It's rather easy to guess how it will score. 1GHz Ontario scores 0.19 in ST and 0.39 in MT. 0.25 for ST on Temash means it has a 32% advantage. 1.4GHz quad Bobcat would get ~1.06.
I wouldn't be too sure. 1.6GHz 8 core Jaguar would get 3.2 points in Cinebench, while Core i7 3517U gets 2.7 points.
But there are probably two other factors playing into this:
-AMD is probably more willing to do custom core, while Intel isn't
-Margins and pricing may be too low for Intel and it may not be for AMD, after all, the latter is quite desperate
Last edited by a moderator:
itsmydamnation
Veteran
http://www.engadget.com/2013/02/24/amd-turbo-dock-hands-on/
looks good form a demo perspective. Now just need a way better display and better styling :smile:
looks good form a demo perspective. Now just need a way better display and better styling :smile:
To me it looks too expensive with the duplication of processors and memory.
PS. oops nevermind, I didn't see they added cooling through the connector ... thought they were switching processors.
Ivy Bridge increased to GPU die space percentage to almost 40%. This part of the chip is idling during CPU heavy workloads (such as Cinebench). This change (among the 22nm "3d" transistors) allowed Intel to increase the turbo clocks of their ULV chips drastically. Ivy Bridge ULV has a 3.0 GHz turbo clock, and it can even (moderately) turbo when all the CPU cores (and all the four threads) are crunching heavy AVX loads. Sandy Bridge couldn't turbo in this situation, and this brings Ivy bridge a nice advantage in CPU heavy applications that are not using the GPU. Jaguar cannot compete with Ivy Bridge in this scenario (while it can compete with Sandy). However games are stressing both the GPU and the CPU, and heavy GPU usage drops Ivy bridge CPU clocks to nominal values.
I wouldn't worry about the gaming performance of Jaguar (Temash). According to engadget it beats HD 4000 based ULV processors in DiRT Showdown (1080p): http://www.engadget.com/2013/01/10/amd-temash-reference-laptop-hands-on
For what it's worth, even if SB could turbo in some circumstances the model AMD compared against has turbo completely disabled. Making it kind of a castrated part despite having an i3 branding.
And that's basically how it would perform if it were a console CPU. The integrated GPU would always be utilized by 90%+ (leaving no extra TDP for the CPU), and all four CPU threads would be always used as well (there would always be two threads running on every CPU core, sharing the resources).
The raw flop/s rate of ULV Ivy Bridge (dual core, four threads) and Jaguar (four cores) is identical (when both CPUs are running at nominal clocks). Single IB core can do 16 flops per clock (8 wide vector, add + mul), single Jaguar core can do 8 flops per clocks (4 wide vector, add + mul). Jaguar has twice the number of cores. If the goal is to maximize the CPU flops and the TDP budget for the CPU is under 20 watts, the Jaguar is a very good contender.
However for the PC laptop market the Ivy Bridge ULV is a very good chip indeed. Many applications are still mainly single threaded, and basically only games are taxing the GPU. When running anything else than games (and CAD software), the CPU part will be running at 2.5 GHz - 3.0 GHz, providing a desktop like experience. Jaguar can't obviously match that. But then again Ivy Bridge 3667U costs 346$ a piece (http://ark.intel.com/products/64898/). Ivy Bridge based tablets cost around 1000$ mostly because of the CPU. That's a bit too much to ask, when you are competing against the iPad... even if the IB provides you 5-10x better performance compared to the best ARM chips
The SB couldn't turbo because it's an i3 surely? The 3667U is an i5.
edit - something else seems a bit off. The top end Temash is supposed to be a 5.9W chip (let's assume this chip at 1GHz is that chip). With 40% higher clocks it surely wouldn't be expected to reach 15W would it?
Some new vids, dual core is ~4W apparently and running torchlight 2.
http://www.youtube.com/watch?v=Vu35ePgBTzU
http://www.youtube.com/watch?v=8N7LNHRl2jE
http://www.youtube.com/watch?v=K0i7yB6EjBk
edit - something else seems a bit off. The top end Temash is supposed to be a 5.9W chip (let's assume this chip at 1GHz is that chip). With 40% higher clocks it surely wouldn't be expected to reach 15W would it?
Some new vids, dual core is ~4W apparently and running torchlight 2.
http://www.youtube.com/watch?v=Vu35ePgBTzU
http://www.youtube.com/watch?v=8N7LNHRl2jE
http://www.youtube.com/watch?v=K0i7yB6EjBk
Last edited by a moderator:
Very roughly speaking, dynamic power consumption is proportional to voltage squared times frequency. If voltage scales linearly with frequency that means power consumption is proportional to the cube of frequency, and 1.4^3 = 2.744, 5.9 * 2.744 = 16.2W. Dynamic power consumption probably dominates at the higher clock.
I'm guessing voltage scaling isn't nearly so fine grained so you probably see more discontinuous jumps, and there are other factors like power consumption increasing with temperature which increases with power consumption depending on your cooling setup. But you can see that a 40% increase in clock resulting in a 170% increase in power consumption isn't that crazy.
http://www.engadget.com/2013/02/24/amd-turbo-dock-hands-on/
Though I would say the logical progression is to disable the APU graphics, tune the full (boosted) tablet TDP to the CPU for higher speeds and maybe activate more cores.
Though I would say the logical progression is to disable the APU graphics, tune the full (boosted) tablet TDP to the CPU for higher speeds and maybe activate more cores.
It'd be really nice if they made system boards for APUs like GPUs boards with integrated ram. Get rid of DIMM slots and make system memory unified GDDR5. Most tablet consumers aren't going to want to upgrade their ram, and they'd certainly notice the performance increase from the added bandwidth. *nudge nudge*
If you're looking at Tamesh it doesn't really seem that limited by memory bandwidth, if you're assuming lpddr3-1600. If you compare that with a HD 7750, it's got only ~1/6 the bandwidth (and this is shared with the cpu too), however it only has less than 1/10 the compute power (and compared with hd 7770, it is still 1/6 the bandwidth but only 1/16 the compute power). Of course if you clock those CUs up to 500Mhz then the flop/bandwidth ratio will change but it's still about the same as the flop/bandwidth ratio of 7750.
Arguably though I guess you'd typically run less complex shaders so a higher bandwidth/flop ratio would probably make sense.
Though if you go up to higher power levels (say 800Mhz gpu clock or more CUs) then you'd definitely really wanted more bandwidth. Maybe that's what that 16bit mobile gddr5 listed in that hynix datasheet is supposed to be for.
That's fair, but it'll likely be cheaper than fabricating commodity dimm slots because of less board complexity. The only disadvantage is that memory densities will a bit worse, but that'll fix itself as lithography advances.
Last edited by a moderator:
AMD Temash™, now with up to 60 times more fishes than the competition!
itsmydamnation
Veteran
i went and ran that demo, you have to be careful it is very browser dependant. firefox was rubbish, IE wasn't bad, chrome was the best.
itsmydamnation
Veteran
looking at it more, it appears to be a single threaded app and from what i can see not very good at hitting the GPU because of it.
I get 1300 fish on my I7 920 @ 3.8
single thread at 100% of one core
6970 GPU hits around 40% load according to afterburner.
applying overclock to GPU sees no increase in fish count
So its very likely intel poor GPU drivers, that said AMD could easily be CPU limited.
I get 1300 fish on my I7 920 @ 3.8
single thread at 100% of one core
6970 GPU hits around 40% load according to afterburner.
applying overclock to GPU sees no increase in fish count
So its very likely intel poor GPU drivers, that said AMD could easily be CPU limited.