[XBIT]AMD Richland vs. Intel Haswell: Integrated Graphics Performance Review

[Jaydip]

http://www.xbitlabs.com/articles/cpu/display/amd-richland-intel-haswell-gpu.html

Mods please delete if repost.

 

[frozentundra123456]

Pretty much as expected on the desktop. The test makes the apus look better than they really are compared to a discrete card though because if you look closely they are comparing to a DDR 3 HD7750, which I can't imagine anyone who wants to game would buy.

On the desktop, apus still are a poor solution for gaming compared to a low end CPU like the athlon x4 750k and a gddr 5 HD7750.

It would be more interesting to see a test like this in laptops compared to Intel hd4600, hd5000, and iris pro.

 

[SPBHM]

Pretty much as expected on the desktop. The test makes the apus look better than they really are compared to a discrete card though because if you look closely they are comparing to a DDR 3 HD7750, which I can't imagine anyone who wants to game would buy.

On the desktop, apus still are a poor solution for gaming compared to a low end CPU like the athlon x4 750k and a gddr 5 HD7750.

It would be more interesting to see a test like this in laptops compared to Intel hd4600, hd5000, and iris pro.

laptops should be closer, just like it was with Ivy vs Trinity...

but I agree, that DDR3 1600 (even the system ram they choose is 2133MHz lol) 7750 is really bad.... DDR5 model (reference model) is a lot faster...

 

[ShadowVVL]

n the desktop, apus still are a poor solution for gaming compared to a low end CPU like the athlon x4 750k and a gddr 5 HD7750.

Also you can get the athlon x4 750k and a gddr 5 HD7750 for only a tiny $25 more then the 6800k on newegg

 

[Insert_Nickname]

With regards to that article, I'd point out that the power consumption charts do not necessarily reflect reality. I have a 6800K+ITX board here that does not go over 40W doing 1080p decoding. Minimum power for my system is only 22W. Absolute maximum power consumption for my system is only 120W.

Also you can get the athlon x4 750k and a gddr 5 HD7750 for only a tiny $25 more then the 6800k on newegg

And even then, you'll still have to factor in the 2133MHz memory to make Richland perform optimally.

 

[littlebluesmurf]

It is a bit weird that they used a DDR3 7750, however if Kaveri has performance levels similar to a DDR3 7750 this review could give us a rough idea of what to expect from the top desktop Kavari APU's.

 

[Erenhardt]

On the desktop, apus still are a poor solution for gaming compared to a low end CPU like the athlon x4 750k and a gddr 5 HD7750.

Technically you are still buying A10 APU, but with graphics part disabled. The only reason this solution gives more performance/$ is because AMD want people to spend those extra $30 on AMD hardware. Whole plan fails when someone buys athlon 750k + gtx650

And even then, you'll still have to factor in the 2133MHz memory to make Richland perform optimally.

In my country 2x4GB 2300MHz is $5 more than 2x4GB 1600MHz or sometimes even cheaper! WTF?

 

[ShintaiDK]

In my country 2x4GB 2300MHz is $5 more than 2x4GB 1600MHz or sometimes even cheaper! WTF?

Both 1.5V or 1.35V?

On geishalz the cheapest 2133 cost twice as much as the 1600.

 

[Erenhardt]

Both 1.5V or 1.35V?

On geishalz the cheapest 2133 cost twice as much as the 1600.

Kingston HyperX Predator DIMM XMP Kit 8GB PC3-19200U CL11-13-13 (DDR3-2400) 1.65V

 

[Insert_Nickname]

In my country 2x4GB 2300MHz is $5 more than 2x4GB 1600MHz or sometimes even cheaper! WTF?

Both 1.5V or 1.35V?

On geishalz the cheapest 2133 cost twice as much as the 1600.

I -do- believe he's partially right. A quick check on proshop.dk, revealed the cheapest 8GB 2400MHz kit cost only 143DKK (~$25) more then the cheapest 1600MHz kit. 2133MHz is more expensive then 2400MHz...

 

[ShintaiDK]

Well 1.65V kinda ruins it. Its out of spec overvolted.

I dont know what AMDs stance is on 1.65V. But for Intel, your warranty is gone.

 

[Insert_Nickname]

Well 1.65V kinda ruins it. Its out of spec overvolted.

Wouldn't use 1.65V on anything Intel, but that kit appears to use Samsung ICs. So it might very well do 1866MHz on 1.35V.

 

[Erenhardt]

Well 1.65V kinda ruins it. Its out of spec overvolted.

I dont know what AMDs stance is on 1.65V. But for Intel, your warranty is gone.

I'm using Kindston HyperX 1600MHz 1.65V in my FX-6300 on cheapest am3+ mo-bo possible (M5A78L-LM lx)and it is working just fine (other than throttling from turbo to stock speed under Heavy load in OCCT - 3phase VRM)

 

[frozentundra123456]

Technically you are still buying A10 APU, but with graphics part disabled. The only reason this solution gives more performance/$ is because AMD want people to spend those extra $30 on AMD hardware. Whole plan fails when someone buys athlon 750k + gtx650


In my country 2x4GB 2300MHz is $5 more than 2x4GB 1600MHz or sometimes even cheaper! WTF?

True, about the GTX650, but I ususally use the HD7750 for comparison because it doesnt require a 6 pin connector, so you can use it as an add-in card with a cheap off the shelf system.

 

[monstercameron]

Also you can get the athlon x4 750k and a gddr 5 HD7750 for only a tiny $25 more then the 6800k on newegg

you can get an a10-6800k for $135 on amazon...

super dissapointed with the review

The Richland design has its downsides, though. Such APUs have high practical power consumption and heat dissipation. They need twice more power than the Intel Core i5 in a lot of applications, both 3D and multimedia ones. The 4000 series Core i3 is going to widen the gap even more. That’s why the A10, A8 and A6 series APUs may not be the optimal choice for small and quiet home computers or HTPCs, especially as the Richland’s multimedia engine looks outdated in comparison with the Haswell’s. It cannot cope with 4K content and has rather low performance in video transcoding tasks.

video transcode done on cpu vs. in hardware like intel
power difference is negligible when not under max load
overblows the 150W max stressed power consumption
even after outperformng intls parts reviewer goes negative on the amd part not being fast enough to play all the newest shooters in FHD.

 

[SiliconWars]

So 45% overall based on those. It's pretty hard to believe AMD could maintain a gap like that by basically doing nothing much. It's a real shame Kaveri hasn't been out by now because that would make the HD4600 performance totally unnacceptable.

 

[jvroig]

@SiliconWars,
Since you mentioned Kaveri, can you tell me what Kaveri has that will solve/mitigate the memory bandwidth limitation, like Intel's Crystalwell?

Because without anything comparable to Crystalwell, I'm not sure how Kaveri will increase significantly from Richland or Trinity on the GPU side, given that the A10 SKU's are pretty much bandwidth-limited.

I'm not completely up to date with Kaveri info, and it's hard sorting through fact (official announcements) vs completely unfounded rumors (I think the soldered-on-mobo GDDR5 thing is completely out of the picture?), so I'm asking because maybe you have much better info than me.

Much appreciated

 

[frozentundra123456]

So 45% overall based on those. It's pretty hard to believe AMD could maintain a gap like that by basically doing nothing much. It's a real shame Kaveri hasn't been out by now because that would make the HD4600 performance totally unnacceptable.

Who cares about igp performance on the desktop as long as it is good enough for normal tasks, which the HD4600 is. Just add a discrete gpu if you want good graphical performance. APUs for the desktop are still in limbo: more than needed for everyday use, but far exceeded by a low end current generation AMD discrete card if you have a graphically demanding application.

 

[Abwx]

Who cares about igp performance on the desktop as long as it is good enough for normal tasks, which the HD4600 is. Just add a discrete gpu if you want good graphical performance. APUs for the desktop are still in limbo: more than needed for everyday use, but far exceeded by a low end current generation AMD discrete card if you have a graphically demanding application.

That s all circular arguments to elude the point of said review..

Actualy what is obvious is that an A10 can be used for games
up to 1920/1080 while the 4600 can not, sometimes even
at lower res.

 

[SiliconWars]

@SiliconWars,
Since you mentioned Kaveri, can you tell me what Kaveri has that will solve/mitigate the memory bandwidth limitation, like Intel's Crystalwell?

Because without anything comparable to Crystalwell, I'm not sure how Kaveri will increase significantly from Richland or Trinity on the GPU side, given that the A10 SKU's are pretty much bandwidth-limited.

I'm not completely up to date with Kaveri info, and it's hard sorting through fact (official announcements) vs completely unfounded rumors (I think the soldered-on-mobo GDDR5 thing is completely out of the picture?), so I'm asking because maybe you have much better info than me.

Much appreciated

I don't believe we'll see anything different for bandwidth, so basically what you're seeing of the 7750 in that xbitlabs article will be close to the low-end what we get. There may be improvements from hUMA on top of it. I reckon +30% on Richland at a conservative guess.

Obviously I'd love to be wrong on this and see AMD really pushing it, but I think we're waiting on DDR4 tbh.

 

[Abwx]

I reckon +30% on Richland at a conservative guess.

33% bigger GPU with GCN and a CPU that is more or less announced
at 15-20% better perf at same frequency should yield more than 30%
better perfs provided the bandwith issue is solved , surely that hUMA
will help improve the effeciency of the available bandwidth.

 

[SiliconWars]

Yeah it's all about the bandwidth. Obviously if there is anything new at all, ie quad channel or some GDDR5 sideport or anything that basically isn't DDR3, the performance could easily more than double.

 

[monstercameron]

what if the took the 8mb l3 cache from fx8xxx and stuffed it in the apu?

 

[SiliconWars]

Doubtful it would make any improvement past the normal improvement the cache gives in some games (which to be fair is quite substantial, but also takes up a ton of die space).

 

[Insert_Nickname]

33% bigger GPU with GCN and a CPU that is more or less announced
at 15-20% better perf at same frequency should yield more than 30%
better perfs provided the bandwith issue is solved , surely that hUMA
will help improve the effeciency of the available bandwidth.

I doubt we'll see much more then 30% more performance with current DDR3. The IGP is going to bottlenecked no matter how you cut it, we need DDR4 at high frequencies to really make a dent in that bandwidth limitation. Even then the maximum semi-official announcement for 4266MHz DDR4 only gives ~34GB/s per channel. With dual channel that's only ~68GB/s, even my HD7870 has a 154GB/s available... :|

In short, I would not get my hopes up for massive gains in IGP performance...