AMD Ryzen (Summit Ridge) Benchmarks Thread (use new thread)

[Abwx]

Well, I didn't draw a conclusion, I just asked if it COULD hide a poor branch predictor. But thanks for answering

As hinted by a member above it could as well hide a superior branch predictor...

 

[sirmo]

If turbo isn't ready, it's hard to see zen (I like this name better) would be available in January. What do you people think?

Might be ready today.. I don't think you can say either way. They are obviously working on it. It could be completed but needs to pass validation, it could be in various stages of completion. But every team will cover their ass and say it's not ready until it's tested. This is just common practice in the industry. You don't demo something until it's thoroughly tested. What if it crashes mid demo? Can you imagine?

Lisa Su has been very good about her promises in the past. She said rx480 middle of the year, and lo and behold it fell exactly on the middle of the year.

About Zen she said possible limited shipments at the end of 2016 and volume shipments in Q1 2017. So far she hasn't been wrong.

 

[blublub]

Might be ready today.. I don't think you can say either way. They are obviously working on it. It could be completed but needs to pass validation, it could be in various stages of completion. But every team will cover their ass and say it's not ready until it's tested. This is just common practice in the industry. You don't demo something until it's thoroughly tested. What if it crashes mid demo? Can you imagine?

Lisa Su has been very good about her promises in the past. She said rx480 middle of the year, and lo and behold it fell exactly on the middle of the year.

About Zen she said possible limited shipments at the end of 2016 and volume shipments in Q1 2017. So far she hasn't been wrong.

You really think AMD will ship final silicon of Ryzen in the remaining days of December - I doubt That.
Something blew the December shipments, let it be a bug or the chance for another performance improvement.
I just really hope that the thing is in stores in quantity at the end of march

 

[CatMerc]

As hinted by a member above it could as well hide a superior branch predictor...

As Nothingness said, no conclusions can be made

 

[The Stilt]

If turbo isn't ready, it's hard to see zen (I like this name better) would be available in January. What do you people think?

Since Excavator these things are being controlled by the SMU, so in case there are any actual issues it is most likely a software (firmware) one.

 

[CatMerc]

Since Excavator these things are being controlled by the SMU, so in case there are any actual issues it is most likely a software (firmware) one.

What does SMU stand for?

 

[krumme]

@wit
First you base the asumption of a die if 220mm2. Meaning same size as bwe 8c at 224mm2. Can you say with a straight face that thats what you will have used if it wasnt just posted?
I dont think so and with all due respect thats a bit hard for me to accept when i think you use numbers in your calculation you dont quite think yourself is right. You migh help me here.

I think its fair to asume its a bit smaller than p10 going by Dresdenboys estimates but also that there can be more defects and perhaps 4c and 6c will make up for it plenty. But at the end of the day its more or less the same ballpark. By all means - We have a similar business case here for our intends and purposes..

Now if we look at amd portfolio we have the consoles making a profit and seriously i think the rest sans polaris is selling at basement prices not making any money. But somewhere some of the money is comming from and polaris is a good guess.

We have 460 selling at sub 99usd retail. Make profit. Take markup at oem and retail and look at bom for eg. Board vrm ram. There is no way a polaris p10 die cost more than 50usd. And 25usd for the p11. I would asume an upper estimate for zen is those 50usd after 2 or 3 months startup time.

 

[kraatus77]

What does SMU stand for?

Single Multiple Unit


/s

 

[sirmo]

You really think AMD will ship final silicon of Ryzen in the remaining days of December - I doubt That.
Something blew the December shipments, let it be a bug or the chance for another performance improvement.
I just really hope that the thing is in stores in quantity at the end of march

At this point all it's left is fine tuning, microcode updates.. they could already be shipping product to OEMs like SuperMicro or similar.

 

[The Stilt]

What does SMU stand for?

System management unit.

 

[realibrad]

and I think you are completely wrong in your suggestion.

AMD need Zen to:
1. Allow them to reduce their leverage which has hung over them since 2006.
2. Allow them to have 2 or 3 solid iterations for continued profit and CPU growth over the next 3-5 years.
3. Allow them to have a solid x86 base for forays into HPC APUs and custom silicon/FPGAs based on x86 baselines.


If Zen is as good as is hoped, then it demonstrates small teams well led by technically competent folks rather than the powerpoint mafia with their MBA can compete with behemoths that are inefficient due to their size.

AMD could stand up quite a few more small teams of the same ilk based on Zen profits without having to go near banks. Bigger is not always better and more money is often not the solution.



Not really.
The amount of consumers shelling out for a HEDT compared to the overall market size is miniscule. The subset of those that don't know what they are looking for and don't compare across genuine competitors is vanishingly small.




No. They will look at performance and go with whoever offers them best performance or best performance/watt.

This comment gives me insight into your understanding, or lack of, of the market.





AMD are still paying up for the ATi decision. That monkey needs to be removed from their back.

I'm not suggesting going to zero-loans, or anything like it, but the bonds need to be brought under control. If Zen flops, the company probably dies due to the pending bond payment - that is not a healthy situation.




Which is what everyone else is suggesting. But not undercut them by half, which would lead to AMD not maximizing their profit - both in 2017 and over 2017-2020. That error could, and likely would, impact their long term prospects. Which is very bad for consumers and commercial alike.

AMD has over $1.6 billion in debt. Zen alone will not get their debt to equity low enough. What AMD needs to do is take Zen and build future assets. Waiting for cash from sales will take a very long time. You are arguing for AMD to wait for the cash which limits R&D to only the future. AMD is holding back in sectors because they don't have the money to do them all at once.

The reason we only have the 480x is because of money constraints. If they can get cash to launch more products they can actually increase profits on the long run. This is the main principle for loans. AMD looks to finally have competitive products across multiple segments but can't launch them all right now.

Also, companies will not buy something new because it simply has higher performance/watt or per $. If the cost to switch is higher than the net benefit, companies won't switch. That wild be like buying a new car because it gets slightly better mpg vs buying new parts. Depending on the cost and returns it may be more cost efficient to buy parts instead.

 

[sirmo]

AMD has over $1.6 billion in debt. Zen alone will not get their debt to equity low enough. What AMD needs to do is take Zen and build future assets. Waiting for cash from sales will take a very long time. You are arguing for AMD to wait for the cash which limits R&D to only the future. AMD is holding back in sectors because they don't have the money to do them all at once.

The reason we only have the 480x is because of money constraints. If they can get cash to launch more products they can actually increase profits on the long run. This is the main principle for loans. AMD looks to finally have competitive products across multiple segments but can't launch them all right now.

Also, companies will not buy something new because it simply has higher performance/watt or per $. If the cost to switch is higher than the net benefit, companies won't switch. That wild be like buying a new car because it gets slightly better mpg vs buying new parts. Depending on the cost and returns it may be more cost efficient to buy parts instead.

Zen is an important part of a range of products for AMD. It cuts the cost of development of Jaguar/Puma cores as those no longer need for them to be developed (they can probably shift resources into improvements like Zen+ and beyond and K12),

Their debt was recently restructured. And it's looking much better than it did last quarter. Their debt was $2.3B and they managed to lower it down to $1.8B but not only that, they also lowered the interest significantly and pushed the term further down the road. They also have $1B in cash in the bank, which they need, but their financials aren't looking as dire as they used to.

I think Zen and its derivatives could pay down the debt in a few years. Their graphics division I think is on a right track as well, to start taking some deep learning business. As well as strong rumors of Intel licensing deals and other semi-custom wins, like a potential custom Zen SoC for Apple.

 

[KTE]

Here we go:

Performance counter stats for 'ghb':

  1130485,461040  task-clock (msec)  #  3,393 CPUs utilized  
  483 439  context-switches  #  0,428 K/sec  
  48 996  cpu-migrations  #  0,043 K/sec  
  400 587  page-faults  #  0,354 K/sec  
2 599 208 890 203  cycles  #  2,299 GHz  
  <not supported>  stalled-cycles-frontend 
  <not supported>  stalled-cycles-backend  
3 523 266 638 003  instructions  #  1,36  insns per cycle  
  416 464 420 076  branches  #  368,394 M/sec  
  6 172 858 731  branch-misses  #  1,48% of all branches  

  333,149310230 seconds time elapsed

That's on a Haswell 4600U.

Branch density is low at 11.8%. Branch prediction is good. This means the branch MPKI is low at 1.75.

That's about the same MPKI as found on blender at RWT.

So both blender and handbrake as shown by AMD have low branch density and low branch misprediction per thousand instructions.

The blender render scene AMD chose...

//cloud.tapatalk.com/s/5857035d475fc/VT1.png?
//cloud.tapatalk.com/s/5857036a096a8/VT2.png?

...does not test the ALUs/branch predictors and is cache bound, being fully SSE ops with a L/S ratio of 1.52. It even avoids testing the FPU on Intels chip to a great extent.

It is mostly an L1 cache + scalar single precision 128b SSE bench with a some
packed FP. A lot of MOV/LEA too.

On Broadwell its FPU utilization was <1% and IPC 0.7.

When I profiled the test on Bobcat and Jaguar, every cache access at every level was a miss with 100% DRAM hits. Probably due to the cache size/associativity.

Branch Rate 3.93%
Mispredict Rate 1.01%
IC Cache Miss 100%
DC Accesses 100%
DC Misses 100%
L2 Misses 100%
Cycles FPU Empty 100%
Retired x87 FP Ops 0%
Retired SSE Ops 100%
DRAM Accesses 100%

//cloud.tapatalk.com/s/5856603332711/CA Blender1.jpg?
//cloud.tapatalk.com/s/585660378f9bb/CA Blender2.jpg?

Sent from HTC 10
(Opinions are own)

 

[KTE]

Blah, it's 2017iiish but using mobiles to post on forums is still not comparable to PC!


image post





Sent from HTC 10
(opinions are own)

 

[CatMerc]

Gah, been trying to work with Intel vTune for hours now, trying to see how to profile code branches and branch mispredictions.

I'm getting over 50% branch misprediction in Blender on my 6600K, so I doubt I did it right.

Anyone more familiar with it mind lending a helping hand?

 

[Nothingness]

These measures are useless: they just show what percentage of the events happened in blender vs the whole system, or in which function they happen; this does not tell how many such events happened

I don't think you use that tool correctly.

 

[realibrad]

Zen is an important part of a range of products for AMD. It cuts the cost of development of Jaguar/Puma cores as those no longer need for them to be developed (they can probably shift resources into improvements like Zen+ and beyond and K12),

Their debt was recently restructured. And it's looking much better than it did last quarter. Their debt was $2.3B and they managed to lower it down to $1.8B but not only that, they also lowered the interest significantly and pushed the term further down the road. They also have $1B in cash in the bank, which they need, but their financials aren't looking as dire as they used to.

I think Zen and its derivatives could pay down the debt in a few years. Their graphics division I think is on a right track as well, to start taking some deep learning business. As well as strong rumors of Intel licensing deals and other semi-custom wins, like a potential custom Zen SoC for Apple.

There is nothing incorrect in what you just said as far as I know. His argument is that AMD has giant amounts of debt so taking any more would be bad. My argument is that financially they can (and in my opinion should) take on more and speed things up.

Zen looks* great, but let's be honest, Intel has so much more money to throw at coming back if they so choose. If AMD can build enough of a lead, then the cost to come back would/might disuad Intel from things that have margins. The desktop market is diying market. They could very well let AMD have more of it simply because it's not worth long term investment. That would leave AMD to then focus on more profitable things like servers and even mobile.

But, my side is that AMD is not stuck behind a mountain of debt like they were, so they could easily take on more.

 

[Arachnotronic]

I hope AMD sends a Ryzen CPU to somebody in the tech press to benchmark against Broadwell-E/Skylake soon. Right now, all we have are two marketing demos.

 

[Arachnotronic]

Zen looks* great, but let's be honest, Intel has so much more money to throw at coming back if they so choose. If AMD can build enough of a lead, then the cost to come back would/might disuad Intel from things that have margins. The desktop market is diying market. They could very well let AMD have more of it simply because it's not worth long term investment. That would leave AMD to then focus on more profitable things like servers and even mobile.

Intel throws >$12B/year at R&D (and growing), more than 10x what AMD does. And that R&D is spent in service of building products that are better than what the competition is building so that they can reap the financial benefits.

Intel also spends a ton on sales and marketing (nearly as much as it does on R&D annually) -- just building a decent chip isn't enough to win and support the many system designs in the market. There is a lot of work that goes into winning/supporting designs and then a ton of effort that goes into working with channel/PC OEM partners to get these systems into the hands of consumers.

Really, it's just amazingly complex, and I think individuals whose perspective is that of an enthusiast who builds her or his own computers tend to miss the overarching business picture.

I don't think the argument that Intel is just letting AMD come in and rip away wads of profitable revenue really stands up to scrutiny. If AMD wins market share from Intel, it won't be because Intel let AMD have it. The published financials simply don't tell that tale.

 

[jpiniero]

At the same time it's pretty clear that Intel is deliberately slowing down to cut costs...

 

[The Stilt]

Did some profiling on the newer build I've released (on Haswell-EP):

The code itself has no obvious hotspots. Most of the time is spent in SSE kernels.

AVX2 kernel

Spoiler

OP, frequency, percentage

mov 20648 12.9911
vmovups 16045 10.0950
vmovss 8688 5.4662
vinsertps 7810 4.9138
vmulss 7338 4.6168
vmovaps 6733 4.2362
lea 6631 4.1720
vshufps 5231 3.2912
vxorps 5039 3.1704
add 4396 2.7658
movsxd 3215 2.0228
and 3202 2.0146
cmp 3170 1.9945
vmulps 3099 1.9498
jz 3007 1.8919
jmp 2799 1.7610
shl 2685 1.6893
test 2531 1.5924
vaddps 2290 1.4408
vcomiss 2224 1.3993
or 1962 1.2344
shr 1904 1.1979
vsubss 1819 1.1445
call 1798 1.1312
movzx 1786 1.1237
vcvtsi2ss 1755 1.1042
jnz 1669 1.0501
vaddss 1449 0.9117
vbroadcastss 1431 0.9003
vdpps 1424 0.8959
vdivss 1396 0.8783
xor 1129 0.7103
sub 1104 0.6946
vpshufd 1039 0.6537
vfmsub213ss 881 0.5543
vfmadd213ss 864 0.5436
imul 733 0.4612
vucomiss 708 0.4455
vsubps 677 0.4259
inc 631 0.3970
vfmadd231ps 618 0.3888
jb 610 0.3838
pop 594 0.3737
jnb 589 0.3706
vmovdqu 587 0.3693
jbe 577 0.3630
push 570 0.3586
vandps 563 0.3542
ja 517 0.3253
vmaxss 496 0.3121
retn 459 0.2888
vmovmskps 410 0.2580
cdqe 406 0.2554
cmovl 405 0.2548
vsqrtss 404 0.2542
vminss 363 0.2284
cmovg 353 0.2221
cdq 349 0.2196
idiv 346 0.2177
vcvttss2si 340 0.2139
cmovns 336 0.2114
vdivps 329 0.2070
vcmpps 313 0.1969
nop 308 0.1938
vsqrtps 303 0.1906
vextractps 266 0.1674
vzeroupper 249 0.1567
vfnmadd213ss 216 0.1359
dec 216 0.1359
vblendvps 198 0.1246
vmovq 195 0.1227
vinsertf128 183 0.1151
vmaxps 179 0.1126
jl 177 0.1114
jle 174 0.1095
vunpcklps 173 0.1088
jge 146 0.0919
setb 145 0.0912
vfmsub231ps 142 0.0893
vminps 126 0.0793
vmovsd 119 0.0749
vpinsrd 96 0.0604
vorps 88 0.0554
js 82 0.0516
vpminsd 78 0.0491
vfmadd132ss 78 0.0491
vpmaxsd 78 0.0491
bt 78 0.0491
cmovle 76 0.0478
vpextrq 74 0.0466
vmovd 71 0.0447
cmovz 62 0.0390
tzcnt 60 0.0378
vpermilps 60 0.0378
vpermps 60 0.0378
jns 53 0.0333
vpshufb 51 0.0321
cmovnb 49 0.0308
blsr 45 0.0283
vpslld 39 0.0245
vpor 38 0.0239
vfnmadd231ps 38 0.0239
vpsrld 38 0.0239
vpxor 38 0.0239
vpsubd 35 0.0220
xchg 34 0.0214
sar 34 0.0214
vhsubps 30 0.0189
vfmsub132ss 28 0.0176
jg 28 0.0176
shlx 26 0.0164
setz 24 0.0151
vextractf128 24 0.0151
vpaddd 24 0.0151
vmovdqa 22 0.0138
bts 22 0.0138
setnz 20 0.0126
vpcmpgtd 19 0.0120
vpextrd 17 0.0107
div 12 0.0076
rol 10 0.0063
sarx 9 0.0057
vpand 9 0.0057
vbroadcastf128 8 0.0050
vblendps 8 0.0050
blsi 8 0.0050
vcvtdq2ps 7 0.0044
setnbe 7 0.0044
vcvttps2dq 6 0.0038
btr 6 0.0038
cmovs 6 0.0038
vfnmadd132ss 6 0.0038
cmovnz 4 0.0025
vpcmpeqd 4 0.0025
bsr 4 0.0025
ror 4 0.0025
vhaddps 4 0.0025
not 2 0.0013
vfnmsub213ss 2 0.0013
vandnps 1 0.0006
vcvtps2dq 1 0.0006
vcvtps2ph 1 0.0006
setnle 1 0.0006
mul 1 0.0006
vmovlps 1 0.0006
bsf 1 0.0006
vmovhlps 1 0.0006

AVX kernel

Spoiler

OP, frequency, percentage

mov 20660 12.8232
vmovups 16350 10.1481
vmulss 9756 6.0553
vmovss 8681 5.3881
vinsertps 7973 4.9487
lea 6696 4.1561
vshufps 6207 3.8526
vxorps 5068 3.1456
vmovaps 5054 3.1369
add 4412 2.7384
vmulps 3869 2.4014
vsubss 3486 2.1637
movsxd 3347 2.0774
and 3165 1.9644
cmp 3109 1.9297
vaddps 2949 1.8304
jz 2932 1.8198
jmp 2795 1.7348
shl 2729 1.6938
test 2468 1.5318
vcomiss 2401 1.4902
vaddss 2385 1.4803
or 1933 1.1998
shr 1904 1.1818
call 1798 1.1160
movzx 1786 1.1085
vcvtsi2ss 1753 1.0880
jnz 1646 1.0216
vdpps 1424 0.8838
vdivss 1394 0.8652
xor 1135 0.7045
vpshufd 1120 0.6952
sub 1106 0.6865
vsubps 781 0.4847
imul 731 0.4537
vucomiss 705 0.4376
jb 670 0.4159
inc 657 0.4078
pop 647 0.4016
push 623 0.3867
jnb 619 0.3842
ja 608 0.3774
jbe 576 0.3575
vandps 562 0.3488
vmovdqu 546 0.3389
vmaxss 499 0.3097
retn 450 0.2793
cmovl 405 0.2514
cdqe 403 0.2501
vsqrtss 402 0.2495
vminss 363 0.2253
cmovg 353 0.2191
vmovmskps 350 0.2172
cdq 349 0.2166
idiv 346 0.2148
vcvttss2si 341 0.2117
cmovns 336 0.2085
vdivps 330 0.2048
vsqrtps 304 0.1887
nop 294 0.1825
vcmpps 283 0.1757
vbroadcastss 265 0.1645
vextractps 264 0.1639
dec 219 0.1359
vzeroupper 217 0.1347
vblendvps 198 0.1229
vmovq 195 0.1210
vmaxps 179 0.1111
jl 177 0.1099
jle 174 0.1080
vunpcklps 173 0.1074
jge 146 0.0906
setb 145 0.0900
vminps 126 0.0782
vmovsd 119 0.0739
vorps 88 0.0546
js 82 0.0509
vpmaxsd 78 0.0484
vpminsd 78 0.0484
bt 78 0.0484
cmovle 76 0.0472
vpextrq 74 0.0459
vmovd 71 0.0441
cmovz 62 0.0385
bsf 61 0.0379
jns 53 0.0329
vpshufb 51 0.0317
cmovnb 49 0.0304
sar 43 0.0267
vpslld 39 0.0242
vpsrld 38 0.0236
vpor 38 0.0236
vpxor 38 0.0236
xchg 36 0.0223
vpsubd 35 0.0217
vinsertf128 29 0.0180
jg 28 0.0174
vpaddd 25 0.0155
setz 24 0.0149
vmovdqa 22 0.0137
bts 22 0.0137
setnz 20 0.0124
vpcmpgtd 20 0.0124
vpextrd 17 0.0106
div 12 0.0074
vpand 11 0.0068
rol 10 0.0062
vblendps 8 0.0050
neg 8 0.0050
setnbe 7 0.0043
vcvtdq2ps 7 0.0043
vpinsrd 6 0.0037
vcvttps2dq 6 0.0037
cmovs 6 0.0037
btr 6 0.0037
vhaddps 4 0.0025
ror 4 0.0025
cmovnz 4 0.0025
bsr 4 0.0025
vpcmpeqd 4 0.0025
not 2 0.0012
vpsrad 1 0.0006
vmovhlps 1 0.0006
mul 1 0.0006
setnle 1 0.0006
vcvtps2dq 1 0.0006
vandnps 1 0.0006
vpackssdw 1 0.0006
vpandn 1 0.0006
vmovlps 1 0.0006

 

[sirmo]

Intel throws >$12B/year at R&D (and growing), more than 10x what AMD does. .

Yup.. which is all the more reason Ryzen is shaping up to be a small miracle.

 

[Arachnotronic]

At the same time it's pretty clear that Intel is deliberately slowing down to cut costs...

Slowing down what?

 

[scannall]

At the same time it's pretty clear that Intel is deliberately slowing down to cut costs...

Slow down what? Until some amazing new material (Unobtanium and pixie dust?) comes along there really isn't any easy pickings left in x86. The low hanging fruit was picked a long time ago. Then put some back with Netburst. After a DOH! moment they picked that low fruit again. At this point, all they really have is throwing more transistors at things as process shrinks allow. I'm not an Intel 'fan'. Nor a 'fan' of any other company really. Further, I'd love to see x86 die. Won't happen, and I know that. But hey, everyone has to dream?

 

[Arachnotronic]

Slow down what? Until some amazing new material (Unobtanium and pixie dust?) comes along there really isn't any easy pickings left in x86. The low hanging fruit was picked a long time ago. Then put some back with Netburst. After a DOH! moment they picked that low fruit again. At this point, all they really have is throwing more transistors at things as process shrinks allow. I'm not an Intel 'fan'. Nor a 'fan' of any other company really. Further, I'd love to see x86 die. Won't happen, and I know that. But hey, everyone has to dream?

More performance isn't the problem. More performance subject to the constraint of improved energy efficiency is.