Overclock.net › Forums › Industry News › Rumors and Unconfirmed Articles › [WCCF]AMD On Kaveri APU x86 Performance Against Intel Haswell,”We’ll Lose” – Flagship APU To Cost Around $150
New Posts  All Forums:Forum Nav:

[WCCF]AMD On Kaveri APU x86 Performance Against Intel Haswell,”We’ll Lose” – Flagship APU To Cost Around $150 - Page 5

post #41 of 230
This is only mind-blowing news (on the x86 front) to people who had unrealistic expectations to begin with. There was no way they would magically catch up to Haswell's performance in one leap.
post #42 of 230
Quote:
Originally Posted by NaroonGTX View Post

This is only mind-blowing news (on the x86 front) to people who had unrealistic expectations to begin with. There was no way they would magically catch up to Haswell's performance in one leap.

Actually ,I have big expectation of that leaked module from last spring. I know seronx insists it is SR but (unless AMD is trolling us when telling us what to expect) this level of performance is underwhelming for such a design. So I expect it to be EX instead.
Mastodon
(11 items)
 
HP Z220
(8 items)
 
 
CPUMotherboardGraphicsRAM
FX 8320 M5A99FX Pro 2.0 Asus R9 290X DirectCU II Crucial Ballistix 
Hard DriveCoolingOSMonitor
Pny SSD 240GB CM Nepton 280L Win 8.1 Acer Predator XG270HU Freesync 
PowerCaseMouse
XFX 750W Pro HAF XM Logitech G300 
CPUMotherboardGraphicsCooling
i7 3770 HP Quadro K2000 HP 
OSPowerCaseMouse
Win 7  HP 400W HP CMT RAT 7 
  hide details  
Reply
Mastodon
(11 items)
 
HP Z220
(8 items)
 
 
CPUMotherboardGraphicsRAM
FX 8320 M5A99FX Pro 2.0 Asus R9 290X DirectCU II Crucial Ballistix 
Hard DriveCoolingOSMonitor
Pny SSD 240GB CM Nepton 280L Win 8.1 Acer Predator XG270HU Freesync 
PowerCaseMouse
XFX 750W Pro HAF XM Logitech G300 
CPUMotherboardGraphicsCooling
i7 3770 HP Quadro K2000 HP 
OSPowerCaseMouse
Win 7  HP 400W HP CMT RAT 7 
  hide details  
Reply
post #43 of 230
That module has a lot of doubled-up units and logic which isn't in SR-B, so it could only be EX or something that succeeds it (unless AMD has been trolling with these controlled leaks and suddenly SR-B has doubled up existence, though I doubt this lol.)
post #44 of 230
Quote:
Originally Posted by Kuivamaa View Post

Actually ,I have big expectation of that leaked module from last spring. I know seronx insists it is SR but (unless AMD is trolling us when telling us what to expect) this level of performance is underwhelming for such a design. So I expect it to be EX instead.

What level of performance? We don't even have reliable benchmarks.

I was expecting performance somewhere between Ivy i3 and locked i5's (in general). People who were expecting more are being unrealistic. Of course it'll be interesting to see how hybrid Crossfire performs and what developers cook up using HSA and Mantle.
post #45 of 230
Quote:
Originally Posted by MrJava View Post

What level of performance? We don't even have reliable benchmarks.

I was expecting performance somewhere between Ivy i3 and locked i5's (in general). People who were expecting more are being unrealistic. Of course it'll be interesting to see how hybrid Crossfire performs and what developers cook up using HSA and Mantle.

Kozak said said "we'll lose" when asked about the cpu part of kaveri.That leaked die shot had doubled up cache/AGU/ALU/FMA (a hint of AVX2 support? someone more knowledgable could comment) so I expect a processor of that class to match or even surpass a haswell quad (non hyperthreaded) on certain MT workloads (If we accept there is no ST regression).
Mastodon
(11 items)
 
HP Z220
(8 items)
 
 
CPUMotherboardGraphicsRAM
FX 8320 M5A99FX Pro 2.0 Asus R9 290X DirectCU II Crucial Ballistix 
Hard DriveCoolingOSMonitor
Pny SSD 240GB CM Nepton 280L Win 8.1 Acer Predator XG270HU Freesync 
PowerCaseMouse
XFX 750W Pro HAF XM Logitech G300 
CPUMotherboardGraphicsCooling
i7 3770 HP Quadro K2000 HP 
OSPowerCaseMouse
Win 7  HP 400W HP CMT RAT 7 
  hide details  
Reply
Mastodon
(11 items)
 
HP Z220
(8 items)
 
 
CPUMotherboardGraphicsRAM
FX 8320 M5A99FX Pro 2.0 Asus R9 290X DirectCU II Crucial Ballistix 
Hard DriveCoolingOSMonitor
Pny SSD 240GB CM Nepton 280L Win 8.1 Acer Predator XG270HU Freesync 
PowerCaseMouse
XFX 750W Pro HAF XM Logitech G300 
CPUMotherboardGraphicsCooling
i7 3770 HP Quadro K2000 HP 
OSPowerCaseMouse
Win 7  HP 400W HP CMT RAT 7 
  hide details  
Reply
post #46 of 230
Quote:
Originally Posted by Kuivamaa View Post

Kozak said said "we'll lose" when asked about the cpu part of kaveri.That leaked die shot had doubled up cache/AGU/ALU/FMA (a hint of AVX2 support? someone more knowledgable could comment) so I expect a processor of that class to match or even surpass a haswell quad (non hyperthreaded) on certain MT workloads (If we accept there is no ST regression).

You can be underwhelmed by anything if you don't keep expectations in check. I have no idea where the leaked die shot came from, but it could easily be a fake.
Keep in mind that the goal of Bulldozer was to increase multithreaded throughput per mm^2 in modern applications while giving up as little single threaded performance as possible. If you go and wildly double everything in the module, this doesn't achieve that goal. In fact, you'd probably be better off doing separate cores at that point.

I'm still waiting on a SWOG (software optimization guide) to find out all the juicy tidbits about what exactly has changed in Steamroller. I'm told its coming soon.
post #47 of 230
Quote:
Originally Posted by MrJava View Post

You can be underwhelmed by anything if you don't keep expectations in check. I have no idea where the leaked die shot came from, but it could easily be a fake.
Keep in mind that the goal of Bulldozer was to increase multithreaded throughput per mm^2 in modern applications while giving up as little single threaded performance as possible. If you go and wildly double everything in the module, this doesn't achieve that goal. In fact, you'd probably be better off doing separate cores at that point.

I'm still waiting on a SWOG (software optimization guide) to find out all the juicy tidbits about what exactly has changed in Steamroller. I'm told its coming soon.

http://semiaccurate.com/forums/showpost.php?p=184500&postcount=932

People with great expertise find it legit.
Mastodon
(11 items)
 
HP Z220
(8 items)
 
 
CPUMotherboardGraphicsRAM
FX 8320 M5A99FX Pro 2.0 Asus R9 290X DirectCU II Crucial Ballistix 
Hard DriveCoolingOSMonitor
Pny SSD 240GB CM Nepton 280L Win 8.1 Acer Predator XG270HU Freesync 
PowerCaseMouse
XFX 750W Pro HAF XM Logitech G300 
CPUMotherboardGraphicsCooling
i7 3770 HP Quadro K2000 HP 
OSPowerCaseMouse
Win 7  HP 400W HP CMT RAT 7 
  hide details  
Reply
Mastodon
(11 items)
 
HP Z220
(8 items)
 
 
CPUMotherboardGraphicsRAM
FX 8320 M5A99FX Pro 2.0 Asus R9 290X DirectCU II Crucial Ballistix 
Hard DriveCoolingOSMonitor
Pny SSD 240GB CM Nepton 280L Win 8.1 Acer Predator XG270HU Freesync 
PowerCaseMouse
XFX 750W Pro HAF XM Logitech G300 
CPUMotherboardGraphicsCooling
i7 3770 HP Quadro K2000 HP 
OSPowerCaseMouse
Win 7  HP 400W HP CMT RAT 7 
  hide details  
Reply
post #48 of 230
Quote:
Originally Posted by Kuivamaa View Post

http://semiaccurate.com/forums/showpost.php?p=184500&postcount=932

People with great expertise find it legit.

Hans is very smart. But smart people can be fooled too. wink.gif

He interprets it is have 2 256-bit pipes. I would expect Excavator to have 4 128-bit pipes instead. But I could be wrong.
post #49 of 230
Wish I could identify more than integer parts and cache personally ,die shots are largely a mystery to me biggrin.gif 4 pipes per module you say? 2 pipes per core, if they can be "unified" for AVX duties as 256 units per core, wouldn't that be doubled up over what BD/PD has?

Edit.
http://browser.primatelabs.com/geekbench3/223722

L1 dcache=16KBx4 , further indicating that this die shot isn't Steamroller. I think we are looking at "Carrizo".
Edited by Kuivamaa - 11/30/13 at 1:59pm
Mastodon
(11 items)
 
HP Z220
(8 items)
 
 
CPUMotherboardGraphicsRAM
FX 8320 M5A99FX Pro 2.0 Asus R9 290X DirectCU II Crucial Ballistix 
Hard DriveCoolingOSMonitor
Pny SSD 240GB CM Nepton 280L Win 8.1 Acer Predator XG270HU Freesync 
PowerCaseMouse
XFX 750W Pro HAF XM Logitech G300 
CPUMotherboardGraphicsCooling
i7 3770 HP Quadro K2000 HP 
OSPowerCaseMouse
Win 7  HP 400W HP CMT RAT 7 
  hide details  
Reply
Mastodon
(11 items)
 
HP Z220
(8 items)
 
 
CPUMotherboardGraphicsRAM
FX 8320 M5A99FX Pro 2.0 Asus R9 290X DirectCU II Crucial Ballistix 
Hard DriveCoolingOSMonitor
Pny SSD 240GB CM Nepton 280L Win 8.1 Acer Predator XG270HU Freesync 
PowerCaseMouse
XFX 750W Pro HAF XM Logitech G300 
CPUMotherboardGraphicsCooling
i7 3770 HP Quadro K2000 HP 
OSPowerCaseMouse
Win 7  HP 400W HP CMT RAT 7 
  hide details  
Reply
post #50 of 230
Quote:
Originally Posted by Kuivamaa View Post

Wish I could identify more than integer parts and cache personally ,die shots are largely a mystery to me biggrin.gif 4 pipes per module you say? 2 pipes per core, if they can be "unified" for AVX duties as 256 units per core, wouldn't that be doubled up over what BD/PD has?

That's the idea. You'll only gain advantage of BD/PD/Steamroller with 2 256-bit pipes if all the code is AVX2. With 4 128-bit pipes (which can fuse for 256-bit ops) you'll get higher throughput for AVX and SSE code as well.

He also mentioned 4 ALU/AGU is overkill which is probably true in most situations.
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Rumors and Unconfirmed Articles
Overclock.net › Forums › Industry News › Rumors and Unconfirmed Articles › [WCCF]AMD On Kaveri APU x86 Performance Against Intel Haswell,”We’ll Lose” – Flagship APU To Cost Around $150