Haswell vs Kaveri

Of course by the time that Cherry Trail is out AMD's Nolan will be on the horizon and of course it's competition will not be Cherry Trail but Intel next next low power core.

 

Mullins will face Cherry Trail and according to a Roadmap AMD doesn't bring anything new for tablets 2015, means it will face Willow Trail as well:





Don't get fooled by AMDs marketing. Even if they officially launch Mullins May/June like Temash last year, we have to wait a couple of months until we see a couple of devices in stores. It happened with Temash, Trinity, Richland and will not change with Mullins. AMD probably knows it:

http://www.techradar.com/reviews/pc...t-discovery-tablet-review-1212897/review#null


@Topic:




http://www.pugetsystems.com/labs/articles/AMD-A10-7850K-Performance-Review-529/



I'm surprised about the small gap between Richland and Kaveri.

 

Looks like it's very memory bandwidth limited (in fact Just Cause 2 shows a 25% perf improvement for 17% more bandwidth?). I thought though Sea Islands (and just GCN too) had some new tricks in the bandwidth saving area, but based on these results it doesn't look like it would help much... And they were using mostly low settings, thus probably making the benchmarks generally more bandwidth limited.

 

This is one problem it seems and the Gflops downgrade from 1050 to 856 Gflops didn't help either. On higher settings Kaveri might get bigger gains though.

 

According to the numbers in the same review, 1600->1866 is a decent jump but 1866->2133 isn't huge. If the GPU perf numbers are real and translate into other games similarly (and that's sort of the range that is expected from the specs) then while I'm sure Kaveri will beat Iris Pro 5200 when MSAA is used, I'm not sure it will without. Will be interesting to see in any case as the have fairly similar theoreticals on paper (ignoring EDRAM).

 

How well does Kaveri perform against the Intel Iris 5100? (The version without eDRAM)

All the comparisons I've seen in the past few days are against CPUs with the weaker HD 4600 graphics core.

 

Not seeing anything new for tablets on a roadmap clearly marked "Desktop roadmap" isn't terribly surprising.

 

One of the Kaveri Slides on this url http://imgur.com/a/IQT27
is entitled "Data Pointers".
Is the 50,000 nodes/ms for HSA code impressive?
What would the Haswell figures look like.

 

My understanding is ARM solutions will replace Beema and Mullins in 2015 and beyond...

 

They will just use the same cpu for the next 2 or 3 years like vishera . What could possibly go wrong

 

Depends also on which Iris Pro you compare it. The 65W model is additionally 7% faster than the 55W model with same turbo frequency at least in 3dmark: http://www.anandtech.com/show/7648/gigabyte-brix-pro/2

Without MSAA they should be pretty close.

 

IIRC Iris 5100 is only 10-20% faster than the regular destop variants and it's even more rare than the 5200. This is both due to no EDRAM but also it's pretty power-constrained... I think it only shows up in a 25W variant (half the power of the mobile 5200) and I only know one laptop (the highest end Asus UX301) that uses it.

Yup definitely, pretty much the entire line of mobile chips is power-limited so the additional headroom has a direct impact on performance. It will definitely be interesting to see the comparison - I don't think either is going to win by much despite what the various marketing departments might be saying

 

I've recently bought a MacBook Pro with a Core i5-4258U and this CPU also has the Iris 5100 core. It's a 28W part with CPU running at 2.4/2.9GHz and GPU running at 1100MHz. It has 8GB of DDR3L-1600

I still haven't had the time to install Windows on it but I was wondering if it would be a better laptop for the occasional gaming 'on the road' than my current PC, which has a AMD A6-3410 (Llano APU) also coupled to DDR3-1600.

 

It's not impossible that it's faster than your desktop... basically the 5100 is a little faster than the 15W 4250U/HD5000 in the Macbook Air, so you can sort of compare the AMD numbers to that and add a small delta. Some quick googling seems to indicate that indeed the 4250U is probably a little faster than the A6, so I imagine yours would be a little faster still. If you give it a try let me know - I'm curious.

 

It would seem that Kaveri, as a chip, supports four memory channels, even though FM2+ implementations can only use two:


More details and further evidence from Hans de Vries here: http://semiaccurate.com/forums/showpost.php?p=204808&postcount=2925

So, quad-channel DDR3L in laptops? Or for servers/HPC? Something SteamBoxy? Dave?

 

I will. I'm also curious because all my past experiences with Intel integrated graphics were... Well, rather forgettable

 

Yeah well, it has gotten a lot better in the past few years once Intel started taking it seriously

 

Intel HD graphics game compatibility is incredibly better the GMA days. It seems like AMD and NV have little advantage over them in that area anymore.

 

I always thought the PHYs were required to be on the outside of the chip due to the i/o pads being there, so how is that supposed to work?

 

Laptops with embedded 2GB DDR3 that only the GPU can access?

Or it could simply be broken and Kaveri will never, ever use the third and fourth channels.
It looks like a big waste of transistors and die area, though.