[AT] Iris Pro in a socket

[Enigmoid]

We are not evaluating APUs here as a product but only the iGPU performance. Playable frame rates is not what we are seeking for.
Eliminating the CPU influence (lower resolutions and image quality settings) really show the iGPU performance of each architecture. And those AT graphs clearly demonstrate that GPU wise, Haswell IrisPro is not that much more power efficient than Kaveri.

Wow, um lets pick Xtreme settings to eliminate CPU bottleneck. Its not like there is you know, a low setting where the frame rate is about 6x higher clearly showing that a 7850k can power bioshock infinite at 68 fps. It clearly must be tremendously CPU limited at medium at 24 fps.

And when we test mid-low range GPUs like the R260X and 750Ti we must always test at 4K + 8x MSAA because we are not looking at inconsequential things like playable frame rates and target markets. Really gotta let those GPUs go.

 

[NTMBK]

We are not evaluating APUs here as a product but only the iGPU performance. Playable frame rates is not what we are seeking for.
Eliminating the CPU influence (lower resolutions and image quality settings) really show the iGPU performance of each architecture. And those AT graphs clearly demonstrate that GPU wise, Haswell IrisPro is not that much more power efficient than Kaveri.

The eDRAM has a "sweet spot" where it has the most performance impact; this happily falls into the region where performance is actually playable. When you crank the settings up to the point where it is unplayable on all competitors, the eDRAM has less impact.

And no comment on the fact that the 4770R was thermally throttling in those tests due to the poor chassis?

 

[Phynaz]

So, where did you see that Iris Pro creams Kaveri at the same power consumption

Right here:

http://techreport.com/review/26166/gigabyte-brix-pro-reviewed/3

They didn't look at settings that result in 9 FPS frame rates though, if you consider that playable.

 

[Phynaz]

We are not evaluating APUs here as a product but only the iGPU performance.

Correction - you may not be, but we are. What's the point to not look at the entire product? Can you buy an APU without the CPU part enabled?

 

[AtenRa]

Actually, all cranking up to maximum settings does is negate the majority of the advantage derived from the eDRAM due to its limited size. Which is why there's a pretty obvious trend of the AMD parts tested with 1866/2133 MHz DDR3 pull ahead of Iris Pro with 1600 MHz DDR3 as scene complexity/graphics memory footprint increases.

Not true,

That could only be said for 1080p Sleeping Dogs where every Intel iGPU almost produce the same fps.It is also the only game that IrisPro is the Slower Intel iGPU and that could mean that the eDRAM is having a negative impact or the driver is under performing with this eDRAM implementation.

BUT on the other hand, AMD Kaveri scales with TDP and number of Shaders (TMUs) as A10-7850K is faster than 65W and 45W TDP SKUs. Also, Since both 5800K and 6800K also using 2133MHz memory they are far bellow Kaveris performance. That is clearly an architecture/design advantage and not memory bandwidth.

Lets have a look at the other games,

Bioshock, Iris Pro @ 1080p is faster than A10-5800K that using 2133MHz memory in AT review
Tomb Raider, Iris pro @ 1080p is faster than A10-6800K using 2133MHz memory in AT review
Company Of Heroes, Iris Pro is faster than A10-6800K using 2133MHz memory at AT review.

So as we can see, Iris Pro with 1600MHz memory is faster than any Richland with 2133MHz memory (except in Sleeping Dogs). If what you said was correct 5800K/6800K with 2133MHz memory would be faster than IriPro.

Kaveri 95W TDP is faster than Iris Pro across all games at 1080p. Every Kaveri SKU from 45W to 95W using the same 2133MHz memory sees a nice scaling due to more resources of the iGPU (7850K has more Shaders/TMUs than 7700K/7600) and or higher turbo frequency due to higher TDP limit.

This is a GPU performance advantage, not memory bandwidth advantage. At low resolutions Kaveri and IriPro are Memory bandwidth limited. As the resolution and image quality gets higher the different parts of the GPU (Shaders, TMUs, ROPs etc) are becoming the limited factor bottle-necking the GPU. Simple, the memory bandwidth is not you limiting factor anymore because the entire GPU is waiting for the Shaders, the TMUs or the ROPs to finish their calculations first.

 

[AtenRa]

And when we test mid-low range GPUs like the R260X and 750Ti we must always test at 4K + 8x MSAA because we are not looking at inconsequential things like playable frame rates and target markets. Really gotta let those GPUs go.

When you test Discrete GPUs, you have the SAME CPU for every system. Clearly you knew that

 

[Sweepr]

Barely faster at unplayable settings (vs a throtthling Intel system) and still gets smoked CPU-wise by a lower TDP quad-core Haswell with Iris Pro (Core i7-4770R is roughly 2x faster @ Cinebench 11.5 x64 than A10-7850K - 7.13 vs 3.61 pts). Pricing aside, there's no denying that Haswell GT3e is still impressive even with outdated Gen 7.5 graphics.
Cant wait to see what a 95W LGA Broadwell-K with Gen 8 based Iris Pro will do to Excavator.

http://us.hardware.info/reviews/515...ew-amds-new-apu-benchmarks-igpu-cinebench-115
http://us.hardware.info/reviews/518...o-5200-graphics-benchmarks-igpu-cinebench-115

 

[AtenRa]

The eDRAM has a "sweet spot" where it has the most performance impact; this happily falls into the region where performance is actually playable. When you crank the settings up to the point where it is unplayable on all competitors, the eDRAM has less impact.

Yes, at higher Resolutions/Image Quality more parts of the GPU becomes the bottleneck negating the eDRAM advantage. IriPro clearly doesnt have the GPU performance and it entire relies in the eDRAM at lower resolutions/Image Quality. Raise the Resolutions and the GPU architecture/design becomes the Bottleneck, simple as that.
Also at lower resolutions the CPU advantage of the Core i7 is too much to ignore vs Kaveri.

And no comment on the fact that the 4770R was thermally throttling in those tests due to the poor chassis?

The only game that Iris Pro is slower than Vanilla HD4600 is Sleeping Dogs and i dont believe it is due to throttling.

 

[AtenRa]

Correction - you may not be, but we are. What's the point to not look at the entire product? Can you buy an APU without the CPU part enabled?

Well, if you are talking about the Core i5/7R product then yes it is faster than Kaveri at lower playable resolutions/image Quality.

But you only said,

Probably because Haswell Iris Pro creams Kaveri at the same power consumption. Broadwell is only going to get better.

That my friend as Intel refers to, is the iGPU alone (GT3 with eDRAM) and not any SKU, thus why i only evaluated the iGPU alone.

 

[Lepton87]

The pricing listed on ARK for tray pricing is not far off from retail box pricing. For example, the i5 4670K has a box price of $243, whereas the tray price is a whopping... $242.

If Intel's worked out discounts with various OEMs, they're undisclosed, and obviously not listed on the ARK page.

And yes, I fully expect an i5 model to cost less than a 4770K.

I see no mention whether it will be i5 or i7. I would expect an i7 with full cache and HT, otherwise this product will have very limited appeal.

 

[Enigmoid]

When you test Discrete GPUs, you have the SAME CPU for every system. Clearly you knew that

No, you need enough CPU to put the load 100% on the GPU.

 

[Homeles]

I see no mention whether it will be i5 or i7. I would expect an i7 with full cache and HT, otherwise this product will have very limited appeal.

One Broadwell-K GT3 SKU seems illogical to me.

 

[PPB]

Shoot in the dark here, but GT4 might has been a failed attempt by Intel to win console APU.

Yeah, design a sku that will ship more than 1 year later than what the console makers needed to reach their launch date, and that probably costs as much as the 50 to 90% of the console price by itself.

Totally makes sense.

Iris pro was designed per apple's request. Every sucessor only accomodates to continue that request.

 

[Shivansps]

Broadwell was scheduled for Q4 2013, GT4 for all we know its not gona exist, if that happens its because it was planned and cancelled.

 

[bunnyfubbles]

If this is a Haswell Iris Pro performance then i dont see whats the fuss about it.

a lot of it has to do with the 128MB L4 Crystal Well cache meant to improve memory bandwidth for the iGPU, but is also available to the CPU... its pretty impressive stuff, allows lower power mobile CPUs able to come close to stock 4770K performance in many areas, and win when the cache can be leveraged (ie QuickSync), imagining what might be done with an unlocked chip is very enticing indeed

 

[PPB]

a lot of it has to do with the 128MB L4 Crystal Well cache meant to improve memory bandwidth for the iGPU, but is also available to the CPU... its pretty impressive stuff, allows lower power mobile CPUs able to come close to stock 4770K performance in many areas, and win when the cache can be leveraged (ie QuickSync), imagining what might be done with an unlocked chip is very enticing indeed

Wasnt this debunked with the R series reviews done here on AT? As far as those tests showed that L4 cache was pretty much innocuous on that regard.

I have never seen the point of having such a big cache to leverage more performance if it tends to fall off really dramatically as you tend to up the settings (specially the resolution). Considering those iris pro skus ship on rather well specced machines (usually with higher than 1080p displays) i dont see the point in needing to reduce the resolution beyond the native res of the display panel so it isnt a slideshow. You may as well go for a dGPU solution with similar costs but without of the fear of suffering such a dramatic performance degradation for just trying to play at the native res your +$1000 portable/AIO is sporting.

 

[Enigmoid]

Wasnt this debunked with the R series reviews done here on AT? As far as those tests showed that L4 cache was pretty much innocuous on that regard.

I have never seen the point of having such a big cache to leverage more performance if it tends to fall off really dramatically as you tend to up the settings (specially the resolution). Considering those iris pro skus ship on rather well specced machines (usually with higher than 1080p displays) i dont see the point in needing to reduce the resolution beyond the native res of the display panel so it isnt a slideshow. You may as well go for a dGPU solution with similar costs but without of the fear of suffering such a dramatic performance degradation for just trying to play at the native res your +$1000 portable/AIO is sporting.

Its not just the cache. Intel igp architecture (gen 7) has poor texturing performance meaning it sucks at higher settings (or AA).

Even on hd 4600 as you up the settings you will notice this and there is no L4 there.

 

[Shivansps]

Gen 6 has the same problem too, OK performance (more than a GT520/GT610 on most games) but horrible texture fill rate.

 

[mikk]

Its not just the cache. Intel igp architecture (gen 7) has poor texturing performance meaning it sucks at higher settings (or AA).

Exactly. Broadwell has almost twice the sampler per EU if the GenX schematic from Tom Piazza belongs to Broadwell. There are big gains possible especially with MSAA where Haswell is really poor.

 

[IntelUser2000]

And assuming that the HD4600 and Iris Pro numbers from here - http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/15 - are simply 30x those in the review you linked then we can add...
4 samplers HD 5200: 15.56 @ 1.3GHz - 11.97 ponits/GHz
2 samplers HD 4600: 7.64 @ 1.35GHz - 5.66 points/GHz

Take a closer look at the Iris Pro review Khato, its stated in Gigatexels/s. There's no way it can get that high. That's a typo. We don't know what they are really substituting there. Megapixels/s would make it too low.

http://www.anandtech.com/show/7677/amd-kaveri-review-a8-7600-a10-7850k/14

This is better, unfortunately we don't see Iris Pro there. The good thing is now we see HD 4600. The increase in Texel Fill is nearly exactly the clock speed difference between HD 4600 and HD 4000. The funny thing is no one bothers to make proper theoretical comparisons anymore. Why not test the fillrate on the Iris Pro system. Almost feels like its being manipulated or something...

if it was then why do the AMD and NVIDIA parts also show fractional pixel/clock?

Actually, other than the Gen 7 iGPUs, the others achieve 80-90% of the theoretical fillrate. The Gen 7 iGPUs are significantly off others in "efficiency". They are at 66%.

Plus, Intel has already done it before. The most important is though how it fares relatively to similar architecture, that is, against their own. I've correctly predicted back in the GMA X3000 days based on 3DMark results that the fillrates were "off".

Once I found the presentation... It quite clearly states Fill rate in MPixel/sec there as well - what does that have to do with texture sampling? It's purely a ROP measurement.

Because it says "Pixels/s" does not mean ROP, because its technically not right to refer to as pixels/s since the unified pipeline, it changed to Single Texture performance versus Multi-texture fillrate.

Bay Trail has half the texture performance of GT1 in Ivy Bridge. I can't show that to you right now, because I can't find it but when I do I will.

I have another theory that's even harder to prove. I think the effective FLOPs is also 3/4 the peak as well. Funny thing is Tom's presentation about Future GenX chip was showing 96EUs having 2TFlop throughput at 1GHz. That makes no sense considering how current GenX EU's work. Perhaps they meant its "2TFlop equivalent"?(because 96EUs x 1GHz x boosting clock to full which means 33% increase = 2TFlops).

 

[IntelUser2000]

Its not just the cache. Intel igp architecture (gen 7) has poor texturing performance meaning it sucks at higher settings (or AA).

Even on hd 4600 as you up the settings you will notice this and there is no L4 there.

There's probably more than that. The AA implementation doesn't seem top notch either.

Another thing is drivers. It seems trivial, but the different arrangement of caching and memory on the Iris Pro may not play so nice with some games. It's quite likely we'll see some games see rather nice boost as they get such things fixed.

The fact is the GPU manufacturers do quite specific optimizations down to settings and resolutions, and Intel still isn't putting much effort as Nvidia and AMD do for drivers. I mean, look at the GUI for the drivers. It's bland and featureless. What is this from SiS or something?

 

[mikk]

That makes no sense considering how current GenX EU's work.

You know how Gen8 works? How does it work?

 

[Khato]

Take a closer look at the Iris Pro review Khato, its stated in Gigatexels/s. There's no way it can get that high. That's a typo. We don't know what they are really substituting there. Megapixels/s would make it too low.

http://www.anandtech.com/show/7677/amd-kaveri-review-a8-7600-a10-7850k/14

This is better, unfortunately we don't see Iris Pro there. The good thing is now we see HD 4600. The increase in Texel Fill is nearly exactly the clock speed difference between HD 4600 and HD 4000. The funny thing is no one bothers to make proper theoretical comparisons anymore. Why not test the fillrate on the Iris Pro system. Almost feels like its being manipulated or something...

And the number there for the i7-4770k's HD 4600 is 7.69 versus the 7.64 that I derived by dividing the Iris Pro review numbers by 30 - all the other numbers match up as well with that manipulation, so what reason is there to believe that it won't be correct for Iris Pro? As well, Texel Fill rate shown in your link has the i7-3770k's HD 4000 at 6.12 while the i7-4770k's HD 4600 is at 7.69 - that's a 1.256x gain for a 1.087x clock speed increase.

Actually, other than the Gen 7 iGPUs, the others achieve 80-90% of the theoretical fillrate. The Gen 7 iGPUs are significantly off others in "efficiency". They are at 66%.

And that's precisely the point. You're basing your speculation off how the competition behaves compared to theoretical and assuming the same should apply to Intel's design. It doesn't.

Bay Trail has half the texture performance of GT1 in Ivy Bridge. I can't show that to you right now, because I can't find it but when I do I will.

Thanks for making me dig around since it yielded a far easier way to demonstrate what I'm saying in the form of GLBenchmark 2.7's texel fill rate test. In the 3DMark test the i7-4770k's HD4600 gets 7.69 GigaTexels/s - divide that by the 1.25 GHz clock frequency and 2 samplers and you arrive at 3.076 texels/clock. Meanwhile if you take the surface pro 2 results in GLBenchmark 2.7's texel fill rate test - https://gfxbench.com/device.jsp?benchmark=gfx27&os=Windows&api=dx&D=Microsoft Surface Pro 2 - of 7,432 MTexels/s and divide by its peak 1 GHz frequency and 2 samplers you arrive at 3.716 texels/clock (93% of the 'ideal' 4 texels/clock). Even more amusing is the AMD side of the equation - the 8670D goes from 17.81 GTexels/s in the 3DMark test to 4.27 GTexels/s - https://gfxbench.com/device.jsp?benchmark=gfx27&os=Windows&api=dx&D=AMD Radeon HD 8670D - in GLBenchmark 2.7. I do believe that such does an excellent job of demonstrating how dependent texture sampling throughput is upon the specific workload, no?

Oh, and as for Baytrail it only ends up at around 1.5 texels/clock in GLBenchmark 2.7 - https://gfxbench.com/device.jsp?benchmark=gfx27&os=Windows&api=dx&D=Asus T100TA Transformer Book - which is around 2/3 of the ~2.3 texels/clock of a sampler in HD4000. There is a reason for such, but it doesn't have to do with clock frequency.

 

[Hugo Drax]

Sounds cool but then as long as the iGPU is only used as a GPU and basically dead silicon with a dGPU, it's not that interesting. I would rather prefer hexa- or octo cores in mainstream with weaker or no iGPU at all.

Most cores sit idle in today's consumer desktop.

 

[Hugo Drax]

I wonder if they will release an i7 non K edition with iris pro.

The problem with the K version is they neutered the TSX feature.