AMD Carrizo Pre-release thread

[DrMrLordX]

So many BristolRidge geekbench score appears:
https://browser.primatelabs.com/geekbench3/search?q=AMD+family+21+model+101

Those scores look really weird. And very high given the listed clockspeed.

 

[Dresdenboy]

Those scores look really weird. And very high given the listed clockspeed.

The reported clockspeed for recent AMD CPUs seems to be half of the true one.

 

[The Stilt]

They did in fact improve the power management, where else would the clock increases materialize from? This presentation from the ISSCC this year has more info, for those who have access. For those who don't, here are some of the highlights:

- Reliability Tracker (+100MHz)

- DLDO on the cores for C-State Boost (+6%)

- improved STAPM

- Boot-time calibration for the power supply (saves 30mV)

- Shadow P-States (+100MHz)

Most of these changes were implemented through changes in the SMU Firmware and required no special hardware, except the already existing.

The clock increases don't appear to have materialized from anywhere, since they seem to only exist on paper

FX-9800P seems to score (Geekbench) the same as FX-8800P at the stock TDP (15/25W on both), eventhou it has higher nominal clocks (2.7GHz base, 3.6GHz boost vs. 2.1GHz base, 3.4GHz boost on FX-8800P). This was expected since the clocks are fully limited by the power budget.

 

[DrMrLordX]

The reported clockspeed for recent AMD CPUs seems to be half of the true one.

Hmm okay. There seems to be some kind of a scaling issue then? That 2m/4t chip ("cores: 4") at 3.5 GHz has a ST score of 2174 which is respectable, but an MT score of 5139? That's awful. 59% scaling? I got 78% on my SR in Geekbench 3.

 

[The Stilt]

Hmm okay. There seems to be some kind of a scaling issue then? That 2m/4t chip ("cores: 4") at 3.5 GHz has a ST score of 2174 which is respectable, but an MT score of 5139? That's awful. 59% scaling? I got 78% on my SR in Geekbench 3.

3.6GHz in ST and around 2.5GHz average in MT, due power limitation.

 

[dark zero]

According to Sweclockers, AMD hade activated some features that were inactivated in Carrizo.

They did not say which features, though.

One of them is supposed to be the full L2 cache, other is the HDMI 2.0.

Other thing I am hearing is that the process is different too.

Also... Bristol Ridge entering too early... even to Intel... Kabylake was supposed to give the last shot to AMD... and it failed.

 

[The Stilt]

There is no difference in silicon between CZ and BR, so obviously they both use the same L2 too (1MB per compute unit physically present).

Carrizo / Bristol Ridge / Stoney Ridge are made on GF28HPP.

HDMI 2.0 would be pretty hard, since no APU supports any kind of HDMI (only DP and VGA through FCH)

 

[DrMrLordX]

3.6GHz in ST and around 2.5GHz average in MT, due power limitation.

Oh okay. Geez I hope end-users can adjust or disable the TDP limitations on desktop Bristol Ridge.

 

[Dresdenboy]

Oh okay. Geez I hope end-users can adjust or disable the TDP limitations on desktop Bristol Ridge.

You can see this as a typical high ST turbo mode effect under power constraints, which spoils your MT scaling analysis.

 

[monstercameron]

Oh okay. Geez I hope end-users can adjust or disable the TDP limitations on desktop Bristol Ridge.

Oems wont expose those uefi features...nor will they give it beefy enough cooling.

 

[Dresdenboy]

The clock increases don't appear to have materialized from anywhere, since they seem to only exist on paper

FX-9800P seems to score (Geekbench) the same as FX-8800P at the stock TDP (15/25W on both), eventhou it has higher nominal clocks (2.7GHz base, 3.6GHz boost vs. 2.1GHz base, 3.4GHz boost on FX-8800P). This was expected since the clocks are fully limited by the power budget.

Did you see that for chips running at the same cTDP and mem/mem channel settings? Most FX-8800P GB scores seem to be lower (even with 2ch mem) and some significantly higher (35W then).

 

[el etro]

Look at how the low MC scores impact on Carrizo GB result. AMD needs to improve the CON cache and memory performance by 100% or something like this. Looks like the memory performance is the number one bottleneck for its cores.

 

[DrMrLordX]

You can see this as a typical high ST turbo mode effect under power constraints, which spoils your MT scaling analysis.

Yeah, I get it now.

Oems wont expose those uefi features...nor will they give it beefy enough cooling.

If they don't, they're idiots. Cmon man, I'm willing to pony up for the next generation of A88x-Pro or whatever it is that comes down the pike that can let Bristol Ridge strut its stuff. I'd actually like to have one of those chips, but if it's going to be that TDP-constrained without any ability to overcome that behavior then I'm out. I'll wait for Raven Ridge or . . . whatever.

Look at how the low MC scores impact on Carrizo GB result. AMD needs to improve the CON cache and memory performance by 100% or something like this. Looks like the memory performance is the number one bottleneck for its cores.

Actually I don't think we can conclude that per se. Look at the SuperPi results for the x4 845 to see what I mean. As others have said, it looks like Bristol Ridge is hitting TDP limits running Geekbench 3.

 

[monstercameron]

Yeah, I get it now.



If they don't, they're idiots. Cmon man, I'm willing to pony up for the next generation of A88x-Pro or whatever it is that comes down the pike that can let Bristol Ridge strut its stuff. I'd actually like to have one of those chips, but if it's going to be that TDP-constrained without any ability to overcome that behavior then I'm out. I'll wait for Raven Ridge or . . . whatever.



Actually I don't think we can conclude that per se. Look at the SuperPi results for the x4 845 to see what I mean. As others have said, it looks like Bristol Ridge is hitting TDP limits running Geekbench 3.

I was referring to the mobile platforms, who knows how they would treat the desktop.

 

[DrMrLordX]

I was referring to the mobile platforms, who knows how they would treat the desktop.

Oh, okay. Well color me a bit disappointed but I do expect some throttling in mobile platforms. Maybe I'm just jaded.

 

[The Stilt]

Did you see that for chips running at the same cTDP and mem/mem channel settings? Most FX-8800P GB scores seem to be lower (even with 2ch mem) and some significantly higher (35W then).

Here's a FX-8800P result at the stock settings: 15/25W TDP limit, 100% STAPM scalar, 200s STAPM duration, DDR-1600 dual rank, dual channel.

Tested today, on the AMD reference system, using the same version of Geekbench as on FX-9800P results (x86 too).

The FX-9800P results are definitely single channel, but DDR-1866 instead of 1600MHz. AFAIK 9800P does not support higher than 15W cTDP (i.e cTDP can be only lower than stock, unlike on Carrizo) but is supports DDR-1866 at the default TDP. On Carrizo higher than DDR-1600 required increasing the cTDP: >= 16W for 1866MHz & 20W for 2133MHz. FX-9800P doesn't support higher than 1866MHz even with hacked cTDP limit.

Benchmark Summary
  Integer Score              2493   7556
  Floating Point Score       2156   5972
  Memory Score               1941   2692

  Geekbench Score            2247   5949

System Information
  Operating System      Microsoft Windows 7 Ultimate (64-bit)
  Model                 Standard LCL Series
  Motherboard           Standard LCL Series
  Processor             AMD FX-8800P Radeon R7, 12 Compute Cores 4C+8G @ 1.70 GHz
                        1 Processor, 2 Cores, 4 Threads
  Processor ID          AuthenticAMD Family 21 Model 96 Stepping 1
  Processor Package     
  Processor Codename    Carrizo
  L1 Instruction Cache  96.0 KB x 2
  L1 Data Cache         32.0 KB x 4
  L2 Cache              1.00 MB x 2
  L3 Cache              0.00 B
  Memory                7.70 GB  -1MHz
  Northbridge           AMD ID1576 00
  Southbridge           AMD ID790E 11
  BIOS                  American Megatrends Inc. 5.011
  Compiler              Visual C++ 170061030

Integer Performance
  AES
    single-core          2954 ||||||||||
    multi-core           5900 |||||||||||||||||||||
  Twofish
    single-core          2798 ||||||||||
    multi-core           9441 |||||||||||||||||||||||||||||||||||
  SHA1
    single-core          2816 ||||||||||
    multi-core           9366 ||||||||||||||||||||||||||||||||||
  SHA2
    single-core          3369 ||||||||||||
    multi-core          10758 ||||||||||||||||||||||||||||||||||||||||
  BZip2 Compress
    single-core          2276 ||||||||
    multi-core           6822 |||||||||||||||||||||||||
  BZip2 Decompress
    single-core          2555 |||||||||
    multi-core           7850 |||||||||||||||||||||||||||||
  JPEG Compress
    single-core          2594 |||||||||
    multi-core           8372 |||||||||||||||||||||||||||||||
  JPEG Decompress
    single-core          2978 |||||||||||
    multi-core           9657 |||||||||||||||||||||||||||||||||||
  PNG Compress
    single-core          2228 ||||||||
    multi-core           7035 ||||||||||||||||||||||||||
  PNG Decompress
    single-core          2187 ||||||||
    multi-core           6998 ||||||||||||||||||||||||||
  Sobel
    single-core          2106 |||||||
    multi-core           6591 ||||||||||||||||||||||||
  Lua
    single-core          2224 ||||||||
    multi-core           6728 |||||||||||||||||||||||||
  Dijkstra
    single-core          1794 ||||||
    multi-core           4939 ||||||||||||||||||

Floating Point Performance
  BlackScholes
    single-core          2650 |||||||||
    multi-core           7614 ||||||||||||||||||||||||||||
  Mandelbrot
    single-core          2265 ||||||||
    multi-core           7638 ||||||||||||||||||||||||||||
  Sharpen Filter
    single-core          2065 |||||||
    multi-core           5420 ||||||||||||||||||||
  Blur Filter
    single-core          1951 |||||||
    multi-core           4841 |||||||||||||||||
  SGEMM
    single-core          1747 ||||||
    multi-core           5233 |||||||||||||||||||
  DGEMM
    single-core          1688 ||||||
    multi-core           4842 ||||||||||||||||||
  SFFT
    single-core          2575 |||||||||
    multi-core           5358 |||||||||||||||||||
  DFFT
    single-core          2280 ||||||||
    multi-core           5590 ||||||||||||||||||||
  N-Body
    single-core          2085 |||||||
    multi-core           6531 ||||||||||||||||||||||||
  Ray Trace
    single-core          2489 |||||||||
    multi-core           7638 ||||||||||||||||||||||||||||

Memory Performance
  Stream Copy
    single-core          2036 |||||||
    multi-core           2771 ||||||||||
  Stream Scale
    single-core          1850 ||||||
    multi-core           2668 |||||||||
  Stream Add
    single-core          1910 |||||||
    multi-core           2626 |||||||||
  Stream Triad
    single-core          1974 |||||||
    multi-core           2708 ||||||||||

 

[Dresdenboy]

Here's a FX-8800P result at the stock settings: 15/25W TDP limit, 100% STAPM scalar, 200s STAPM duration, DDR-1600 dual rank, dual channel.

Tested today, on the AMD reference system, using the same version of Geekbench as on FX-9800P results (x86 too).

The FX-9800P results are definitely single channel, but DDR-1866 instead of 1600MHz. AFAIK 9800P does not support higher than 15W cTDP (i.e cTDP can be only lower than stock, unlike on Carrizo) but is supports DDR-1866 at the default TDP. On Carrizo higher than DDR-1600 required increasing the cTDP: >= 16W for 1866MHz & 20W for 2133MHz. FX-9800P doesn't support higher than 1866MHz even with hacked cTDP limit.

Memory Performance
  Stream Copy
    single-core          2036 |||||||
    multi-core           2771 ||||||||||
  Stream Scale
    single-core          1850 ||||||
    multi-core           2668 |||||||||
  Stream Add
    single-core          1910 |||||||
    multi-core           2626 |||||||||
  Stream Triad
    single-core          1974 |||||||
    multi-core           2708 ||||||||||

Your mem scores are really high, like for example these likely 2ch ones. There are lots of Carrizo mem scores in the 1xxx range (speeds of 4-5 GB/s) and many, which are clearly >2k, with speeds of 8 to 11 GB/s. Your scores seem to belong to the latter. I don't think, that DDR3-1866 and a related cTDP adaption is enough to explain this.

The AES MT score seems to correlate the most with the mem score. So does yours vs. the BR one.

If I compare the BR to FX-8800P results, which don't look like 35W ones (MT>6k), and which have very similar memory scores, there are significant improvements:
https://browser.primatelabs.com/geekbench3/compare/4384049?baseline=5792324
https://browser.primatelabs.com/geekbench3/compare/4259446?baseline=5792324

If we'd extract all FX scores from the GB database (if it's worth the hassle), we might see clusters of 15W/1ch, 15W/2ch, 35W/1ch, 35W/2ch, 64b, 32b, bad cooling, x/y, etc.

 

[USER8000]

Some tests have been released:

http://www.hardwarecanucks.com/foru...md-details-bristol-ridge-am4-performance.html
http://semiaccurate.com/2016/04/05/amd-announces-its-7th-generation-of-a-series-apus/

 

[Dresdenboy]

Some tests have been released:

http://www.hardwarecanucks.com/foru...md-details-bristol-ridge-am4-performance.html
http://semiaccurate.com/2016/04/05/amd-announces-its-7th-generation-of-a-series-apus/

Basically the same numbers, taken from AMD's footnotes, but in a more convinient way to read.

 

[dark zero]

Seems that BR won't be as competitive as expecting... hoping that in desktops, they have the full 4 MB L2 cache, it would help a little.

 

[Dresdenboy]

Seems that BR won't be as competitive as expecting... hoping that in desktops, they have the full 4 MB L2 cache, it would help a little.

They don't. It's CZ silicon.

 

[The Stilt]

Your mem scores are really high, like for example these likely 2ch ones. There are lots of Carrizo mem scores in the 1xxx range (speeds of 4-5 GB/s) and many, which are clearly >2k, with speeds of 8 to 11 GB/s. Your scores seem to belong to the latter. I don't think, that DDR3-1866 and a related cTDP adaption is enough to explain this.

The AES MT score seems to correlate the most with the mem score. So does yours vs. the BR one.

If I compare the BR to FX-8800P results, which don't look like 35W ones (MT>6k), and which have very similar memory scores, there are significant improvements:
https://browser.primatelabs.com/geekbench3/compare/4384049?baseline=5792324
https://browser.primatelabs.com/geekbench3/compare/4259446?baseline=5792324

If we'd extract all FX scores from the GB database (if it's worth the hassle), we might see clusters of 15W/1ch, 15W/2ch, 35W/1ch, 35W/2ch, 64b, 32b, bad cooling, x/y, etc.

Removed the second memory module and repeated the test:

15/25W, 200s / 100% STAPM, DDR-1600 DR Single Channel:

Benchmark Summary
  Integer Score              2414   7204
  Floating Point Score       2176   6090
  Memory Score               1337   1408

  Geekbench Score            2103   5599

System Information
  Operating System      Microsoft Windows 7 Ultimate (64-bit)
  Model                 Standard LCL Series
  Motherboard           Standard LCL Series
  Processor             AMD FX-8800P Radeon R7, 12 Compute Cores 4C+8G @ 1.70 GHz
                        1 Processor, 2 Cores, 4 Threads
  Processor ID          AuthenticAMD Family 21 Model 96 Stepping 1
  Processor Package     
  Processor Codename    Carrizo
  L1 Instruction Cache  96.0 KB x 2
  L1 Data Cache         32.0 KB x 4
  L2 Cache              1.00 MB x 2
  L3 Cache              0.00 B
  Memory                3.70 GB  -1MHz
  Northbridge           AMD ID1576 00
  Southbridge           AMD ID790E 11
  BIOS                  American Megatrends Inc. 5.011
  Compiler              Visual C++ 170061030

Integer Performance
  AES
    single-core          1968 |||||||
    multi-core           3254 |||||||||||
  Twofish
    single-core          2794 ||||||||||
    multi-core           9569 |||||||||||||||||||||||||||||||||||
  SHA1
    single-core          2813 ||||||||||
    multi-core           9579 |||||||||||||||||||||||||||||||||||
  SHA2
    single-core          3361 ||||||||||||
    multi-core          10925 ||||||||||||||||||||||||||||||||||||||||
  BZip2 Compress
    single-core          2285 ||||||||
    multi-core           6923 |||||||||||||||||||||||||
  BZip2 Decompress
    single-core          2552 |||||||||
    multi-core           8026 |||||||||||||||||||||||||||||
  JPEG Compress
    single-core          2575 |||||||||
    multi-core           8423 ||||||||||||||||||||||||||||||
  JPEG Decompress
    single-core          2986 ||||||||||
    multi-core           9384 ||||||||||||||||||||||||||||||||||
  PNG Compress
    single-core          2227 ||||||||
    multi-core           7051 |||||||||||||||||||||||||
  PNG Decompress
    single-core          2170 |||||||
    multi-core           6790 ||||||||||||||||||||||||
  Sobel
    single-core          2109 |||||||
    multi-core           6507 |||||||||||||||||||||||
  Lua
    single-core          2264 ||||||||
    multi-core           7111 ||||||||||||||||||||||||||
  Dijkstra
    single-core          1763 ||||||
    multi-core           4446 ||||||||||||||||

Floating Point Performance
  BlackScholes
    single-core          2698 |||||||||
    multi-core           7756 ||||||||||||||||||||||||||||
  Mandelbrot
    single-core          2265 ||||||||
    multi-core           7718 ||||||||||||||||||||||||||||
  Sharpen Filter
    single-core          2061 |||||||
    multi-core           5439 |||||||||||||||||||
  Blur Filter
    single-core          1948 |||||||
    multi-core           4920 ||||||||||||||||||
  SGEMM
    single-core          1747 ||||||
    multi-core           5277 |||||||||||||||||||
  DGEMM
    single-core          1695 ||||||
    multi-core           4919 ||||||||||||||||||
  SFFT
    single-core          2543 |||||||||
    multi-core           5531 ||||||||||||||||||||
  DFFT
    single-core          2471 |||||||||
    multi-core           5786 |||||||||||||||||||||
  N-Body
    single-core          2089 |||||||
    multi-core           6651 ||||||||||||||||||||||||
  Ray Trace
    single-core          2500 |||||||||
    multi-core           7938 |||||||||||||||||||||||||||||

Memory Performance
  Stream Copy
    single-core          1349 ||||
    multi-core           1400 |||||
  Stream Scale
    single-core          1333 ||||
    multi-core           1409 |||||
  Stream Add
    single-core          1321 ||||
    multi-core           1389 |||||
  Stream Triad
    single-core          1349 ||||
    multi-core           1435 |||||

15/25W, 200s / 100% STAPM, DDR-1866 DR Single Channel:

Benchmark Summary
  Integer Score              2446   7295
  Floating Point Score       2113   6026
  Memory Score               1446   1508

  Geekbench Score            2112   5630

System Information
  Operating System      Microsoft Windows 7 Ultimate (64-bit)
  Model                 Standard LCL Series
  Motherboard           Standard LCL Series
  Processor             AMD FX-8800P Radeon R7, 12 Compute Cores 4C+8G @ 1.70 GHz
                        1 Processor, 2 Cores, 4 Threads
  Processor ID          AuthenticAMD Family 21 Model 96 Stepping 1
  Processor Package     
  Processor Codename    Carrizo
  L1 Instruction Cache  96.0 KB x 2
  L1 Data Cache         32.0 KB x 4
  L2 Cache              1.00 MB x 2
  L3 Cache              0.00 B
  Memory                3.70 GB  -1MHz
  Northbridge           AMD ID1576 00
  Southbridge           AMD ID790E 11
  BIOS                  American Megatrends Inc. 5.011
  Compiler              Visual C++ 170061030

Integer Performance
  AES
    single-core          2128 |||||||
    multi-core           3400 ||||||||||||
  Twofish
    single-core          2824 ||||||||||
    multi-core           9500 ||||||||||||||||||||||||||||||||||
  SHA1
    single-core          2818 ||||||||||
    multi-core           9620 ||||||||||||||||||||||||||||||||||
  SHA2
    single-core          3372 ||||||||||||
    multi-core          11045 ||||||||||||||||||||||||||||||||||||||||
  BZip2 Compress
    single-core          2289 ||||||||
    multi-core           6958 |||||||||||||||||||||||||
  BZip2 Decompress
    single-core          2527 |||||||||
    multi-core           8015 |||||||||||||||||||||||||||||
  JPEG Compress
    single-core          2604 |||||||||
    multi-core           8447 ||||||||||||||||||||||||||||||
  JPEG Decompress
    single-core          2991 ||||||||||
    multi-core           9692 |||||||||||||||||||||||||||||||||||
  PNG Compress
    single-core          2226 ||||||||
    multi-core           6996 |||||||||||||||||||||||||
  PNG Decompress
    single-core          2190 |||||||
    multi-core           7130 |||||||||||||||||||||||||
  Sobel
    single-core          2149 |||||||
    multi-core           6741 ||||||||||||||||||||||||
  Lua
    single-core          2244 ||||||||
    multi-core           6882 ||||||||||||||||||||||||
  Dijkstra
    single-core          1860 ||||||
    multi-core           4570 ||||||||||||||||

Floating Point Performance
  BlackScholes
    single-core          2254 ||||||||
    multi-core           7685 |||||||||||||||||||||||||||
  Mandelbrot
    single-core          2265 ||||||||
    multi-core           7683 |||||||||||||||||||||||||||
  Sharpen Filter
    single-core          2075 |||||||
    multi-core           5453 |||||||||||||||||||
  Blur Filter
    single-core          1953 |||||||
    multi-core           4884 |||||||||||||||||
  SGEMM
    single-core          1752 ||||||
    multi-core           5244 ||||||||||||||||||
  DGEMM
    single-core          1712 ||||||
    multi-core           4938 |||||||||||||||||
  SFFT
    single-core          2454 ||||||||
    multi-core           5355 |||||||||||||||||||
  DFFT
    single-core          2216 ||||||||
    multi-core           5645 ||||||||||||||||||||
  N-Body
    single-core          2090 |||||||
    multi-core           6607 |||||||||||||||||||||||
  Ray Trace
    single-core          2523 |||||||||
    multi-core           7769 ||||||||||||||||||||||||||||

Memory Performance
  Stream Copy
    single-core          1475 |||||
    multi-core           1521 |||||
  Stream Scale
    single-core          1441 |||||
    multi-core           1518 |||||
  Stream Add
    single-core          1424 |||||
    multi-core           1473 |||||
  Stream Triad
    single-core          1445 |||||
    multi-core           1524 |||||

Still I'm not impressed.

After all they claim 28.6% increase in base clock on FX-9800P (2.7GHz base) compared to FX-8800P (2.1GHz base).

 

[AtenRa]

Just realized that 35W Bristol at 28nm HDL has the same performance as the 22nm Haswell Core i3 4000M at 37W TDP in Cinebench R15 both Single and Multi.

http://www.hardwarecanucks.com/foru...md-details-bristol-ridge-am4-performance.html

Core i3 3110M 35W TDP 22nm

http://www.notebookcheck.net/Intel-Core-i3-3110M-Notebook-Processor.74459.0.html
Cinebench R15
Single = 81,5
Multi = 209

Core i3 4000M 37W TDP 22nm

http://www.notebookcheck.net/Intel-Core-i3-4000M-Notebook-Processor.93560.0.html
Cinebench R15
Single = 94,2
Multi = 239,8

 

[Abwx]

Just realized that 35W Bristol at 28nm HDL has the same performance as the 22nm Haswell Core i3 4000M at 37W TDP in Cinebench R15 both Single and Multi.

http://www.hardwarecanucks.com/foru...md-details-bristol-ridge-am4-performance.html

Core i3 3110M 35W TDP 22nm

http://www.notebookcheck.net/Intel-Core-i3-3110M-Notebook-Processor.74459.0.html
Cinebench R15
Single = 81,5
Multi = 209

Core i3 4000M 37W TDP 22nm

http://www.notebookcheck.net/Intel-Core-i3-4000M-Notebook-Processor.93560.0.html
Cinebench R15
Single = 94,2
Multi = 239,8

Just that the Bristol ridge score are not for 35W, for whatever reason this site did put 35W on the CB graphs while it s 15W for other graphs...

Edit : It s at 15W and it s impossible that they missed it since they have the slides on the note 6 :

http://wccftech.com/amd-bristol-ridge-apu-computex-launch/

 

[AtenRa]

Just that the Bristol ridge score are not for 35W, for whatever reason this site did put 35W on the CB graphs while it s 15W for other graphs...

Those are 15W TDP scores and not 35W TDP ???

Do we have a link for that ??