Haswell vs Kaveri

[Blazkowicz]

To rephrase, it's like a taller Intel NUC with a high TDP Intel CPU (as long as 65W is considered "high" in this context) with Iris 5200 graphics.
It's awesome, but it costs a fucking ton of money as it uses that almost experimental eDRAM die with high speed special connection to the CPU. From above review, it also has not enough cooling.

 

[DSC]

http://www.anandtech.com/show/7875/new-unlocked-iris-pro-cpu-broadwell

Intel is gonna destroy Kaveri in performance.

 

[Wynix]

Time for a Broadwell vs Carrizo thread imo.

Also;

http://www.anandtech.com/show/7873/new-haswell-k-overclocking-enhancements

 

[3dilettante]

Given the rumored 28nm standard bulk process and TDP reduction for Carrizo, a Kaveri vs. Carrizo thread may be a more appropriate.

 

[Wynix]

Given the rumored 28nm standard bulk process and TDP reduction for Carrizo, a Kaveri vs. Carrizo thread may be a more appropriate.

Spoiler

 

[Andrew Lauritzen]

It's simply impossible, no matter what the situation.

It's completely possible; different test scenarios make comparisons between them meaningless (let alone different patches of BF4). In particular, the scene they used in the Brix Pro review is known to be one of the heaviest in BF4 (outdoors, lots of objects and effects) single player whereas the scene they used in the Kaveri review is pretty light (indoors, small area).

 

[Priyadarshi]

But doesn't that make the review a bit meaningless since the reviewer can cherry pick the areas to show a greater performance difference between two GPUs. The only way to settle this is making a more comprehensive review using different maps.

 

[Lightman]

It's completely possible; different test scenarios make comparisons between them meaningless (let alone different patches of BF4). In particular, the scene they used in the Brix Pro review is known to be one of the heaviest in BF4 (outdoors, lots of objects and effects) single player whereas the scene they used in the Kaveri review is pretty light (indoors, small area).

BF4 map they tested Brix Pro on is the best case scenario for Iris Pro vs Kaveri as not only they used most CPU intensive scenario, they also dropped effects to Low. AMD tend to do relatively better the higher shader load as then it becomes more compute limited and less bandwidth constrained.

Anyway it is only one measuring point, but it shows potential of Intel Graphics and what good solution to memory bandwidth can bring to socketed GPU's

 

[Andrew Lauritzen]

But doesn't that make the review a bit meaningless since the reviewer can cherry pick the areas to show a greater performance difference between two GPUs.

No? Unless every review is meaningless...

BF4 map they tested Brix Pro on is the best case scenario for Iris Pro vs Kaveri as not only they used most CPU intensive scenario, they also dropped effects to Low.

Agreed on the "low" part, but I don't think that area is particularly CPU intensive. I doubt Kaveri is CPU limited at 25-30fps - even on low... that would be ultra-sad for single player (and you'd see a much larger difference with Mantle in that case I think). More likely the difference in that scene is memory bandwidth-related as there are a lot of blend/post-FX things going on with the wind and rain.

Easy to test though, just someone with a Kaveri load it up with those settings and check the GPU load.

 

[DSC]

Intel 10.18.10.3496 drivers released.

https://communities.intel.com/thread/49681

32bit
https://downloadcenter.intel.com/Detail_Desc.aspx?DwnldID=23713

64bit
https://downloadcenter.intel.com/Detail_Desc.aspx?DwnldID=23714

 

[firstminion]

But doesn't that make the review a bit meaningless since the reviewer can cherry pick the areas to show a greater performance difference between two GPUs. The only way to settle this is making a more comprehensive review using different maps.

Pretty much, yes. We can't say much with a sample size of one. OFC that doesn't apply only to tech reviews.

 

[Paran]

This is new to me, Intel recommends multi-rank memory.

 

[homerdog]

This is in reference to IGP performance right? Doubt it would make much difference with a discreet GPU in there.

 

[Paran]

Integrated GPU of course.

 

[Andrew Lauritzen]

Wow, 7% is more different than I would have thought. Granted this is a specific graphics benchmark but still.

 

[Psycho]

Just like computerbase found for kaveri: http://www.computerbase.de/2014-01/amd-kaveri-arbeitsspeicher/2/ - but NOT for 4670K!?

 

[Andrew Lauritzen]

but NOT for 4670K!?

The 4670k is a GT2 GPU whereas the one Intel tested is GT3 (i.e. twice as big). It's definitely expected that the latter will be much more significantly bottlenecked by DRAM bandwidth than the former.

I'm curious how much it matters for Iris Pro though... i.e. how much of that "scatter" parallelism is the EDRAM just able to eat directly? Someone should do a test

 

[mczak]

The 4670k is a GT2 GPU whereas the one Intel tested is GT3 (i.e. twice as big). It's definitely expected that the latter will be much more significantly bottlenecked by DRAM bandwidth than the former.

That would be true if they'd have the same TDP but they do not. From some quick googling it seems the 28W GT3 is slightly slower than a 84W GT2 (which is most likely always running at max IGP clocks) in vantage (p score). Even the projected performance in that slide with 31/35W would only make that roughly equal, so I would expect them to be similarly affected.
The slide doesn't mention it, is the memory really running at ddr3-1600 or something lower to artificially exaggerate the differences?

 

[Paran]

28W Iris is heavily memory bottlenecked.

 

[mczak]

28W Iris is heavily memory bottlenecked.

Yes but the math doesn't work out the slightest bit if you assume GT3 is heavily bottlenecked but GT2 is not, if they both have roughly the same performance in the end (it would essentially mean the GT2 would be not bottlenecked at all by bandwidth, whereas any additional shader capacity the GT3 has is completely wasted due to bandwidth constraints). Because both have very similar performance in the end, both should be equally bottlenecked by bandwidth, therefore both should benefit similarly from dual rank memory. At least unless there's a specific reason why GT2 could not benefit from more open pages (which is possible maybe GT3 having twice the ROPs could play a role here).