Anandtech Haswell review is up!

[BallaTheFeared]

Yes, but Intel didn't spend 5.4 billion dollars on their gpu tech either.

 

[f1sherman]

Yes, but Intel didn't spend 5.4 billion dollars on their gpu tech either.

+ they produce all of their silicion
+ size increase because of edram does not affect the yield
+ mm2 "inefficiency" is a conscious decision in order to skyrocket performance/TDP

 

[notty22]

Really ? Discrete gpu's are 'better' than APU, gpu's! Who would have ever known this revelation.

 

[StinkyPinky]

Intel have quite rightly changed focus to the mobile market. why wouldn't they? That's where the money is these days as desktop pcs are slowly dying. There will always be a niche for them for power users but that world has moved to mobile and Haswell is pretty impressive there.

 

[Abwx]

What's with
the random line breaks?
I always picture
you composing your posts
under a cherry Tree wearing
victorian costume while wracked with
Consumption.
The content is usually
interesting but the
format is challenging.

Well , as hinted by the member just below the width
of the writing area is optimal with about 12-15 words
per line , hence the stanza apparent but unwillingful style.

Some crap browser on a non-win system is my guess...

Firefox + WXP on a T4400.

 

[Haserath]

Yes, just to show you how inefficient(performance per die size) the Intel Iris is, at almost the same die size of 160mm2, AMDs HD7790 (including 128bit memory controller) should be 50% faster or more at 28nm process.

You're comparing a full blown GPU@ higher TDP vs a mobile CPU/GPU. How do you know what clocks the 7790 would be at in a mobile APU?

 

[Magic Carpet]

You're comparing a full blown GPU@ higher TDP vs a mobile CPU/GPU. How do you know what clocks the 7790 would be at in a mobile APU?

Of course, it's a very incorrect comparison. But it gives food for thought, nonetheless.

 

[Arkaign]

Im not the one produced a 174mm2(for the GT3 iGPU) + 87mm2 for the eDRAM die that perform lower than the 118mm2 GTX650M (including the 128bit memory controller)

Are you silly? Most of the performance difference can be accounted for the dedicated GDDR memory in discrete CPUs If you want to compare iGPUs with system DDR3, that's one thing, comparing them to discrete cards with GDDR5 is apples/oranges. Of course there is a tremendous amount more bandwidth to work with on a dedicated card with dedicated memory designed entirely to the task.

Intel's new integrated graphics are actually incredible considering the limitations inherent in iGPU.

 

[Arkaign]

Of course, it's a very incorrect comparison. But it gives food for thought, nonetheless.

It does, but to complete the thought, one would have to hypothetically build an Intel HD discrete card with the same bus and memory (as well as cooling!) to give the same clock ceiling. iGPUs have a low ceiling with shared DDR3 and thermal constraints. AMD's iGPUs are no different, look at Llano/Trinity on DDR3-1066 vs. DDR3-1866 even. HUGE HUGE differences. With 1066 they look like total garbage.

 

[AtenRa]

You're comparing a full blown GPU@ higher TDP vs a mobile CPU/GPU. How do you know what clocks the 7790 would be at in a mobile APU?

I was talking about performance per die size. But, another way to see how inefficient the Intel Iris is, have a look at the AMD Trinity A10-4600M at 35W and 1600MHz memory against the Desktop Iris 4600 in Core i7 4770K with 84W TDP and 2400MHz memory.

The iGPU in Trinity is close to 98mm2 at 32nm, Iris 4600 is close to 87mm2 at 22nm. Trinity iGPU at 22nm TriGate would be close to 55-60mm2.

And this is Trinity, if we put 35W Richland in the same graphs it will catch or even surpass the Intel Iris 4600.

 

[TidusZ]

So a few years from now will the only way to get meaningful pc upgrades be multi-processor PCs?

Sick of this mobile/smartphone bullshit

 

[ShintaiDK]

Only the GT3e is called Iris...

 

[AtenRa]

Are you silly? Most of the performance difference can be accounted for the dedicated GDDR memory in discrete CPUs If you want to compare iGPUs with system DDR3, that's one thing, comparing them to discrete cards with GDDR5 is apples/oranges. Of course there is a tremendous amount more bandwidth to work with on a dedicated card with dedicated memory designed entirely to the task.

Intel's new integrated graphics are actually incredible considering the limitations inherent in iGPU.

Sorry but you forgetting that GT3e has Crystallwell. The comparison I made is valid.

 

[Arkaign]

It's misleading to use low-performance CPUs with more focus on GPU as a basis for that conclusion. The compute portion of Llano/Trinity is a joke. You could show graphs for video encoding (wattage for frames/second), and the Trinity would look like a power-sucking slug.

It's just not a reasonable apples/oranges comparison.

 

[AtenRa]

Only the GT3e is called Iris...

Yes my mistake, Intel 4600 is HD graphics

 

[Khato]

One of the poster said it was surprising for Intel to surpass AMD on iGPU so quickly, its not surprising when they dedicate such massive die space for their iGPU, and on a node advantage.

But the fact is, if Iris Pro is ~165mm2 or so, that's within range of a decent dedicated GPU die, which it has no hope of matching.

All that adds to the $$.

I believe Anand's lack of time/sleep showed through just a bit with that estimate of his. We know the Haswell 4C GT2 die size and have a high-resolution die shot of it. From such we can easily derive what portion of its 177mm^2 die is devoted to graphics - the answer being roughly 58mm^2. GT3 is nothing more than a doubling of GT2, so roughly 116mm^2.

There's little question that Anand's estimate of 174mm^2 for GT3 is incorrect as, from the above, we know that 4 cores, their associated L3, and system agent in the 4C GT2 die take up roughly 119mm^2. Subtract that from the measured 264mm^2 for the 4C GT3 die and you arrive at a maximum of 145mm^2 for the graphics.

I expect that the graphics area is closer to the 116mm^2 figure and the remaining 29mm^2 in the 4C GT3 configuration is due to using a less optimal die configuration (more square instead of the increasingly skinny rectangle that's been with us since SNB) and the additional logic and IOs necessary for the eDRAM L4.

For comparison, Trinity is using approximately 106mm^2 on a 32nm process. So compensating for the node advantage GT3 is probably around 2x the die area. But the other way to look at it is that this is the first time Intel is using a similar percentage of die space on graphics as AMD has been for some time now.

 

[Arkaign]

Sorry but you forgetting that GT3e has Crystallwell. The comparison I made is valid.

No, 128MB eDRAM is not apples/apples.

Iris on discrete with 1GB or 2GB of GDDR5 would kick the pure crap out of Iris discrete with 128MB eDRAM + shared DDR3.

 

[AtenRa]

It's misleading to use low-performance CPUs with more focus on GPU as a basis for that conclusion. The compute portion of Llano/Trinity is a joke. You could show graphs for video encoding (wattage for frames/second), and the Trinity would look like a power-sucking slug.

It's just not a reasonable apples/oranges comparison.

Again, the comparison was only for the iGPU. Intel HD4600 87mm2 at 22nm vs AMD Trinity a10-4600M 98mm2 at 32nm.

 

[Arkaign]

Again, the comparison was only for the iGPU. Intel HD4600 87mm2 at 22nm vs AMD Trinity a10-4600M 98mm2 at 32nm.

Okay, and are they only measuring the power usage of the iGPU, or of the package?

 

[ViRGE]

I've cleaned up this thread once already. I'd rather not have to do it again.

Please stay on topic, and above all please remain respectful. Should you have a problem with another poster, report it and we will deal with it.

-ViRGE

 

[loccothan]

Just one thing fo reading right TDP --> http://en.wikipedia.org/wiki/Thermal_design_power generally is not volt from Power plug, but Heat distribution. Overclocked Haswell has 85 celcius degree ! but Overclocked my Phenom II x6 1105 has on 4.1 max 55 celsius. AMD always put Highest Maximum TDP for its processor - Intel shows TDP different

 

[ShintaiDK]

Just one thing fo reading right TDP --> http://en.wikipedia.org/wiki/Thermal_design_power generally is not volt from Power plug, but Heat distribution. Overclocked Haswell has 85 celcius degree ! but Overclocked my Phenom II x6 1105 has on 4.1 max 55 celsius. AMD always put Highest Maximum TDP for its processor - Intel shows TDP different

Your Phenom II got its sensor another place and reports with an offset.

Same goes for FX chips for that matter. They report close to the same temperature as the heatsink actually is when measured with an IR meter.

And AMD do not put the highest TDP. Its already proven by even mobo makers and IDC that FX chips hit 140W+ and run "out of spec". Even tho AMD refuse to publish any specs on the area.

 

[AtenRa]

Okay, and are they only measuring the power usage of the iGPU, or of the package?

The A10-4600M is within a 15" Laptop, constraint to 35W TDP cooler and way higher thermals. Install the same APU to an open environment(Desktop) and you will get more performance at even lower power usage.
It's not that difficult to see that Intels iGPU is not that efficient, if they didnt have the process lead they would be in serious problems.

 

[ShintaiDK]

The A10-4600M is within a 15" Laptop, constraint to 35W TDP cooler and way higher thermals. Install the same APU to an open environment(Desktop) and you will get more performance at even lower power usage.
It's not that difficult to see that Intels iGPU is not that efficient, if they didnt have the process lead they would be in serious problems.

I think you forget the CPU part is over twice as fast as well. And I am sure even the 15W Haswell CPUs will easily beat the A6 4400M.

 

[galego]

Real world tests suggest a greater power delta, however. 7W TDP difference is rather misleading. Haswell isn't power efficient past certain Ghz level, at all.

Idle power is nice, but it isn't as important to me.

All us know how Intel measures TDPs

Average 12-15% higher power consumption for 8% performance gain is a step backward of Failwell over ivy. Look at the toms review, the figure on total power consumed shows that Haswell is more inefficient