Intel adds Iris Pro Skylake-H - 128Mb of eDRAM, 30% graphical improvement

[CommandMan7]

After the introduction of 128Mb eDRAM in select Haswell and Broadwell processors, Intel has brought eDRAM forward into new Skylake-H BGA (Soldered) chips. Since Haswell, the amount of execution units of the top-end Iris Pro Graphics have increased, going from 40 to 48 to 72 execution units. The jump from Broadwell's 48 to Skylake's 72 makes the Skylake 9th Generation GT4e Iris Pro around 30% faster. This puts the Skylake Iris Pro just a bit faster than the GTX 750 and just a hair slower than the R7 250X. Of course, all of the announced chips are BGA, AKA non-socketed, so these processors will only be available inside laptops and mobile workstations. 

 

There is one significant difference between Broadwell and Skylake eDRAM, however:

Quote

The eDRAM for Skylake is different to that found in previous Crystal Well implementations: in the last generation, the eDRAM acted as a victim cache to the L3 cache, meaning that evicted cache lines from L3 would add up in the eDRAM and be quick for re-reading without having to access main memory. The downside to this is that data could not end up in eDRAM without being used first, giving initial data read latencies the same performance as previous processors. Ultimately this is still good for graphics and gaming, where textures are re-read from memory frequently. The new arrangement for the eDRAM in these Skylake processors has placed the eDRAM in a different part of the chain, between the System Agent and the DDR memory. This means that the eDRAM acts as a DRAM buffer, with 50 GBps bandwidth in each direction to the LLC, but is also accessible for early reads/writes by any device that needs memory access through the system agent (i.e. anything through PCIe). Previously this was not possible, but now it means that Skylake's eDRAM implementation should offer a speedup in many more scenarios that before. 

What this means is that the eDRAM is no longer acting as a L4 cache, but instead a completely separate buffer for DRAM. This shouldn't have any notable impact on performance, but still a notable change nonetheless.

 

Opinion: I think this is making it clear that large on-die caches are definitely the way to go about integrated graphics. After all of AMD's effort to focus on APU's, they're getting beaten at what many consider to be one of the only strengths of AMD's processors at the moment - their graphics. Of course, AMD still holds the graphical advantage on all the important price points, but there's no reason they shouldn't be able to claim the crown as well. Hopefully AMD will use the faster speed of DDR4 and their experience in HBM to make their Zen APU's break far into the mid-range enthusiast market. If AMD can pull off the same integrated graphics improvements as Intel (and they easily should be able to) I'd reckon that we could see the highest end Zen APU's matching GTX 960 or perhaps even GTX 970 performance. 

 

Source:

http://www.anandtech.com/show/10281/intel-adds-crystal-well-skylake-processors-65w-edram

[Sakkura]

9 minutes ago, CommandMan7 said:

Opinion: I think this is making it clear that large on-die caches are definitely the way to go about integrated graphics. After all of AMD's effort to focus on APU's, they're getting beaten at what many consider to be one of the only strengths of AMD's processors at the moment - their graphics. Of course, AMD still holds the graphical advantage on all the important price points, but there's no reason they shouldn't be able to claim the crown as well. Hopefully AMD will use the faster speed of DDR4 and their experience in HBM to make their Zen APU's break far into the mid-range enthusiast market. If AMD can pull off the same integrated graphics improvements as Intel (and they easily should be able to) I'd reckon that we could see the highest end Zen APU's matching GTX 960 or perhaps even GTX 970 performance. 

Intel is just barely beating AMD on GPU performance, despite a crushing advantage in process node. That means AMD is doing much better than Intel on the actual architecture side (and/or driver optimization). Once AMD gets a modern process node again, expect their APUs to easily outperform Intel again.

[Notional]

3 minutes ago, Sakkura said:

Intel is just barely beating AMD on GPU performance, despite a crushing advantage in process node. That means AMD is doing much better than Intel on the actual architecture side (and/or driver optimization). Once AMD gets a modern process node again, expect their APUs to easily outperform Intel again.

Not just that, AMD also supports things like 10 bit colour and HDR. I doubt Intel will. And of course Intel chips are useless for anything gaming, unlike AMD APU's. In reality Intel's iGPU's are really just useful for super basic desktop image or some sort of low end compute, like streaming.

[Prysin]

48 minutes ago, CommandMan7 said:

 The jump from Broadwell's 48 to Skylake's 72 makes the Skylake 9th Generation GT4e Iris Pro around 30% faster. This puts the Skylake Iris Pro just a bit faster than the GTX 750 and just a hair slower than the R7 250X.

 

Oh, but @patrickjp93, wasnt GT4e supposed to have 50% higher performance. Oh wait, you were saying it was supposed to match up to a GTX 950. You did so more then once. Hell you were adamant it would be AT LEAST faster then a 750Ti, which is slightly faster then a 250X.

 

GET REKT

 

[Prysin]

1 hour ago, Sakkura said:

Intel is just barely beating AMD on GPU performance, despite a crushing advantage in process node. That means AMD is doing much better than Intel on the actual architecture side (and/or driver optimization). Once AMD gets a modern process node again, expect their APUs to easily outperform Intel again.

14nm AMD APUs will most likely beat GT4e and the subsequent next gen of Intel "APUs" by 20-25%

[deviant88]

Intel sucks, it doesnt look like it because theres no competition due to other suckers that didnt know how to sail the AMD ship.

[zMeul]

1 hour ago, Sakkura said:

Intel is just barely beating AMD on GPU performance, despite a crushing advantage in process node. That means AMD is doing much better than Intel on the actual architecture side (and/or driver optimization). Once AMD gets a modern process node again, expect their APUs to easily outperform Intel again.

despite providing better iGPs in their APUs, AMD had a hard time selling APUs and their market strategy failed hard

Intel could use some random iGPs in their CPUs and still sell a ton of them - what AMD didn't get: people don't buy integrated graphics for graphical performance

[MrDynamicMan]

This is still dumb. Their best iGPUs should be on CPUs that would not typically be paired with a dgpu. When they put anything other than the most basic iGPU on a top tier chip theyre just wasting die space.

[Prysin]

30 minutes ago, zMeul said:

despite providing better iGPs in their APUs, AMD had a hard time selling APUs and their market strategy failed hard

Intel could use some random iGPs in their CPUs and still sell a ton of them - what AMD didn't get: people don't buy integrated graphics for graphical performance

mostly because they dont KNOW how good the APUs are.

 

AMD needs to get their heads outta their asses and start to properly market their stuff.

[patrickjp93]

2 hours ago, Sakkura said:

Intel is just barely beating AMD on GPU performance, despite a crushing advantage in process node. That means AMD is doing much better than Intel on the actual architecture side (and/or driver optimization). Once AMD gets a modern process node again, expect their APUs to easily outperform Intel again.

It's winning handily, but also remember Intel's iGPUs are much smaller than AMD's (until GT4e where Intel now has 576 SPs to AMD's 512).

 

Intel isn't trying to really compete in high-end gaming either.

 

1 hour ago, Prysin said:

Oh, but @patrickjp93, wasnt GT4e supposed to have 50% higher performance. Oh wait, you were saying it was supposed to match up to a GTX 950. You did so more then once. Hell you were adamant it would be AT LEAST faster then a 750Ti, which is slightly faster then a 250X.

 

Hardly. Care to look at the throttling charts? There's plenty of performance left on the table by my count. I'll wait for Gigabyte's superior SFFPC before pinning down a verdict.

[Mew]

I wish laptop makers would use iris and iris pro cpus in all of their products, much like Apple did with Haswell. 

[Sakkura]

37 minutes ago, patrickjp93 said:

It's winning handily, but also remember Intel's iGPUs are much smaller than AMD's (until GT4e where Intel now has 576 SPs to AMD's 512).

 

Intel isn't trying to really compete in high-end gaming either.

It's not winning handily. And it's only smaller because of the massive process node advantage. You cannot compare the number of shaders, since there are different approaches to the architecture (getting many low-performance shaders or fewer high-performance shaders). Performance per transistor would be a much more fair way to measure how well each architecture utilizes what it's given, but that's hard to do because transistor counts are fluffy.

[patrickjp93]

Just now, Sakkura said:

It's not winning handily. And it's only smaller because of the massive process node advantage. You cannot compare the number of shaders, since there are different approaches to the architecture (getting many low-performance shaders or fewer high-performance shaders). Performance per transistor would be a much more fair way to measure how well each architecture utilizes what it's given, but that's hard to do because transistor counts are fluffy.

Intel also uses far fewer transistors, so really, where are you going with this? It's 1.7 billion transistors to a Skylake Quad with GT2, and it's 2.3 to a Skylake Quad with GT4e, including the eDRAM.

 

It's 3.1 billion to a Carrizo quad.

[patrickjp93]

3 hours ago, Notional said:

Not just that, AMD also supports things like 10 bit colour and HDR. I doubt Intel will. And of course Intel chips are useless for anything gaming, unlike AMD APU's. In reality Intel's iGPU's are really just useful for super basic desktop image or some sort of low end compute, like streaming.

Intel already announced full 10-bit support for Kaby Lake. As for HDR, it's unknown, but there's no reason to doubt Intel will support it. Stealing the entry professional photo and video editing market from both AMD and Nvidia is a decent chunk of change.

[Notional]

7 minutes ago, patrickjp93 said:

Intel already announced full 10-bit support for Kaby Lake. As for HDR, it's unknown, but there's no reason to doubt Intel will support it. Stealing the entry professional photo and video editing market from both AMD and Nvidia is a decent chunk of change.

Most photo pros use Apple anyways. Workstations are AMD bases sure, but they also tend to use laptops which are only Intel, so I doubt there is a huge market potential there.

[Prysin]

59 minutes ago, patrickjp93 said:

It's winning handily, but also remember Intel's iGPUs are much smaller than AMD's (until GT4e where Intel now has 576 SPs to AMD's 512).

 

Intel isn't trying to really compete in high-end gaming either.

 

Hardly. Care to look at the throttling charts? There's plenty of performance left on the table by my count. I'll wait for Gigabyte's superior SFFPC before pinning down a verdict.

Actually, if you go by the size of the GPU, in comparison to the rest of the package. Intels iGPU uses a higher % of the die then AMDs. It is 60% (Broadwell SKUs) vs 40-50% (Godavari).

 

So intel has less shaders on "more" space. That means their GPU design is of MUCH MUCH lower density. Then again, they are using a Nvidia based IP are they not? That would explain the low shader count + high clock speed combo.

 

Sure, Intel has MUCH more advanced CPU cores, but even at more then HALF the node, they still need a larger portion of their die for iGPU.

[patrickjp93]

18 minutes ago, Prysin said:

Actually, if you go by the size of the GPU, in comparison to the rest of the package. Intels iGPU uses a higher % of the die then AMDs. It is 60% (Broadwell SKUs) vs 40-50% (Godavari).

 

So intel has less shaders on "more" space. That means their GPU design is of MUCH MUCH lower density. Then again, they are using a Nvidia based IP are they not? That would explain the low shader count + high clock speed combo.

 

Sure, Intel has MUCH more advanced CPU cores, but even at more then HALF the node, they still need a larger portion of their die for iGPU.

Using a higher % of the die doesn't mean much when the die is still less than half the size of AMD's. Intel has very dense CPU libraries that don't require nearly the transistor counts AMD's do.

 

Intel has less shaders on less space but in greater proportions of real estate.

 

Again, you can say nothing about density unless you actually get the transistor counts and exact die areas. You're assuming far too much without having the necessary information. A Godavari die is more than 300mm sq.. Skylake is around 150.

[Prysin]

6 minutes ago, patrickjp93 said:

Using a higher % of the die doesn't mean much when the die is still less than half the size of AMD's. Intel has very dense CPU libraries that don't require nearly the transistor counts AMD's do.

 

Intel has less shaders on less space but in greater proportions of real estate.

 

Again, you can say nothing about density unless you actually get the transistor counts and exact die areas. You're assuming far too much without having the necessary information. A Godavari die is more than 300mm sq.. Skylake is around 150.

Kaveri/Godavari APU package are 245mm2.

 

A Skylake i5 6600 is 177mm2

https://www.techpowerup.com/cpudb/1828/core-i5-6600

 

Iris Pro GT4e SKUs will be larger. So they will be closer to 190-210mm2 for sure

[Sakkura]

39 minutes ago, patrickjp93 said:

Intel also uses far fewer transistors, so really, where are you going with this? It's 1.7 billion transistors to a Skylake Quad with GT2, and it's 2.3 to a Skylake Quad with GT4e, including the eDRAM.

 

It's 3.1 billion to a Carrizo quad.

Source please.

[patrickjp93]

7 minutes ago, Prysin said:

Kaveri/Godavari APU package are 245mm2.

 

A Skylake i5 6600 is 177mm2

https://www.techpowerup.com/cpudb/1828/core-i5-6600

 

Iris Pro GT4e SKUs will be larger. So they will be closer to 190-210mm2 for sure

Sorry, was thinking Carrizo (stupid half-generation BS AMD). Regardless, the premise stands.

 

Also, no, the 6600K is 122.4

http://www.anandtech.com/show/9505/skylake-cpu-package-analysis

[patrickjp93]

4 minutes ago, Sakkura said:

Source please.

Carrizo 3.1 billion: http://wccftech.com/amd-carrizo-apu-isscc-2015-presentation-leaked-5-ipc-gain-x86-steamroller-die-consists-31-billion-transistors/

 

Skylake: 1.75 billion: 

https://en.wikipedia.org/wiki/Transistor_count

[vanished]

Since the early days, the whole point of the integrated graphics (in laptops at least) was to give people something just "good enough" for general desktop use that would consume far less power than the dedicated graphics card.  I'm all for them making more powerful versions, but I hope they don't lose sight of that original goal; if the integrated graphics start getting as power hungry as the cards they are supposed to replace, what's the point?  Even if it's an ultra book or budget computer with only integrated graphics, it makes me think there should be two stages of it (one for max performance, one for battery life).  Unless of course they can make it more powerful while retaining the ability to scale down to super low power when possible.  And if, so, I'd like them to just licence that to nvidia or someone so we can do away with integrated and just have a GTX 1180 (or whenever it would come out ) on all the time that uses less than a fraction of a watt at idle   

[Prysin]

2 minutes ago, Sakkura said:

Source please.

Kaveri is around 2.41Billion Transistors

There is no reliable source of Carrizo transistor count out yet. But around 3 billion isnt unlikely, as AMD increased density of their CPU liberaries by a huge margin

[Sakkura]

4 minutes ago, patrickjp93 said:

Carrizo 3.1 billion: http://wccftech.com/amd-carrizo-apu-isscc-2015-presentation-leaked-5-ipc-gain-x86-steamroller-die-consists-31-billion-transistors/

 

Skylake: 1.75 billion: 

https://en.wikipedia.org/wiki/Transistor_count

Your source does not say anything about the transistor count of Skylake GT4e. There's not even any actual source for the Skylake GT2 number.

[patrickjp93]

1 minute ago, Prysin said:

Kaveri is around 2.41Billion Transistors

There is no reliable source of Carrizo transistor count out yet. But around 3 billion isnt unlikely, as AMD increased density of their CPU liberaries by a huge margin

AMD disclosed the exact number as 3.1 billion at ISSCC 2015. Where the Hell have you been?