hypothetical die sizes. Even a hypothetical 5:1 die doesn't drastically reduce the total dies/wafer. Not bad, AMD, not bad.

2.25:1
T5B1sKX.png


vs.

1.75:1
qjAdH80.png


vs.

1.00:1

90nbxeO.png


vs.

5.00:1

0hkb9Sk.png


It's too bad the CCX's aren't at the ends with everything else in the middle, then they could use those fringe dies by disabling the one incomplete CCX and salvage a Quad core die.

Probably more trouble than it's worth.
 
  • Like
Reactions: Drazick and Peicy

bjt2

Senior member
Sep 11, 2016
784
180
86
Interesting to see the aspect ratio of the die be so rectangular... By my math, it's roughly 2.25:1. If the die size is 200 mm^2, then it's dimensions are roughly 9.4 mm x 21.3 mm.

I wonder if they could have eeked out more yields per wafer if they rearranged the CCXs to have the long sides facing each other instead of having the short sides facing each other, which results in something like 1.75:1 (10.6 mm x 17.8 mm). One thinks that a more square-shaped die has better packing on a single wafer than a die with a 2.25:1 aspect ratio...

Hmmm, nevermind, the online calculators actually suggest very similar dies/wafer between those two hypothetical die sizes. Even a hypothetical 5:1 die doesn't drastically reduce the total dies/wafer. Not bad, AMD, not bad.

2.25:1
T5B1sKX.png


vs.

1.75:1
qjAdH80.png


vs.

1.00:1

90nbxeO.png


vs.

5.00:1

0hkb9Sk.png

There is a paper, that dates to Carrizo, in which the positions on the die was investigated. It turns out that putting the most consuming parts (FPUs and core in general) toward the centre of the die, let you gain up to 5 degrees. So the positioning is not casual, i think...
 

Tuna-Fish

Golden Member
Mar 4, 2011
1,324
1,462
136
Interesting to see the aspect ratio of the die be so rectangular... By my math, it's roughly 2.25:1.

A common reason for more rectangular dies is that the IO wants to be on the edges and as you add IO to a die you want to increase the circumference. As visible on the picture, Ryzen has a crapload of IO structures on the edges. Maybe they wouldn't fit with a more square die? (Although, the left edge is almost clear of IO.)