Overclock.net › Forums › Industry News › Hardware News › [anand] NVIDIA Logan SoC: < 1W Kepler, 1H 2014, More Energy Efficient than A6X?
New Posts  All Forums:Forum Nav:

[anand] NVIDIA Logan SoC: < 1W Kepler, 1H 2014, More Energy Efficient than A6X? - Page 4

post #31 of 38
Quote:
Originally Posted by Alatar View Post

And Anand seems to disagree on all points. Apparently he actually did the math and the numbers check out pretty well...

FROM ANANDTECH ARTICLE

http://www.anandtech.com/show/7169/nvidia-demonstrates-logan-soc-mobile-kepler/2

Isolating GPU power alone, NVIDIA measured around 900mW for Logan's Kepler implementation running at iPad 4 performance levels (potentially as little as 1/5 of Logan's peak performance). NVIDIA also attempted to find and isolate the GPU power rail going into Apple's A6X (using a similar approach to what we documented here), and came up with an average GPU power value of around 2.6W.

http://www.anandtech.com/show/6426/ipad-4-gpu-performance-analyzed-powervr-sgx-554mp4-under-the-hood

A6X GPU = 76.8 GFLOPS

so to achieve 76.8 GFLOPS Nvidia was drawing 900 mw (close to 1W) . to achieve that perf the GPU must have been running at 200 mhz( 200 Mhz x 2 x 192 = 76.8 GFLOPS) . its well known that power scales exponentially with higher clocks (which often require higher voltages). even assuming linear power increase with linear clock scaling you are talking of 2.7w for a 600 mhz Logan GPU ( 3 x 0.9w = 2.7w) . frankly thats just not going to happen. the most I see is 500 - 550 mhz at less than 3W for GPU. so yeah 500 Mhz and 200 GFLOPS is realistic. 1 Ghz is not happening even in a tablet form factor. thumb.gif
Fragbox
(14 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i5 2400 DH67BL AMD Radeon HD 6900 Series XFX HD 6950 2GB 
RAMRAMHard DriveOptical Drive
Corsair Corsair DDR3 1333 Mhz 2 GB Corsair DDR3 1333 Mhz 2 GB Western Digital Caviar Green SONY DVD-RW AD-7260S 
OSMonitorKeyboardPower
Windows 7 Professional 32 bit BENQ G2420HD Logitech K200 Seasonic VX550 psu 
CaseMouse
ANTEC 200 V2 Logitech mouse 
  hide details  
Reply
Fragbox
(14 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i5 2400 DH67BL AMD Radeon HD 6900 Series XFX HD 6950 2GB 
RAMRAMHard DriveOptical Drive
Corsair Corsair DDR3 1333 Mhz 2 GB Corsair DDR3 1333 Mhz 2 GB Western Digital Caviar Green SONY DVD-RW AD-7260S 
OSMonitorKeyboardPower
Windows 7 Professional 32 bit BENQ G2420HD Logitech K200 Seasonic VX550 psu 
CaseMouse
ANTEC 200 V2 Logitech mouse 
  hide details  
Reply
post #32 of 38
Thread Starter 
And then after the power consumption part anand then says this:
Quote:
If these numbers are believable, you could see roughly 2x the performance of an iPad 4 in a Logan based smartphone, and 4 - 5x the performance of an iPad 4 in a Logan tablet - in as little as 12 months from now if NVIDIA can ship this thing on time.

And then on the first page there's this:
Quote:
That's certainly doable from an architectural standpoint (although we've never seen it done on any low power 28nm process), but it's probably a bit too high for something like a smartphone.

NVIDIA didn't want to talk frequencies but they did tell me that we might see something this fast in some sort of a tablet.

Now how you ask?

The answer lies in what they were benching here;
Quote:
NVIDIA got Logan silicon back from the fabs around 3 weeks ago, making it almost certain that we're dealing with some form of 28nm silicon here and not early 20nm samples.

First round of silicon. Almost a year from release. Absolutely no extra steppings or respins and already matching A6X at 900mW. Guess what happens when they refine the chips with tsmc for 9 months or so?
 
Benching
(17 items)
 
 
CPUMotherboardGraphicsRAM
[i7 5960X @ 4.8GHz] [Rampage V Extreme] [Titan 1400MHz (1500MHz bench)] [Various] 
Hard DriveCoolingCoolingCooling
[250GB 840EVO +2x SpinpointF3 1TB RAID0] [LD PC-V2 SS Phase Change] [XSPC X2O 750 pump/res] [Monsta 360 full copper + EK XT 360 + XT 240] 
MonitorPowerCaseAudio
[Crossover 27Q LED-P 1440p+ASUS 1200p+LG 1080p] [Corsair AX1200] [Dimastech Easy v3.0] [Sennheiser HD558s] 
CPUCPUMotherboardGraphics
FX 8320, FX 8350, Phenom II x2 555BE i7 3930K, i7 860, i7 4770K, 68x Celeron D CVF, commando, 2x RIVE, Z87X-OC Asus 4870x2, Sapphire 4870 
GraphicsGraphicsGraphicsGraphics
2x 5870, 5850, 5830, 5770 2x 3870x2, 3870 GTX Titan, GTX 480, GTX 590 GTX 285, GTX 260, 4x 9800GT, 8800GTX 
RAMHard DriveCoolingCooling
4x4GB vengeance, 2x4GB predatorX, 2x1GB OCZ DDR2 Intel X25-M 80GB LD PC-V2 SS Phase Change OCN Marksman 
CoolingCoolingOSPower
2x old tek slims (GPU) Various watercooling stuff win7, winxp AX1200 
Case
test bench / cardboard box 
  hide details  
Reply
 
Benching
(17 items)
 
 
CPUMotherboardGraphicsRAM
[i7 5960X @ 4.8GHz] [Rampage V Extreme] [Titan 1400MHz (1500MHz bench)] [Various] 
Hard DriveCoolingCoolingCooling
[250GB 840EVO +2x SpinpointF3 1TB RAID0] [LD PC-V2 SS Phase Change] [XSPC X2O 750 pump/res] [Monsta 360 full copper + EK XT 360 + XT 240] 
MonitorPowerCaseAudio
[Crossover 27Q LED-P 1440p+ASUS 1200p+LG 1080p] [Corsair AX1200] [Dimastech Easy v3.0] [Sennheiser HD558s] 
CPUCPUMotherboardGraphics
FX 8320, FX 8350, Phenom II x2 555BE i7 3930K, i7 860, i7 4770K, 68x Celeron D CVF, commando, 2x RIVE, Z87X-OC Asus 4870x2, Sapphire 4870 
GraphicsGraphicsGraphicsGraphics
2x 5870, 5850, 5830, 5770 2x 3870x2, 3870 GTX Titan, GTX 480, GTX 590 GTX 285, GTX 260, 4x 9800GT, 8800GTX 
RAMHard DriveCoolingCooling
4x4GB vengeance, 2x4GB predatorX, 2x1GB OCZ DDR2 Intel X25-M 80GB LD PC-V2 SS Phase Change OCN Marksman 
CoolingCoolingOSPower
2x old tek slims (GPU) Various watercooling stuff win7, winxp AX1200 
Case
test bench / cardboard box 
  hide details  
Reply
post #33 of 38
Quote:
Originally Posted by sumitlian View Post

Unique things about those chips are that they are latest DX11, OpenGL 4.4 with good power consumption, though integer/floating point performance is still unclear.

Not really unique because they are comparing themselves to the current snapdragon SoCs. Heck the 600 and 800 were the first to get opengl es 3.0, so when Nvidia says they'll have things that qualcomm doesnt, they are saying their 2014 chip will have newer features then a 2013 chip. So its more or less Nvidia spin. Also even though the chip is slated for Q1, thats not when any devices will be launching, and by then I guarantee you the next snapdragon chips will be out.

The funniest part about all of this is, how many Tegra slides have you seen over the years? And how 'amazing' did those Tegra chips do in real use? (hint: not very well)
post #34 of 38
two leads on that device? One is power the other is a mini Hdmi input from a titan tongue.gif
post #35 of 38
Quote:
Originally Posted by Alatar View Post

And then after the power consumption part anand then says this:
First round of silicon. Almost a year from release. Absolutely no extra steppings or respins and already matching A6X at 900mW. Guess what happens when they refine the chips with tsmc for 9 months or so?

TSMC 28nm process is well understood by now and very mature. steppings can bring slight improvements but not drastic ones. when a 200 Mhz Logan on first silicon draws 900mw don't expect that to be halved. maybe shave off 50 - 100 mw. so 500 Mhz at sub 3W is maybe possible. you have to understand that clock scaling requires more voltage and power scales exponentially with increased voltage. frankly Nvidia has been all tak with regards to Tegra. They have soured their relationship with OEMs because of overpromising and underdelivering. the lack of high profile / high volume design wins for Tegra 4 is proof.

By the time Nvidia ships Logan in products sometime in mid - 2014 Apple would have released A7X in Oct 2013 and A8X on TSMC 20nm would be close to launching in Oct 2014 . Qualcomm would also be ready with 20nm Snapdragon products in Q3 2014.
Edited by raghu78 - 7/25/13 at 2:07am
Fragbox
(14 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i5 2400 DH67BL AMD Radeon HD 6900 Series XFX HD 6950 2GB 
RAMRAMHard DriveOptical Drive
Corsair Corsair DDR3 1333 Mhz 2 GB Corsair DDR3 1333 Mhz 2 GB Western Digital Caviar Green SONY DVD-RW AD-7260S 
OSMonitorKeyboardPower
Windows 7 Professional 32 bit BENQ G2420HD Logitech K200 Seasonic VX550 psu 
CaseMouse
ANTEC 200 V2 Logitech mouse 
  hide details  
Reply
Fragbox
(14 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i5 2400 DH67BL AMD Radeon HD 6900 Series XFX HD 6950 2GB 
RAMRAMHard DriveOptical Drive
Corsair Corsair DDR3 1333 Mhz 2 GB Corsair DDR3 1333 Mhz 2 GB Western Digital Caviar Green SONY DVD-RW AD-7260S 
OSMonitorKeyboardPower
Windows 7 Professional 32 bit BENQ G2420HD Logitech K200 Seasonic VX550 psu 
CaseMouse
ANTEC 200 V2 Logitech mouse 
  hide details  
Reply
post #36 of 38
Thread Starter 
Except that the proper exponential scaling only really starts when we reach the actual operational frequencies of the chips in question.

See what happens with an i7 2600K for example when it's running closer to the minimum possible operating frequencies.



Percentage wise the speed improvement is actually bigger than the increase in power consumption. The cores aren't the only thing in a CPU pulling power and neither are the ones in GPUs. At some point lowering the frequencies starts having a diminishing effect. And soon after that you reach the point where the chip wont even operate at low frequencies. Kepler as an architecture is designed to scale over 1ghz. 200mhz is very low.

Also it's good to note that even though this is a kepler SMX, it's not the same thing you get on desktops. It's a tweaked variant. FP64 cores removed, and some other tweaks done as well. Getting it right and getting the manufacturing right will take time and some effort. Manufacturing kepler for high power GPUs =/= manufacturing kepler for mobile.

You're the guy who says that AMD can increase their die size by 30%, keep the same clocks and increase performance by 30% and keep their old power consumption numbers. And do this with the same architecture. But apparently Nvidia can't go from one architecture to another, ancient shader architecture to a modern one and improve the design's power consumption by more than 5% because the process is already so well known?

Anyways, this thing is about three times as powerful as the 32nm A6X at the same power consumption levels? Do you think the next apple SoC will get a 200% performance boost? And do you think Nvidia wont also shrink this thing to 20nm as soon as possible?
 
Benching
(17 items)
 
 
CPUMotherboardGraphicsRAM
[i7 5960X @ 4.8GHz] [Rampage V Extreme] [Titan 1400MHz (1500MHz bench)] [Various] 
Hard DriveCoolingCoolingCooling
[250GB 840EVO +2x SpinpointF3 1TB RAID0] [LD PC-V2 SS Phase Change] [XSPC X2O 750 pump/res] [Monsta 360 full copper + EK XT 360 + XT 240] 
MonitorPowerCaseAudio
[Crossover 27Q LED-P 1440p+ASUS 1200p+LG 1080p] [Corsair AX1200] [Dimastech Easy v3.0] [Sennheiser HD558s] 
CPUCPUMotherboardGraphics
FX 8320, FX 8350, Phenom II x2 555BE i7 3930K, i7 860, i7 4770K, 68x Celeron D CVF, commando, 2x RIVE, Z87X-OC Asus 4870x2, Sapphire 4870 
GraphicsGraphicsGraphicsGraphics
2x 5870, 5850, 5830, 5770 2x 3870x2, 3870 GTX Titan, GTX 480, GTX 590 GTX 285, GTX 260, 4x 9800GT, 8800GTX 
RAMHard DriveCoolingCooling
4x4GB vengeance, 2x4GB predatorX, 2x1GB OCZ DDR2 Intel X25-M 80GB LD PC-V2 SS Phase Change OCN Marksman 
CoolingCoolingOSPower
2x old tek slims (GPU) Various watercooling stuff win7, winxp AX1200 
Case
test bench / cardboard box 
  hide details  
Reply
 
Benching
(17 items)
 
 
CPUMotherboardGraphicsRAM
[i7 5960X @ 4.8GHz] [Rampage V Extreme] [Titan 1400MHz (1500MHz bench)] [Various] 
Hard DriveCoolingCoolingCooling
[250GB 840EVO +2x SpinpointF3 1TB RAID0] [LD PC-V2 SS Phase Change] [XSPC X2O 750 pump/res] [Monsta 360 full copper + EK XT 360 + XT 240] 
MonitorPowerCaseAudio
[Crossover 27Q LED-P 1440p+ASUS 1200p+LG 1080p] [Corsair AX1200] [Dimastech Easy v3.0] [Sennheiser HD558s] 
CPUCPUMotherboardGraphics
FX 8320, FX 8350, Phenom II x2 555BE i7 3930K, i7 860, i7 4770K, 68x Celeron D CVF, commando, 2x RIVE, Z87X-OC Asus 4870x2, Sapphire 4870 
GraphicsGraphicsGraphicsGraphics
2x 5870, 5850, 5830, 5770 2x 3870x2, 3870 GTX Titan, GTX 480, GTX 590 GTX 285, GTX 260, 4x 9800GT, 8800GTX 
RAMHard DriveCoolingCooling
4x4GB vengeance, 2x4GB predatorX, 2x1GB OCZ DDR2 Intel X25-M 80GB LD PC-V2 SS Phase Change OCN Marksman 
CoolingCoolingOSPower
2x old tek slims (GPU) Various watercooling stuff win7, winxp AX1200 
Case
test bench / cardboard box 
  hide details  
Reply
post #37 of 38
Kepler is indeed by nature a matured arch, IT'S FROM FERMI!
And it's true that AMD can increase die size by 30% blablabla, HD7790 has already shown us that. But Nvidia? They don't have Kepler 2.0 AFAIK.
And ... comparing a 28nm to 32nm .. Not sure if serious. Why do we seem so pessimistic about Tegra? Well it's because they have overpromised and undelivered all the time, just ask anyone who bought a HTC One X!
AP33 was more power hungry than a 1.7GHz Qualcomm S3 Scorpion and was slower than that as well (thanks to the memory bandwidth)
They totally needed T33 to fix AP33 problems and by then the One X+ was a good 6 months later than the Samsung S3 which still ran much longer
Is it really a surprise that HTC's decision to go with Nvidia for the first quad-core phone led to their demise?
Not really. Nvidia came out with a quad-core A8 first but that was a epic flop

Same goes for Tegra and Tegra 2 as well as Tegra 4
I'm sure i don't have to spell it out

TEGRA SUCKS
Ayase Eli
(21 items)
 
  
CPUMotherboardGraphicsRAM
Intel E5-2675 V3 ASUS X99-A/USB3.1 Sapphire R9 290X ref Corsair Vengeance LPX 2400MHz C14 
Hard DriveHard DriveHard DriveCooling
Zotac SSD Premium 480GB OCZ Vector 180 480GB Hitachi 7K3000 3TB 7200rpm Barrow 250mm Glass 
CoolingCoolingCoolingCooling
EK Supremacy MX EK-D5 X-TOP Rev.2 + Swiftech D5 Aquacomputer R9 290X "Bitspower EML" fittings 
CoolingCoolingMonitorKeyboard
Magicool 360 Ultra Magicool 240 Slim Dell U2713HM CM Storm MECH 
PowerCaseMouseMouse Pad
FSP Aurum PT 1200 Phanteks Enthoo Pro Logitech G502 Tesoro Aegis X3 Virtus.Pro 
Audio
Audio Technica M50X 
  hide details  
Reply
Ayase Eli
(21 items)
 
  
CPUMotherboardGraphicsRAM
Intel E5-2675 V3 ASUS X99-A/USB3.1 Sapphire R9 290X ref Corsair Vengeance LPX 2400MHz C14 
Hard DriveHard DriveHard DriveCooling
Zotac SSD Premium 480GB OCZ Vector 180 480GB Hitachi 7K3000 3TB 7200rpm Barrow 250mm Glass 
CoolingCoolingCoolingCooling
EK Supremacy MX EK-D5 X-TOP Rev.2 + Swiftech D5 Aquacomputer R9 290X "Bitspower EML" fittings 
CoolingCoolingMonitorKeyboard
Magicool 360 Ultra Magicool 240 Slim Dell U2713HM CM Storm MECH 
PowerCaseMouseMouse Pad
FSP Aurum PT 1200 Phanteks Enthoo Pro Logitech G502 Tesoro Aegis X3 Virtus.Pro 
Audio
Audio Technica M50X 
  hide details  
Reply
post #38 of 38
Quote:
Originally Posted by Alatar View Post

Except that the proper exponential scaling only really starts when we reach the actual operational frequencies of the chips in question. Percentage wise the speed improvement is actually bigger than the increase in power consumption. The cores aren't the only thing in a CPU pulling power and neither are the ones in GPUs. At some point lowering the frequencies starts having a diminishing effect. And soon after that you reach the point where the chip wont even operate at low frequencies. Kepler as an architecture is designed to scale over 1ghz. 200mhz is very low.

why do you show a CPU chart when talking of a GPU. secondly the power envelopes are very different. thirdly the intel process is different and the best in the world. their process is tweaked according to their products requirements and their chips are closely designed with their process to extract best results. Intel gets the best out of their process and their IDM model is incomparable to a fabless company like Nvidia working with a foundry like TSMC who has a host of other customers.
Quote:
Also it's good to note that even though this is a kepler SMX, it's not the same thing you get on desktops. It's a tweaked variant. FP64 cores removed, and some other tweaks done as well. Getting it right and getting the manufacturing right will take time and some effort. Manufacturing kepler for high power GPUs =/= manufacturing kepler for mobile.

agreed. but still 900mw at 200 mhz cannot be halved by a couple of steppings. a 10 - 15% improvement is at best.
Quote:
You're the guy who says that AMD can increase their die size by 30%, keep the same clocks and increase performance by 30% and keep their old power consumption numbers. And do this with the same architecture. But apparently Nvidia can't go from one architecture to another, ancient shader architecture to a modern one and improve the design's power consumption by more than 5% because the process is already so well known?

HD 7790 has shown what AMD can do in the same power envelope with better chip design, power optimization (more voltage states) and binning. the core clocks remained the same on HD 7790 as on HD 7770. so just chill out man. rolleyes.gif as for improvements on Logan with future steppings I clearly stated that shaving off 100mw from 900 mw (at 200 mhz clocks) is definitely possible. slightly above 10% reduction in the next couple of steppings. what I said is don't expect a halving of power consumption.

Quote:
Anyways, this thing is about three times as powerful as the 32nm A6X at the same power consumption levels? Do you think the next apple SoC will get a 200% performance boost? And do you think Nvidia wont also shrink this thing to 20nm as soon as possible?

firstly A7X on TSMC or Samsung 28nm will definitely improve GPU performance well above A6X / Tegra 4. Apple is the most aggressive mobile SOC vendor when it comes to GPU performance. There is a very good probability that A7X will sport Power VR Series 6 Rogue which is expected to be in the 200 - 300 GFLOPS range and launch in Q4 2013.

A6X was shipping since last Oct in very high volume and all Nvidia had was Tegra 3. Now that Tegra 4 has just started shipping it will compete with A7X which will destroy it. And when Logan ships it will compete with 20nm A8X.

you have to understand Nvidia lags behind Apple / Qualcomm in transitioning to a new process and has not shown to match Apple in mobile SOC GPU perf thumb.gif

finally OEMs know better than you do of Nvidia overpromising and underdelivering on perf within a given power envelope in the mobile space. Why do you think Qualcomm won the next gen Google Nexus 7 (announced yesterday) and next gen Surface RT due later this year biggrin.gif

http://www.google.com/nexus/7/
http://www.anandtech.com/show/7171/googles-breakfast-with-sundar-pichai-event-live-blog

FROM THE BLOG

"12:19PM EDT - 1.5 GHz Snapdragon S4 Pro SoC (APQ8064 no doubt)"
Edited by raghu78 - 7/25/13 at 9:49pm
Fragbox
(14 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i5 2400 DH67BL AMD Radeon HD 6900 Series XFX HD 6950 2GB 
RAMRAMHard DriveOptical Drive
Corsair Corsair DDR3 1333 Mhz 2 GB Corsair DDR3 1333 Mhz 2 GB Western Digital Caviar Green SONY DVD-RW AD-7260S 
OSMonitorKeyboardPower
Windows 7 Professional 32 bit BENQ G2420HD Logitech K200 Seasonic VX550 psu 
CaseMouse
ANTEC 200 V2 Logitech mouse 
  hide details  
Reply
Fragbox
(14 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i5 2400 DH67BL AMD Radeon HD 6900 Series XFX HD 6950 2GB 
RAMRAMHard DriveOptical Drive
Corsair Corsair DDR3 1333 Mhz 2 GB Corsair DDR3 1333 Mhz 2 GB Western Digital Caviar Green SONY DVD-RW AD-7260S 
OSMonitorKeyboardPower
Windows 7 Professional 32 bit BENQ G2420HD Logitech K200 Seasonic VX550 psu 
CaseMouse
ANTEC 200 V2 Logitech mouse 
  hide details  
Reply
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Hardware News
Overclock.net › Forums › Industry News › Hardware News › [anand] NVIDIA Logan SoC: < 1W Kepler, 1H 2014, More Energy Efficient than A6X?