Kaveri review : tested 13 recent games, vs Intel Iris Pro / Hybrid CF included

[DGLee]

Actually I ran all those benchmarks prior to Kaveri's embargo is lifted. Only translating them in English takes that much time thank to my poor English ability. Anyway, let's have a look:

My review consists of two parts - Part 1 is for Processor Graphics inside Kaveri while Part 2 is being planned to deal with Kaveri's coverage at the domain of traditional CPU as well as its HSA features. Today I will show you the former one.

Test hardwares are as follows:

Before the test begins, I estimated Kaveri's Processor Graphics performance by using VGA calculator I designed. It's a simple, first-order fractional equation consists of GPU's compute performance(number of shader * core clock), texture fillrate(number of TMU * core clock), rendering performance(number of ROP * core clock) and VRAM bandwidth(bitrate * VRAM clock) as its terms. (further about this equation, see this: http://udteam.tistory.com/535 - by using this, I speculated that Hawaii will be faster than GTX TITAN while full-blown GK110 will also retake the crown from Hawaii even prior to Hawaii's release) First result for A10-7850K, which has 8 Graphics Core(= 512SP/32TMU/8ROP) within it.

Second is for A10-7700K / A8-7600 both have 6 Graphics Core(= 384SP/24TMU/8ROP).

Well, the results suggests that Kaveri's Processor Graphics is virtually competitive to entry level discrete graphics card. At the same time, however, it also implies that +128SP is not actually helpful (difference between 7850K and 7700K/7600 is less than 7%) and this means performance hesitator is not SP/TMU part. Rather, ROP partition and poor memory bandwidth are more likely. So why AMD cultivate that much amount of SPs despite all? I think, AMD's intention is aimed for Post HSA era. Huge amount of SPs are employed not only for graphics acceleration but also as compute resources, thus to allow them a whole "856 GFLOPS" in one chip. (further will be dealt in Part 2)

Well, the intro was too long. Let's see the numbers!
All tests are done at 1280x720(HD) / 1600x900(HD+) with highest possible graphics quality settings.
(You can see original results from my blog: http://udteam.tistory.com/611)

-----

1. 3DMark 11
(Entry / Performance Preset)

2. 3DMark 13
(Cloud Gate / Fire Strike)

3. Aliens vs Predator
(Texture : Very High, Shadow : High, AF x16, SSAO On, Tessellation On, Advanced Shadow On, AA Off)

 

[DGLee]

4. Batman : Arkham City
(DX11 Features : MVSS and HBAO, DX11 Tessellation : High, Detail Level : Extreme, AA & PhysX Off)

5. Bioshock : Infinite
(UltraDX11 with Diffusion Depth of Field)

6. Crysis : Warhead
(64bit, DX10, Enthusiast, AA Off)

 

[DGLee]

7. DiRT : Showdown
(Ultra Preset, AA Off)

8. Hitman : Absolution
(Quality Level : Ultra, AA & FXAA Off)

9. Just Cause 2
(Texture/Shadow/SSAO : High, Water/Objects Detail : Very High, All others On except AA)

 

[DGLee]

10. Metro 2033
(DX11 with DOF, Quality : Very High, AF 16x, AA & PhysX Off)

11. Metro : Last Light
(DX11, Quality : Very High, AF 16x, Motion Blur : Normal, Tessellation : Very High, AA & PhysX Off)

12. Sleeping Dogs
(Graphics Level : Extreme, AA : Normal)

13. Tomb Raider : Reboot
(Quality : Ultra, AA Off)

That's all for this review. Here are summarizing graphs:

Thanks for reading (or attempt to read) my article. Have a nice day!

 

[BallaTheFeared]

I'd love to see if I could get a 45% OC out of Iris.

 

[Homeles]

I really like the performance summary charts.

 

[monstercameron]

what about igp + memory +NB oc for kaveri?

 

[Vesku]

Still think the Kaveri A10 is a bit overpriced but it's nice of AMD to provide us with Intel Iris Pro performance at under $200 and available for do it yourself desktop customers.

 

[jj109]

Keep in mind that the Iris Pro on the 4750HQ is TDP limited to 47W for the whole package.

 

[ZGR]

I'd love to see if I could get a 45% OC out of Iris.

I bet it scales pretty well too.

 

[jj109]

I bet it scales pretty well too.

The main question is how well the cooling solution on that platform scales. It's a shame Iris Pro isn't available in LGA 1150.

 

[DGLee]

what about igp + memory +NB oc for kaveri?

For IGP/NB/APU itself, I didn't even tried to overclock because the time Kaveri was arrived was already a couple of hours before the embargo lifts. (Of course I'd like to include Kaveri's overclockability in general as a chapter for Part 2)

For memory, all other compared groups were paired with 2 x 2400MHz DDR3 while only Kaveri has 2 x 2133MHz, since all tests except Kaveri was already done with that clock profile, and after I've got Kaveri, I found that this setting is not compatible with it. (I guess this is a kind of IMC unstability, or probably a BIOS/Mobo problem) In other words, the result is actually a bit unfair for Kaveri. If all compared groups were paired with the same speed memory, Kaveri's position in the graphs may uplift or whatever similar manner.

 

[DGLee]

I'd love to see if I could get a 45% OC out of Iris.

It's my first experience, not only of Kaveri, but also of Iris Pro. Tested with MacBook Pro 15' Retina, the results were reasonably accepted thank to its processor power is similar to my desktop counterpart (i5 4670K).

Actually I knew little about Intel's graphics and had some bias that they have really poor 3D performance, Iris Pro and HD 4600, however, helped me to get rid off that bias. I think AMD would be very thankful for Intel is not planning to ship their Iris Pro as an ordinary desktop product.

 

[DGLee]

I really like the performance summary charts.

Thank you bro

 

[Silverforce11]

Holy shiat that is damn awesome with DG, when it works, its blazing fast. I'm surprise it even works so well when the official DG drivers aren't released yet!

Hmm, a tiny SFF rig with A10 + R250.. tasty!

 

[EndlessWargasm]

I'd like to see results for dual graphics pairing a DDR3 R7 240 with Kaveri. I suspect it would be a better match.

 

[inf64]

What a sick scaling with dual graphics. It looks really good.
Great review DGLee :thumbsup:

 

[frozentundra123456]

My compliments on the review. I especially liked the way you made the summary graphs at the end. The only thing missing, and you probably don't have the equipment to test it , is some measure of microstutter with the crossfire combinations.

As for kaveri, I am not especially impressed, except for niche uses, like SFF systems for light gaming without a discrete card. Even with crossfire seeming to work better, if I were going to use a discrete card , I would just get a cheaper CPU and a 7770 or higher card.

But again, my compliments on your testing and presentation of the data. Nice job.

Edit the A8 when available would be interesting to test in crossfire. It is much more price competitive vs an athlon x4 750k, and doesn't seem to lose too much performance. At its current price, it is hard to justify purchasing a 7850k plus a discrete card, when one could get an athlon x4 plus HD7770 for approximately the same price as the APU alone.

 

[SPBHM]

it's interesting to see how close the 7850K and Iris Pro are, for 2 reasons, one is the fact that AMD can get same/higher performance with no ultra fast cache, just DDR3, but at same time Intel have an IGP basically as fast as the best IGP AMD (with all their GPU expertise) have, but obviously both are not competing at the same price, so it's a clear win for AMD.

if AMD can get frame pacing working perfectly with DG, it becomes an interesting solution.
I would like to see power usage numbers but,
you basically have around the same perf from 7850+2133ram+r7 250 GDDR5 as to another fast enough CPU + 7770 alone

looking on newegg the price difference is $30 (with no rebates),
but it might get worse for the 7770 because it could be EOL soon!?

edit: 8 ROPs?
the 7750 apparently have 16ROPs for the same 512/32

 

[CHADBOGA]

So IGP's continue to be an unsuitable choice for anyone remotely interested in mainstream games.

 

[Enigmoid]

Nice work. Only thing I can say is that quite a few of the tests were run at unplayable fps on the igps.

 

[Erenhardt]

Nice graphs. just what I was looking for!
Let me quote myself from yesterday:

Now I want a laptop with a8-7600 and r7 240/250 running dual graphics!
As shown in tom's review, dual graphics almost doubles the performance of 240 (when it works ). From 22 fps, it boosted to 42 in bioshock:



Here, you can see that it would be where r7 250 is (22->42fps):

[IMG]

Want more? Go for 250 and gain 30+% as shown in hardware.fr (7850, but it shoudn't be much different with 7600):

[IMG]

Suddenly, it performs like hd7770.
No 6pin connectors, possible passive design.


AMD doesn't need special tasks dedicated to igp if they nail dual graphics. Not until CPU performance is on the lower side. Dedicating physics to igp would be beneficial when pairing apu with powerful dgpu. No one will do this with AMD APUs due to their lower CPU performance. AMD targets low end, where more rendering performance is always welcome. The idea would be good with intel chips, because there, the igp is wasted as soon you plug any dgpu. I'm just afriad intel doesn't care enough to do something like this.[/QUOTE]

Now, this review backs my points perfectly. What is worth noting is that the gt 640 in the review is GDDR5 gk208 - which is basically a gtx650. There is not a whole lot between it and kaveri, to the point where one can cherry pick a benchmark in which kaveri beats gtx650. That means Kaveri with 2133MHz memory is as fast/faster than the original gt640 GDDR3, which is where most of the market is believed to be now (gtx650 and below).

DualGraphics performance is amazing, shame it doesn't work in nvidia titles. I wonder if there is any chance it will work someday, but I would not put my hopes high taking latest incidents with TWIMTBP titles.
Now the plan is to get kaveri and use it for a year or so, and then upgrade with r7 240/250 DG when budged recovers.

 

[SPBHM]

Now, this review backs my points perfectly. What is worth noting is that the gt 640 in the review is GDDR5 gk208 - which is basically a gtx650. There is not a whole lot between it and kaveri, to the point where one can cherry pick a benchmark in which kaveri beats gtx650. That means Kaveri with 2133MHz memory is as fast/faster than the original gt640 GDDR3, which is where most of the market is believed to be now (gtx650 and below).
.

it says 64bits, that's half the of what the 650 have.
64bit GDDR5 should be as good as 128bit DDR3 numbers but I think it's worse because the card will only use half the "ROPs"!?

at the same time it's 1250MHz, while most DDR3 cards use 800 or 900.
so it's like 40GB/s, for this 640 64bit GDDR5

 

[DGLee]

My compliments on the review. I especially liked the way you made the summary graphs at the end. The only thing missing, and you probably don't have the equipment to test it , is some measure of microstutter with the crossfire combinations.

As for kaveri, I am not especially impressed, except for niche uses, like SFF systems for light gaming without a discrete card. Even with crossfire seeming to work better, if I were going to use a discrete card , I would just get a cheaper CPU and a 7770 or higher card.

But again, my compliments on your testing and presentation of the data. Nice job.

Edit the A8 when available would be interesting to test in crossfire. It is much more price competitive vs an athlon x4 750k, and doesn't seem to lose too much performance. At its current price, it is hard to justify purchasing a 7850k plus a discrete card, when one could get an athlon x4 plus HD7770 for approximately the same price as the APU alone.

Thanks for your opinion. I also would love to test the whole things above via FCAT but I can't afford them... Later I'll cover frametimes instead.

 

[DGLee]

it's interesting to see how close the 7850K and Iris Pro are, for 2 reasons, one is the fact that AMD can get same/higher performance with no ultra fast cache, just DDR3, but at same time Intel have an IGP basically as fast as the best IGP AMD (with all their GPU expertise) have, but obviously both are not competing at the same price, so it's a clear win for AMD.

if AMD can get frame pacing working perfectly with DG, it becomes an interesting solution.
I would like to see power usage numbers but,
you basically have around the same perf from 7850+2133ram+r7 250 GDDR5 as to another fast enough CPU + 7770 alone

looking on newegg the price difference is $30 (with no rebates),
but it might get worse for the 7770 because it could be EOL soon!?

edit: 8 ROPs?
the 7750 apparently have 16ROPs for the same 512/32

Yes. Processor Graphics of Kaveri has only 2 Render back-ends which are equivalent to 8 Color ROPs. Indeed it can't be paired with 7750(and so are 7770/7730, in other words, Cape Verde in general) which seriously implies that Processor Graphics of Kaveri is more like to Oland(R7 250/240, have 8 ROPs) rather than Cape Verde(7770/7750/7730, 16 ROPs) despite its huge SP amount. And this is why I guess Kaveri's SPs are reserved for GPGPU/computational usage under post-HSA era rather than for graphics output.