Mobile Kepler review

[Olikan]

http://www.pcper.com/reviews/Mobile/Nvidia-GeForce-GT-640M-Review-Kepler-Arrives-Mobile

so....charlie wasn't THAT wrong =P

 

[exar333]

Wow, great read. Looks very promising. Kepler + IB should be a great mobile solution for high=end graphics needs.

 

[Mopetar]

Might be able to get some idea of the possible performance of their new architecture based on how this part compares to their older mobile parts.

The 640M has 384 CUDA cores and the clock rate is 625 MHz for the GPU, 900 MHz for the memory, and 1250 MHz for the shaders.

Here are the scores that it got for various benchmarks:

3D Mark 06: 9591
3D Mark 11: 1785
DoW 2:      44.53
JC 2:       53.25 
BF 3:       45.21

They compare it against a 555M that has 144 CUDA cores and clock rates of 590 MHz, 1800 MHz, and 1180 MHz for GPU, memory, and shader respectively* **.

Here are the scores that it got on the benchmarks:

3D Mark 06: 12176
3D Mark 11: 1449
DoW 2:      60.62
JC 2:       56.8 
BF 3:       39.18

The comparison isn't exactly straight as the notebook with the 555M is using a more powerful processor (2.4 GHz vs. 1.6 GHz) and the reviewer even notes that the DoW benchmark is CPU limited. This is less than ideal for attempting to draw performance conclusions based on these results alone.

* I looked this information up on Wikipedia, but the 555M that this list as having a 753 MHz GPU clock only has 94 CUDA cores, and the 144 CUDA core variations have lower GPU clocks so I'm not sure if this is a part that isn't listed on Wikipedia or the reviewer is somewhat mistaken.

** Based on information from busydude, it seems more likely that the particular configuration for the GPU used in the review is wrong. The original numbers used in the post were 753 MHz GPU clock, 1506 MHz shader clock, and 3138 MHz memory clock.

 

[blastingcap]

http://www.pcper.com/reviews/Mobile/Nvidia-GeForce-GT-640M-Review-Kepler-Arrives-Mobile

so....charlie wasn't THAT wrong =P

For what it's worth, Charlie is now saying that the launch got pushed back a week and a half. I guess that means something like March 23.

http://semiaccurate.com/2012/03/08/...o-nvidia-keplergk104gtx680-launch-activities/

 

[busydude]

^^ There are two variations that are shipped with the same moniker.

From AT:

NVIDIA GeForce GT 555M "A"
96 CUDA Cores, 16 TMUs, 4 ROPs, Core Clock: 753MHz, Shader Clocks: 1506MHz
128-bit Memory Bus, GDDR5, Effective Memory Clocks: 3138MHz
Desktop Counterpart: GeForce GT 440 GDDR5 (GF108)
And this is where NVIDIA's mobile lineup completely loses its mind. The GeForce GT 555M is actually two completely different chips and configurations; the "A" and "B" are our designation. Our "A" configuration is essentially just a souped-up version of the GT 525M/540M/550M, with a higher core clock and the benefit of GDDR5. While NVIDIA lists both versions on their site (though lacking an explanation as to why this split was made), a glance at NewEgg suggests this "A" version is the more common of the two (powering MSI and Lenovo laptops while the "B" version resides almost exclusively in an Alienware.) You can recognize the "A" version by the use of GDDR5, but since it and the "B" version are so bizarrely matched we can't really tell definitively which one would be the faster of the two. (No review available.)

NVIDIA GeForce GT 555M "B"
144 CUDA Cores, 24 TMUs, 24 ROPs, Core Clocks: 590MHz, Shader Clocks: 1180MHz
192-bit Memory Bus, DDR3, Effective Memory Clocks: 1.8GHz
Desktop Counterpart: None (GF106)
The other configuration of the GT 555M is a substantially beefier chip with six times the ROPs, but it operates at lower clocks and lower memory bandwidth due to the use of DDR3 instead of GDDR5. It's essentially a die-harvested version of GF106, and is identifiable by both the use of DDR3 and memory configurations of either 1.5GB of 3GB. It remains inexplicable why NVIDIA decided to use two completely different chips for the GT 555M, but hopefully this makes it a little easier to tell which is which. Raw calculations of pixel and texture fillrate suggest this "B" configuration to be the faster of the two, and as such it's probably the one to look for. Thus far we've only seen it in the Alienware M14x. (No review available.)

555M 'B' spec is the one that Alienware uses in its M14x.

 

[Silverforce11]

So taking the BF3 results, the new shaders are ~1/3 as effective as the old shaders (taking into account clock speed differences).

Extrapolating this to gk104 vs gtx580, it has 3x more shaders which run much faster. It could well be gtx580 +20%, which puts it right into 7970 territory at stock.

That's excellent for a chip of its size.

 

[tviceman]

So taking the BF3 results, the new shaders are ~1/3 as effective as the old shaders (taking into account clock speed differences).

Extrapolating this to gk104 vs gtx580, it has 3x more shaders which run much faster. It could well be gtx580 +20%, which puts it right into 7970 territory at stock.

That's excellent for a chip of its size.

I am 90% sure that this particular laptop has ddr3 ram strapped with that 640m, no doubt hindering it.

 

[tviceman]

Wow that is weird. Anyways yeah, this particular 640m has pathetic gddr3 ram attached to it. The bandwidth is awful, even for a small chip like gk107.

 

[ocre]

i can very accurately calculate gk104 gtx 680 by this information on mobile part.

1536 virtual cores act more like 768 real cores divided by 512 gk110=150%

the 1004mhz vs 700mhz balance each other out due to the bandwidth constraints.

150% gf110
+6% give
-6% take
--------------

=gfk104 is anywhere from 111% to +132.65% of a 7970 give or take 28%!!!!!


i would put money on it

 

[Mopetar]

I am 90% sure that this particular laptop has ddr3 ram strapped with that 640m, no doubt hindering it.

I feel slightly confused, but how did you reply to a post that comes after yours considering that it does not indicate that your post was edited?

Either that or the board is just messed up for me. :hmm:

 

[tviceman]

I feel slightly confused, but how did you reply to a post that comes after yours considering that it does not indicate that your post was edited?

Either that or the board is just messed up for me. :hmm:

yeah it messed up.

EDIT: OH GOD IT'S DOING IT AGAIN. MAKE IT STOP!

 

[Ventanni]

Wow that is weird. Anyways yeah, this particular 640m has pathetic gddr3 ram attached to it. The bandwidth is pathetic, even for a small chip like gk107.

That is intentional of Nvidia, as they did that with their Geforce 540m line as well. Extrapolating that, that should give us a rough idea of what kind of models the 625m and 655m are as well:

625m - 550mhz core, 1100mhz shader, 384 CUDA cores, DDR3 @ 900mhz
640m - 625mhz core, 1250mhz shader, 384 CUDA cores, DDR3 @ 900mhz
655m - 700mhz core, 1400mhz shader, 384 CUDA cores, DDR5 @ ~800-900mhz

That should also put the 620m at at 700-750mhz core with 192 CUDA cores, albeit on a 64bit bus.

The Geforce 540m isn't the most potent of GPUs, but I will be the first to admit that it is surprisingly capable of delivering a very adequate and comfortable gaming experience for most games I've thrown at it (WoW, SC2, Deus Ex: Human Revolution, BF3 etc).

Looks like the 640m will perform a little bit better.

 

[X98]

impressive, most impressive

 

[blastingcap]

i can very accurately calculate gk104 gtx 680 by this information on mobile part.

1536 virtual cores act more like 768 real cores divided by 512 gk110=150%

the 1004mhz vs 700mhz balance each other out due to the bandwidth constraints.

150% gf110
+6% give
-6% take
--------------

=gfk104 is anywhere from 111% to +132.65% of a 7970 give or take 28%!!!!!


i would put money on it

That would be awesome but it probably doesn't scale linearly like that, due to various bottlenecks. Maybe even memory bottlenecks.

 

[taltamir]

^^ There are two variations that are shipped with the same moniker.

From AT:



555M 'B' spec is the one that Alienware uses in its M14x.

that is terrible, and I bet it is intentional and that there the two parts are not distinguishable by name alone. (aka, no "A" or "B" on the box anywhere)

 

[boxleitnerb]

What clocks is the GT555m running at? This test
http://www.notebookcheck.com/Test-Alienware-M14x-Notebook.53568.0.html
says 590, the pcper test says 753?

 

[Riek]

Cooling

One of the tactics used by Nvidia to squeeze maximum performance out of the new Kepler mobile chips is performance scaling. Every laptop has a maximum amount of heat it can cool - so why not make sure all of it is being used?

The answer to that question starts to become evident when playing games on this laptop. While the palmrest remains in the mid-80s, the middle of the keyboard becomes extremely warm, reaching temperatures up to 105 degrees.

That’s nothing compared to center-bottom of the laptop, where we read temperatures of up to 128 degrees after playing Battlefield 3 on Ultra. That’s unacceptably warm - it’s hard to imagine using this laptop on anything besides a desk if you want to engage the GPU.

I asked an Nvidia rep what they were going to do to ensure that the dynamic performance scaling did not result in high external temperatures. The answer I received was that it’s up to the laptop manufacturer as the thermal capacity of the chassis is often determined before they select a GPU. It appears Acer has been very liberal with the maximum temperatures it will tolerate.

I'm glad most people are impressed with the performance... but if i read the above statement... I'm more scared than impressed..
(also for less ventilated or lower max temps the performance will be different...(lower))

 

[Mopetar]

What clocks is the GT555m running at? This test
http://www.notebookcheck.com/Test-Alienware-M14x-Notebook.53568.0.html
says 590, the pcper test says 753?

Based on some homework done by a poster earlier in the thread, it's 590 MHz. The author of the article must have made a mistake.

 

[boxleitnerb]

Ah I overlooked the 753MHz from the 555A. Thanks.

 

[Mopetar]

I'm glad most people are impressed with the performance... but if i read the above statement... I'm more scared than impressed..

Look at the device that it's in, though. Although we don't have a full set of images in the article so that we can see if from all angles, if it's like most other ultrabooks, it won't have much for vents or a cooling system and will be cramming a lot into a small area.

It's really no different then complaining about a 7970 running hot if you put an inadequate fan on it and don't provide enough space for case vents and fans. The chips are probably fine, but anything can run hot if the cooling system is poorly designed.

 

[Riek]

Look at the device that it's in, though. Although we don't have a full set of images in the article so that we can see if from all angles, if it's like most other ultrabooks, it won't have much for vents or a cooling system and will be cramming a lot into a small area.

It's really no different then complaining about a 7970 running hot if you put an inadequate fan on it and don't provide enough space for case vents and fans. The chips are probably fine, but anything can run hot if the cooling system is poorly designed.

Not sure what your point is....
You put such a 7970 (or a 580) or any other hot component on your lap? I don't think so.. because it is not designed to be on your lap. A laptop however is, and such temperatures and heat development is just unacceptable for a laptop. And if its the kepler design (as hinted in that review) that just uses the extra thermal capabilities for his settings (despite its surrounding components) than this is just plain wrong. (not to mention the increase power consumption, changed urability in such temperature environments).
Ofcourse some manufacturers will supply better cooling and hopefully change those (according to the reviewer) thermal settings for the Kepler.... but that will also change the perfromance.

 

[Gikaseixas]

The power consumption should be much lower compared to the 555M too and if it was tested with a similar cpu we could have seen 20/30% improvement across the benchmarks. Kepler is starting to look really good to me. Now AMD needs to show it's hand in the mobile sector.

 

[Broheim]

I feel slightly confused, but how did you reply to a post that comes after yours considering that it does not indicate that your post was edited?

Either that or the board is just messed up for me. :hmm:

aint nuthin but a timewarp



ps. best forum feature ever.

 

[sontin]

Anandtech has his review up: http://www.anandtech.com/show/5672/acer-aspire-timelineu-m3-life-on-the-kepler-verge

Kepler is looking great. Especially in shader bound scenarios will it be much faster than Fermi.

I think the average performance increase over the GTX580 should be something around 60%.

 

[Silverforce11]

I think the average performance increase over the GTX580 should be something around 60%.

LOL good one.