Skylake S SiSoft Benchmark: Confusion and 30% IPC Improvement Over Haswell?[UPDATE]

[dizmo]

It is, because that means ES NDAs are going to start lifting for Broadwell desktop soon, but the premise of this makes no sense.

Yet another thing that I have no idea what it means. *shrug*

[Opcode]

Yet another thing that I have no idea what it means. *shrug*

ES – Engineering Sample (a test chip relative to the final sample)

NDA – Non-Disclosure Agreement (basically if you get a ES you have to agree to NDA which means you cannot speak about it at all e.g. specifications, temperatures, etc)

[Sintezza]

Intel has stated Skylake will support both memory controllers.

 

Source?

 

Because i dont believe a single thing of that, maybe only for the notebook cpu´s.

But not MS desktop cpu´s.

[patrickjp93]

DDR4 isn't exactly all that consumer friendly yet (price/performance) and probably wont be for some time to come. It's a wise choice to stick to supporting DDR3 in order to not break consumer banks. Sales would more than likely fall off for Skylake otherwise. Who really wants to dump $200+ into new ram. FM2 has the north bridge controller completely integrated. Carrizo integrated the south bridge controller as well. FM2r2 chips have supported DDR3 and DDR4 for some time now. Tho AMD seems to not want to transition to DDR4 as it's still not cost effective for the consumer.

Carrizo's not coming to Desktop as a heads up. That said, good to know.

[patrickjp93]

Source?

 

Because i dont believe a single thing of that, maybe only for the notebook cpu´s.

But not MS desktop cpu´s.

http://www.techpowerup.com/205231/how-intel-plans-to-transition-between-ddr3-and-ddr4-for-the-mainstream.html

 

Yes Microsoft desktop CPUs. Why do you people continue to doubt me?

[Opcode]

Carrizo's not coming to Desktop as a heads up. That said, good to know.

It's really all here say right now as some sources say no while others say yes. The problem we are having is AMD hasn't released any information regarding the desktop variant (likely due to it coming later than the mobile variants). All we have right now is older roadmaps that does point to Carrizo being on the FM2r2 package. There could possibly be some major developments for the desktop variant or it could of gotten lost in space. If that's the case we will see a Kaveri refresh. Personally I don't think AMD is going to drop desktop variants of Carrizo. The architecture is the largest step forward out of any Bulldozer based iteration. Besides they will have nothing to bank on for the entire year of 2015. At least Carrizo would serve as a time gap filler for the company other than not pushing out new desktop microprocessors for an entire year.

[patrickjp93]

It's really all here say right now as some sources say no while others say yes. The problem we are having is AMD hasn't released any information regarding the desktop variant (likely due to it coming later than the mobile variants). All we have right now is older roadmaps that does point to Carrizo being on the FM2r2 package. There could possibly be some major developments for the desktop variant or it could of gotten lost in space. If that's the case we will see a Kaveri refresh. Personally I don't think AMD is going to drop desktop variants of Carrizo. The architecture is the largest step forward out of any Bulldozer based iteration. Besides they will have nothing to bank on for the entire year of 2015. At least Carrizo would serve as a time gap filler for the company other than not pushing out new desktop microprocessors for an entire year.

Since the desktop market is still contracting, it wouldn't be surprising. If AMD produces stock it can't sell, that hurts it financially. It's not out of the question.

[Sintezza]

http://www.techpowerup.com/205231/how-intel-plans-to-transition-between-ddr3-and-ddr4-for-the-mainstream.html

 

Yes Microsoft desktop CPUs. Why do you people continue to doubt me?

 

Because your statements, also that from techpower up or what not, is nowhere backedup by intel.

As far as i know, But i readed an article about skylake cpu´s for laptops and what not, who do indeed support DDR3 but No information about the upcomming desktop platform at all.

 

As far as intel road maps go, and information on the net about the new upcomming socket 1151 boards Z170 and H170 chipsets.

Only DDR4 support.

[cesrai]

Carrizo's not coming to Desktop as a heads up. That said, good to know.

No it will probably come to desktop, isn't that they have Carrizo and Carrizo-L ? Carrizo being for desktop and Carrizo-L being for desktop, I couldn't find any information from AMD saying Carrizo is not coming to desktop.

[patrickjp93]

Because your statements, also that from techpower up or what not, is nowhere backedup by intel.

As far as i know, But i readed an article about skylake cpu´s for laptops and what not, who do indeed support DDR3 but No information about the upcomming desktop platform at all.

 

As far as intel road maps go, and information on the net about the new upcomming socket 1151 boards Z170 and H170 chipsets.

Only DDR4 support.

Did you miss where they cited Intel?

[patrickjp93]

No it will probably come to desktop, isn't that they have Carrizo and Carrizo-L ? Carrizo being for desktop and Carrizo-L being for desktop, I couldn't find any information from AMD saying Carrizo is not coming to desktop.

While it is a rumor, it does make a lot of business sense. The desktop market is contracting even now, and AMD is not going to gain market share with it. It would be like throwing money in the toilet. Beyond that, the benchmarks for the first Carrizo-based chips show HBM didn't make it in, so there's not going to be a graphics performance jump of much magnitude. You'll get some of the GCN 1.2 features and be able to do dual graphics with the full array of cards, but it's not ground-breaking enough for desktop people.

 

Servers/supercomputing is the only major market where Carrizo can really make a splash if system engineers pick up HSA and use some HBM as RAM.

[Speedbird]

Dude, Skylake-S is the same as Haswell-DT. It's the desktop, LGA1151 series. It has nothing to do with the S suffix on SKUs.

[patrickjp93]

Dude, Skylake-S is the same as Haswell-DT. It's the desktop, LGA1151 series. It has nothing to do with the S suffix on SKUs.

facepalm* No, but it's damn obvious that it's both in this case! Sheesh... And you people call me a snob... Is it so hard to think multiple steps ahead?

[cesrai]

While it is a rumor, it does make a lot of business sense. The desktop market is contracting even now, and AMD is not going to gain market share with it. It would be like throwing money in the toilet. Beyond that, the benchmarks for the first Carrizo-based chips show HBM didn't make it in, so there's not going to be a graphics performance jump of much magnitude. You'll get some of the GCN 1.2 features and be able to do dual graphics with the full array of cards, but it's not ground-breaking enough for desktop people.

 

Servers/supercomputing is the only major market where Carrizo can really make a splash if system engineers pick up HSA and use some HBM as RAM.

They didn't say Graphics jump but they said power consumption jump, makes sense what you said above.

[patrickjp93]

They didn't say Graphics jump but they said power consumption jump, makes sense what you said above.

but without a graphics jump, their APUs won't sell because the CPU portion still won't compete that well against SandyBridge. If AMD has next to not hope of selling desktop systems well, why make a desktop product?

[cesrai]

but without a graphics jump, their APUs won't sell because the CPU portion still won't compete that well against SandyBridge. If AMD has next to not hope of selling desktop systems well, why make a desktop product?

Yeah that didn't make sense to me but that is what they said in the video, when I asked John Tylor he said it's a big GPU jump, so I'm going with John and hope it's a GPU jump.

[Opcode]

While it is a rumor, it does make a lot of business sense. The desktop market is contracting even now, and AMD is not going to gain market share with it. It would be like throwing money in the toilet. Beyond that, the benchmarks for the first Carrizo-based chips show HBM didn't make it in, so there's not going to be a graphics performance jump of much magnitude. You'll get some of the GCN 1.2 features and be able to do dual graphics with the full array of cards, but it's not ground-breaking enough for desktop people.

 

Servers/supercomputing is the only major market where Carrizo can really make a splash if system engineers pick up HSA and use some HBM as RAM.

It all depends exactly on what Excavator brings. CPU side it's looking to be catching up with Sandy Bridge. Tho that really doesn't matter as even Steamroller is enough to drive the 512 on-board SPU's. GCN should also be updated to GCN 1.2 which brings forth higher tessellation performance and delta compression which would help majorly with slow DDR3 bandwidth. The chip should be an all around upgrade from Kaveri and well worthy upgrade at that. There's so many of us ready to ditch discrete because APU's bring enough grunt for the gaming that we do.

 

but without a graphics jump, their APUs won't sell because the CPU portion still won't compete that well against SandyBridge. If AMD has next to not hope of selling desktop systems well, why make a desktop product?

Their graphics should be bumped up to GCN 1.2 and like stated above we don't know exactly how fast Excavator is in the real world. What we do know from SiSoft validations (rumor) is that a 2.6 GHz Excavator is just as fast as a 3.2 GHz A10-7300. If true Excavator is 19% faster clock for clock than Kaveri. The only problems the desktop variant would face is IPC being lost due to lower base clocks. Tho as long as HDL doesn't hinder overclocking performance any us really won't care as at 4.5 GHz the chip would be phenomenal. Bump up GPU clocks to 1 GHz+ couple it with 2600 MHz DDR3 and you got a beastly little gaming SoC on your hands.

[patrickjp93]

It all depends exactly on what Excavator brings. CPU side it's looking to be catching up with Sandy Bridge. Tho that really doesn't matter as even Steamroller is enough to drive the 512 on-board SPU's. GCN should also be updated to GCN 1.2 which brings forth higher tessellation performance and delta compression which would help majorly with slow DDR3 bandwidth. The chip should be an all around upgrade from Kaveri and well worthy upgrade at that. There's so many of us ready to ditch discrete because APU's bring enough grunt for the gaming that we do.

 

Their graphics should be bumped up to GCN 1.2 and like stated above we don't know exactly how fast Excavator is in the real world. What we do know from SiSoft validations (rumor) is that a 2.6 GHz Excavator is just as fast as a 3.2 GHz A10-7300. If true Excavator is 19% faster clock for clock than Kaveri. The only problems the desktop variant would face is IPC being lost due to lower base clocks. Tho as long as HDL doesn't hinder overclocking performance any us really won't care as at 4.5 GHz the chip would be phenomenal. Bump up GPU clocks to 1 GHz+ couple it with 2600 MHz DDR3 and you got a beastly little gaming SoC on your hands.

HDL will hurt overclocking, no doubt. AMD already had a massive density and transistor count advantage over Intel. It's gonna run hotter, not a doubt in my mind.

 

As per driving 512 SPUs, only if you're running DX 12, OpenCL 2.0, or the not yet released OpenGL 5 built for unified memory, or if the program was built using the HSA API's. Otherwise it'll use the same buffer copying and passing seen for dGPU gaming. That's just a fact.

 

As for 2600MHz memory...MAYBE. Most Kaveri aren't happy past 2400.

 

Bumping the GPU clocks might net you 900 MHz, but 1 GHz I think will be pushing it due to the whole chip using HDL. It's just too bad AMD couldn't get HBM on there. That said, I suspected servers would get it first and the earliest consumers would see it would be with Zen, or on a new enthusiast platform so they can segment their markets a bit the same way Intel does, but with actually legitimate features.

[Opcode]

HDL will hurt overclocking, no doubt. AMD already had a massive density and transistor count advantage over Intel. It's gonna run hotter, not a doubt in my mind.

 

As per driving 512 SPUs, only if you're running DX 12, OpenCL 2.0, or the not yet released OpenGL 5 built for unified memory, or if the program was built using the HSA API's. Otherwise it'll use the same buffer copying and passing seen for dGPU gaming. That's just a fact.

 

As for 2600MHz memory...MAYBE. Most Kaveri aren't happy past 2400.

 

Bumping the GPU clocks might net you 900 MHz, but 1 GHz I think will be pushing it due to the whole chip using HDL. It's just too bad AMD couldn't get HBM on there. That said, I suspected servers would get it first and the earliest consumers would see it would be with Zen, or on a new enthusiast platform so they can segment their markets a bit the same way Intel does, but with actually legitimate features.

Heat shouldn't be a major issue as power consumption of the cores dropped 40% at the same time. Hotter focal points but less overall generated heat (65w TDP vs 95w TDP).

 

Steamroller was more than enough of pushing the iGPU under a full load even in BF4 as demonstrated prior to the Kaveri launch. Ultra settings at 720p maintaining 25-30 FPS while all that being under DirectX 11 as well. No doubt higher tessleation performance and delta compression which can alleviate up to 40% of memory bandwidth will be a big improvement for APU gaming. Top that off with Mantle and the new Omega drivers is just extra gravy on the potatoes.

 

Getting 2600 MHz to work even with Kaveri shouldn't be much of an issue. You can even loosen the timings if need be because timings have no major relevant factor on memory bandwidth which is the most important factor. The problem Kaveri faced was 2600 MHz DDR3 was still expensive and in short supply at it's launch. Plus you need a higher end board that can handle them high frequencies (like my ASRock A85X Extreme6). You can pick up a 8 GB kit of 2666 MHz that will do 2800 MHz on Newegg now for the same price as a 1600 MHz kit.

 

The on-board GCN units already use HDL so there shouldn't be any hindrance to pushing the iGPU past 1 GHz. It's easily obtainable with Kaveri and even with my A10-6800k using VLIW4 I can push 1086 MHz core on stock volts completely stable. I would agree HBM would of made a huge difference in gaming performance tho I guess AMD is more than likely holding out on the big guns until Zen. To add more of a wow factor to the launch of the new architecture.

[patrickjp93]

Heat shouldn't be a major issue as power consumption of the cores dropped 40% at the same time. Hotter focal points but less overall generated heat (65w TDP vs 95w TDP).

 

Steamroller was more than enough of pushing the iGPU under a full load even in BF4 as demonstrated prior to the Kaveri launch. Ultra settings at 720p maintaining 25-30 FPS while all that being under DirectX 11 as well. No doubt higher tessleation performance and delta compression which can alleviate up to 40% of memory bandwidth will be a big improvement for APU gaming. Top that off with Mantle and the new Omega drivers is just extra gravy on the potatoes.

 

Getting 2600 MHz to work even with Kaveri shouldn't be much of an issue. You can even loosen the timings if need be because timings have no major relevant factor on memory bandwidth which is the most important factor. The problem Kaveri faced was 2600 MHz DDR3 was still expensive and in short supply at it's launch. Plus you need a higher end board that can handle them high frequencies (like my ASRock A85X Extreme6). You can pick up a 8 GB kit of 2666 MHz that will do 2800 MHz on Newegg now for the same price as a 1600 MHz kit.

 

The on-board GCN units already use HDL so there shouldn't be any hindrance to pushing the iGPU past 1 GHz. It's easily obtainable with Kaveri and even with my A10-6800k using VLIW4 I can push 1086 MHz core on stock volts completely stable. I would agree HBM would of made a huge difference in gaming performance tho I guess AMD is more than likely holding out on the big guns until Zen. To add more of a wow factor to the launch of the new architecture.

30%, not 40. Also, you underestimate localized heat. That's why running 90s* overclocks isn't safe. The average core temp may be 93, but I guarantee you there are localities at Tjunction. It's exactly the reason Haswell is such an unstable overclocker. The ALU clusters and control logic are all nested together in one place, and the registers are chilling out elsewhere. If the registers at at 60 and the ALUs are at 95, your average readout is only 88.

 

Ultra 720p at 30 fps... Because that's such an accomplishment... 1080p high 45 fps should be the minimum goal for Carrizo honestly. Otherwise it's not a gaming SOC. If AMD invested in more cache the way Intel does, a lot of that bottleneck would be gone. Delta compression saves you a grand total of 14% time and 25% space. The noisier a picture, the less compressible, so you won't get a uniform performance gain for a lot of games. Any time there's weather involved especially, almost no gain. It's the same compression technique VOIP uses, and we all know how clunky that is. Tessellation will help, but I reserve judgment as to how much. How many games use Mantle? Not many. Their Omega drivers are a sham. They refuse to fix DX 11 whereas Nvidia proved it could fix DX 11 support to equal the gains of Mantle. Mantle is proving to be useless actually. AMD's Omega drivers only make Mantle look good against their own DX 11 support.

 

Latency affects bandwidth. Take a networking course and you'll understand why. If you take an additional 2 microseconds to hit the RAM and get data back, the overall bandwidth doesn't help much. That's why I always buy on absolute latency, not just on bandwidth. CL/(RAM clock/2) = absolute latency.

 

6800 was the previous generation: richland, and Kaveri took a massive hit on overclock range. Richland overclocked very well. Don't expect Carrizo to.

 

I suspect AMD held back either due to die space or cost to the consumer sending their products into Intel's upper price ranges, a space they cannot hope to compete in on a bulldozer derivative.

[Opcode]

30%, not 40. Also, you underestimate localized heat. That's why running 90s* overclocks isn't safe. The average core temp may be 93, but I guarantee you there are localities at Tjunction. It's exactly the reason Haswell is such an unstable overclocker. The ALU clusters and control logic are all nested together in one place, and the registers are chilling out elsewhere. If the registers at at 60 and the ALUs are at 95, your average readout is only 88.

 

Ultra 720p at 30 fps... Because that's such an accomplishment... 1080p high 45 fps should be the minimum goal for Carrizo honestly. Otherwise it's not a gaming SOC. If AMD invested in more cache the way Intel does, a lot of that bottleneck would be gone. Delta compression saves you a grand total of 14% time and 25% space. The noisier a picture, the less compressible, so you won't get a uniform performance gain for a lot of games. Any time there's weather involved especially, almost no gain. It's the same compression technique VOIP uses, and we all know how clunky that is. Tessellation will help, but I reserve judgment as to how much. How many games use Mantle? Not many. Their Omega drivers are a sham. They refuse to fix DX 11 whereas Nvidia proved it could fix DX 11 support to equal the gains of Mantle. Mantle is proving to be useless actually. AMD's Omega drivers only make Mantle look good against their own DX 11 support.

 

Latency affects bandwidth. Take a networking course and you'll understand why. If you take an additional 2 microseconds to hit the RAM and get data back, the overall bandwidth doesn't help much. That's why I always buy on absolute latency, not just on bandwidth. CL/(RAM clock/2) = absolute latency.

 

6800 was the previous generation: richland, and Kaveri took a massive hit on overclock range. Richland overclocked very well. Don't expect Carrizo to.

 

I suspect AMD held back either due to die space or cost to the consumer sending their products into Intel's upper price ranges, a space they cannot hope to compete in on a bulldozer derivative.

AMD uses a single package temperature unlike Intel's per core sensor. There hasn't been an issue with heat at all thus far and I doubt that will change all too much either. Like said power consumption dropped 40% on the CPU side of things with TDP at 65w for the entire chip on the same node, heat shouldn't be an issue. My A10-6800k hits 50-55C under a full load (Prime95) with my cooling setup in a non-optimal configuration (exhausting heat out the back). Nothing has really changed with Carrizo other than a possible bump in GCN architecture, denser cores, and the fusion hub being incorporated on die.

The 28 nm chip measures 244.62 mm2 and packs more than 3.1 billion transistors. Its new Excavator core is 23% smaller and uses 40% less power than AMD’s previous x86 core.

 

That's quite an accomplishment especially compared to the competition. Kaveri will run 1080p at acceptable settings and frame rates as well. Tho like said memory bandwidth is the ultimate underlying issue holding the iGPU back. I benched Tomb Raider on my A10-6800k (1600x900 because that's all my monitor supported at the time). Given Kaveri's superior GPU architecture and higher SPU count. 1080p gaming shouldn't be nothing shy of achievable. Compare my A10-6800k results to the results of the A10-7800.

 

Tomb Raider - (Stock DDR3 @ 1866 MHz)

 

1600x900 Normal Settings

• A10-6800k - 29.3 FPS
• A10-7800 - 38 FPS
• A10-7800 (w/ 2133 MHz) - 42 FPS

1920x1080 Normal Settings

• A10-6800k - Untested
• A10-7800 - 28 FPS
• A10-7800 (w/ 2133 MHz) - 31 FPS

2400+ MHz numbers are that much more impressive with Kaveri. With Carrizo stepping the game up with possibly GCN 1.2 the delta compression should reap much larger marginal performance gains than frequency bumps like I outlined above. Delta color compression will be beneficial in all games as there's constantly the same color pixel rendered in sequence. Take drawing things such as SkyBox's as an example. GCN also brings 2-4x tessellation throughput which as said is extra gravy on the potatoes coupled with Omega drivers and Mantle. Your hate speeches towards either of them two things is quite foolish as they've proven to bring a lot of performance improvements in gaming (including Mantle on an Intel system).

 

DRAM latency plays very little role on memory bandwidth. I have tested this time and time again with my APU. System DDR3 with rough timings usually has similar latency to that of GDDR5 (~40 ns). It doesn't matter if the ram refreshes a whole cycle or two faster if the frequency is horrible to begin with. By your theory 1600 MHz @ 6-6-6-16 would be better than 2400 MHz @ 11-11-11-31 for gaming. At this point it seems like you're trying to pick out underlying issues that don't exist in the first place in order to correlate to what sounds like a intelligent response. I've done testing, none of what you implied is true. Frequency is overall primary factor in memory bandwidth. If you're buying an APU go for the highest frequency possible as timings are irrelevant.

 

The overclock headroom of Richland is quite impressive but also is much exaggerated. Trust me I have a decently binned A10-6800k and I can do 5.3 GHz but only with a ridiculous amount of volts 1.55+ and completely unstable (would take degrading volts to stabilize even 5.0 GHz). The 4.8 GHz mark is essentially the high end overclocking bar for Richland while 4.7 GHz seems to be the norm for Kaveri. Carrizo will be a different story because of the use of HDL. Tho we simply don't know how it will impact overclocking.

 

Die space is probably completely consumed with the integration of the south bridge fusion hub. They could of still stacked it on-package tho I presume with HBM only coming in 1 GB flavors it's probably not too cost effective for an APU at this point in time.

[patrickjp93]

AMD uses a single package temperature unlike Intel's per core sensor. There hasn't been an issue with heat at all thus far and I doubt that will change all too much either. Like said power consumption dropped 40% on the CPU side of things with TDP at 65w for the entire chip on the same node, heat shouldn't be an issue. My A10-6800k hits 50-55C under a full load (Prime95) with my cooling setup in a non-optimal configuration (exhausting heat out the back). Nothing has really changed with Carrizo other than a possible bump in GCN architecture, denser cores, and the fusion hub being incorporated on die.

 

That's quite an accomplishment especially compared to the competition. Kaveri will run 1080p at acceptable settings and frame rates as well. Tho like said memory bandwidth is the ultimate underlying issue holding the iGPU back. I benched Tomb Raider on my A10-6800k (1600x900 because that's all my monitor supported at the time). Given Kaveri's superior GPU architecture and higher SPU count. 1080p gaming shouldn't be nothing shy of achievable. Compare my A10-6800k results to the results of the A10-7800.

 

Tomb Raider - (Stock DDR3 @ 1866 MHz)

 

1600x900 Normal Settings

• A10-6800k - 29.3 FPS
• A10-7800 - 38 FPS
• A10-7800 (w/ 2133 MHz) - 42 FPS

1920x1080 Normal Settings

• A10-6800k - Untested
• A10-7800 - 28 FPS
• A10-7800 (w/ 2133 MHz) - 31 FPS

2400+ MHz numbers are that much more impressive with Kaveri. With Carrizo stepping the game up with possibly GCN 1.2 the delta compression should reap much larger marginal performance gains than frequency bumps like I outlined above. Delta color compression will be beneficial in all games as there's constantly the same color pixel rendered in sequence. Take drawing things such as SkyBox's as an example. GCN also brings 2-4x tessellation throughput which as said is extra gravy on the potatoes coupled with Omega drivers and Mantle. Your hate speeches towards either of them two things is quite foolish as they've proven to bring a lot of performance improvements in gaming (including Mantle on an Intel system).

 

DRAM latency plays very little role on memory bandwidth. I have tested this time and time again with my APU. System DDR3 with rough timings usually has similar latency to that of GDDR5 (~40 ns). It doesn't matter if the ram refreshes a whole cycle or two faster if the frequency is horrible to begin with. By your theory 1600 MHz @ 6-6-6-16 would be better than 2400 MHz @ 11-11-11-31 for gaming. At this point it seems like you're trying to pick out underlying issues that don't exist in the first place in order to correlate to what sounds like a intelligent response. I've done testing, none of what you implied is true.

 

The overclock headroom of Richland is quite impressive but is much exaggerated. Trust me I have a decently binned A10-6800k and I can do 5.3 GHz but only with a ridiculous amount of volts 1.55+ and completely unstable (would take degrading volts to stabilize even 5.0 GHz). The 4.8 GHz mark is essentially the high end overclocking bar for Richland while 4.7 GHz seems to be the norm for Kaveri.

 

Die space is probably completely consumed with the integration of the south bridge fusion hub. They could of still stacked it on-package tho I presume with HBM only coming in 1 GB flavors it's probably not too cost effective for an APU at this point in time.

AMD can get away with having package heat sensing because its ALUs (2 vs. Haswell's 4), FPUs (1 per 2 cores vs. Intel's 1 per 1 core), control logic, branch prediction, and other ancillary systems are allocated and spaced in such a way almost no workload can cause a particular region to far exceed the core average. You could say that is an advantage for AMD, but only until you realize they are increasing the density of every single component, crushing them closer together. Without seeing a high-res die shot or looking at the circuits myself, I can't predict with certainty a local heat region will be produced, but the likelihood is greatly increased. The upside at stock clocks is that power consumption is greatly reduced. You could have the same TDP while having less power going in (if the core is smaller but everything is more tightly packed you have increased the heat density, even if raw heat is decreased or left even).

 

How is that impressive exactly? Intel's had these same features for a while. With Intel it's just been about experimenting with GPU architecture, and if Core M is anything to go by, Intel is going to make Iris Pro 6200 competitive to AMD's iGPUs up through Kaveri while still maintaining a huge lead in CPU performance. What many people fail to realize is Intel and AMD both made some of the most critical instructions take only 1 cycle each a long time ago. Software won't speed up through architectural improvements much longer. Intel shaving 2 of its 5 cycles from the floating point operations going from Haswell to Broadwell is a standing anomaly in this, but overall I'm sure you see why. Consumer software doesn't take advantage of the SIMD instructions Intel and AMD made available.years ago. Do you really think Haswell's 4 ALUs could be taken on by Ivy Bridge's 3? In fully optimized software, no. 5% IPC increases on average fail to take into account the treasure trove of extra resources made available. AMD's innovations were long overdue in truth. Getting voltage regulator, memory controller, and fusion hub onboard should have happened years earlier.

 

Skyboxes are easy to compress, and so is grass, but a good game doesn't have a ton of the same color sitting around. We'll have a mixed bag on delta compression benefits. Tessellation is good. Don't get me wrong, but without HBM you're still gonna be choked compared to Intel's 64MB of L4 cache that runs bandwidth in the multiple 100s of Gbps vs. 2400MHz RAM running I think at 62 ish Gbps...

 

I don't hate Mantle. There's just existing proof AMD didn't fix as much as they claimed, and Nvidia's own DX 11 drivers prove this. Before an after comparisons have been made, and AMD's own Mantle games only do well against their own DX 11 drivers. Nvidia's new drivers show a great improvement by fixing interfaces with DX 11 that put them on equal gains with AMD's Mantle, whereas AMD refuses to fix their DX 11 problems in an effort to market Mantle. It's a sham of a product as it exists, even if a good idea.

 

That is not at all what I'm saying. I'm saying absolute latency affects bandwidth. If you shaved half the latency from GDDR5, you'd gain 35% performance because you could keep the GPU better fed. If my bandwidth is 400Mbps, it matters if I deliver you the whole block at the last nanosecond vs. being able to deliver enough chunks per moment to keep the system fed. Latency greatly affects bandwidth. This is why 2400 CL 10 is better than 2666 CL 15. Absolute latency truly does matter.

 

I greatly suspected HBM would be too expensive for the consumer at the start. Anyone who banked on Hynix selling cheap was living in a fantasy world. As per the fusion hub, have you ever seen one bare? It's maybe 10 square mm.

[Opcode]

AMD can get away with having package heat sensing because its ALUs (2 vs. Haswell's 4), FPUs (1 per 2 cores vs. Intel's 1 per 1 core), control logic, branch prediction, and other ancillary systems are allocated and spaced in such a way almost no workload can cause a particular region to far exceed the core average. You could say that is an advantage for AMD, but only until you realize they are increasing the density of every single component, crushing them closer together. Without seeing a high-res die shot or looking at the circuits myself, I can't predict with certainty a local heat region will be produced, but the likelihood is greatly increased. The upside at stock clocks is that power consumption is greatly reduced. You could have the same TDP while having less power going in (if the core is smaller but everything is more tightly packed you have increased the heat density, even if raw heat is decreased or left even).

 

How is that impressive exactly? Intel's had these same features for a while. With Intel it's just been about experimenting with GPU architecture, and if Core M is anything to go by, Intel is going to make Iris Pro 6200 competitive to AMD's iGPUs up through Kaveri while still maintaining a huge lead in CPU performance. What many people fail to realize is Intel and AMD both made some of the most critical instructions take only 1 cycle each a long time ago. Software won't speed up through architectural improvements much longer. Intel shaving 2 of its 5 cycles from the floating point operations going from Haswell to Broadwell is a standing anomaly in this, but overall I'm sure you see why. Consumer software doesn't take advantage of the SIMD instructions Intel and AMD made available.years ago. Do you really think Haswell's 4 ALUs could be taken on by Ivy Bridge's 3? In fully optimized software, no. 5% IPC increases on average fail to take into account the treasure trove of extra resources made available. AMD's innovations were long overdue in truth. Getting voltage regulator, memory controller, and fusion hub onboard should have happened years earlier.

 

Skyboxes are easy to compress, and so is grass, but a good game doesn't have a ton of the same color sitting around. We'll have a mixed bag on delta compression benefits. Tessellation is good. Don't get me wrong, but without HBM you're still gonna be choked compared to Intel's 64MB of L4 cache that runs bandwidth in the multiple 100s of Gbps vs. 2400MHz RAM running I think at 62 ish Gbps...

 

I don't hate Mantle. There's just existing proof AMD didn't fix as much as they claimed, and Nvidia's own DX 11 drivers prove this. Before an after comparisons have been made, and AMD's own Mantle games only do well against their own DX 11 drivers. Nvidia's new drivers show a great improvement by fixing interfaces with DX 11 that put them on equal gains with AMD's Mantle, whereas AMD refuses to fix their DX 11 problems in an effort to market Mantle. It's a sham of a product as it exists, even if a good idea.

 

That is not at all what I'm saying. I'm saying absolute latency affects bandwidth. If you shaved half the latency from GDDR5, you'd gain 35% performance because you could keep the GPU better fed. If my bandwidth is 400Mbps, it matters if I deliver you the whole block at the last nanosecond vs. being able to deliver enough chunks per moment to keep the system fed. Latency greatly affects bandwidth. This is why 2400 CL 10 is better than 2666 CL 15. Absolute latency truly does matter.

 

I greatly suspected HBM would be too expensive for the consumer at the start. Anyone who banked on Hynix selling cheap was living in a fantasy world. As per the fusion hub, have you ever seen one bare? It's maybe 10 square mm.

AMD does package thermal monitoring probably due to it being a bit more cost efficient for mass production. What's the sense of putting a sensor on every core if you can just read package thermals and get an overall temperature reading. I will agree to core density going down 23% will result in hotter focal points with thermals. Tho the shrink in size also brought forth a 40% cut in power consumption. Which is every sense of the word a massive reduction in power consumption. If there was no power cuts at all then I would be on board with there possibly being thermal problems. Tho with the cores consuming essentially almost half the power than before. It's hard to say heat will be an underlying issue. Especially with the cores only shrinking by a full node improvement (essentially 22nm ready).

 

It's impressive when you stack it up against the competition. HD 4600 and Iris Pro 5200 isn't exactly competitive in the gaming market. And before you jump onto the "but Intel is making it bettah" train it still doesn't compare to what's on the market now. If you compare generation to generation Kaveri would stack up against Iris Pro 5200. In which in performance benchmarks Iris Pro 5200 can stand up to Kaveri but falls short in quite a few titles. Iris Pro 6200 will probably level the playing field with Kaveri. Which would of been a tough outcome for AMD if Intel didn't do what they did by not providing lower end CPU's with Broadwell. If Intel launched a i3-5130 or whatever it will be called and if it packs full fledged Iris Pro 6200 then yea AMD could be looking at some problems. Tho really we have not a single clue of what's going on behind closed doors. AMD hasn't released any information about Carrizo. The first lineup of processors we will see will be Carrizo-L (Puma+) sometime this holiday season. The higher end Carrizo and the desktop variant isn't going to be released until next year. So even the possibility of HBM on package isn't left in the dark. We just simply don't know what's going on with the Excavator based APU's. The memory controller has been on die for years now long before Phenom II at that. Voltage regulators are suppose to be queued for addition with Zen. And AMD is actually far ahead of Intel on putting both fusion hubs on die as well. FM2 had the north bridge controller integrated completely on die. Carrizo adds a stripped down version of the south bridge controller. Which is suppose to disable itself when you drop it into existing FM2r2 motherboards.

 

Delta compression will always have mixed results for every single frame rendered. Tho in most games there's a lot of common colors drawn on screen due to textures being more flat because of detail being utilized elsewhere. Games like CS:GO will more than likely take a huge advantage of this technology. While other games like Battlefield 4 not so much. Tho the performance gains will still be there and for a iGPU that is starving from low memory bandwidth to begin with every little bit will help. As you can see from my numbers above that the jump from 1866 MHz to 2133 MHz which is only a 267 MHz jump wielded roughly an extra 4-5 FPS. From my own testing every bump in memory frequency has this effect.

• 1600 -> 1866 = ~5 FPS Gain
• 1866 -> 2133 = ~5 FPS Gain
• 2133 -> 2400 = ~5 FPS Gain

How much more memory bandwidth can only 260 MHz of memory clock bring? Probably not a whole lot in comparison of cutting memory bandwidth nearly in half at best with delta compression. The technology is actually why the new AMD and Nvidia cards run such small memory interfaces. Both AMD and Nvidia use delta compression with Tonga and Maxwell. It should serve for a decent performance boost for APU's.

 

Nvidia's DX11 drivers are far better than AMD's without question. There is kind of a disappointment that AMD couldn't optimize their drivers for DX11 accordingly. Tho the company really has shifted focus more on Mantle than anything. As a majority of games now days and upcoming titles leverage the new API. What's in question isn't really the performance of the API when it comes to frame rates. But more over how much more batches can be sent to the GPU without major frame rate hits. Mantle does bring beneficial performance due to being low level but the underlying key is how hard can you tax the API before it becomes crippled. There's only so much stuff that can be drawn and done in game with DirectX 11 because of it's restricting kernel layers. This is why Microsoft worked in collaboration with AMD on developing DirectX 12 to also be low level. We really can't see how Mantle stands out until we get a game on the market like Oxide Game's tech demo that's built specifically for Mantle. Which we will more than likely have to wait until the launch of DirectX 12.

 

Tho we simply cannot shave latency off of GDDR5 else it would of been done already. The latency with system memory is no longer than GDDR5. The switch from CL9 to CL11 is so minuscule that you'd likely see less than a 0.1 ns difference in latency. With one whole nanosecond being one billionth of a second. You can start to understand the latency involved is not even worth mentioning. The only thing I think we will see these latencies play a role in is with frame times. Tho like said in the real world we're messing with such extremely small numbers that it's change is really better left unmentioned.

 

Carrizo runs a stripped down fusion hub (two SATA 6.0 Gbps ports, four USB 3.0 ports, and four USB 2.0 ports). Which is a good start as the integrated fusion hub is directed towards the mobile market. Personally it's enough for me on the desktop. Tho FM2r2 will fall back on the motherboards south bridge once dropped into the socket.

[patrickjp93]

@Opcode I reserve judgment until I see a die shot, but it's going to become a problem eventually if AMD wants to catch up in CPU compute.

 

Broadwell U will come with HD 5500, 6000, and Iris 6100 (no L4 cache). That much has been confirmed just ahead of CES 2015. Furthermore, you also have to remember the stream processor count Intel has been fighting AMD with: 7 SIMD units * 40 EUs = 280 stream processors. 54.7% of AMD's count. 48*7 = 336 SPs in Iris Pro 6200 (65.6% as many). That's extremely impressive when you think about it.

 

Would have* Also, it can be done, because otherwise we wouldn't have multiple CAS Latencies in RAM modules. Can it be shaved all the way down to 20? I kinda doubt it. Hopefully HBM fixes some of them.

 

CL 9 @ 2400 MHz Absolute latency = 9/(2400/2) = 7.5 nanoseconds

CL 11                                               = 11/(2400/2) = 9.1666... nanoseconds

 

That's 3 cycles wasted on the 4790k at 4.4 GHz.

 

As per the fusion hub, it doesn't matter much outside development platforms, tablets, or embedded systems. It's a good way for AMD to try breaking into those realms via ARM, but I don't see it working out well on the x86 side.

 

If Intel decided it wanted to and dumped 800 million more transistors into graphics (to close the gap in transistor counts between them and AMD), you'd have ~1950 SPs on one processor by some rough estimations. Now add HDL onto that for both the CPU and GPU side, yielding 30% more density... Intel is just warming up honestly. The graphics engineers testify they've always been given the scraps when die space remains. What if Intel told them to have a hayday? I think Zen will be the deciding moment. Either Intel will actually be forced to compete, or the market will permanently stagnate. 2016 will be interesting.

[Opcode]

@Opcode I reserve judgment until I see a die shot, but it's going to become a problem eventually if AMD wants to catch up in CPU compute.

 

Broadwell U will come with HD 5500, 6000, and Iris 6100 (no L4 cache). That much has been confirmed just ahead of CES 2015. Furthermore, you also have to remember the stream processor count Intel has been fighting AMD with: 7 SIMD units * 40 EUs = 280 stream processors. 54.7% of AMD's count. 48*7 = 336 SPs in Iris Pro 6200 (65.6% as many). That's extremely impressive when you think about it.

 

Would have* Also, it can be done, because otherwise we wouldn't have multiple CAS Latencies in RAM modules. Can it be shaved all the way down to 20? I kinda doubt it. Hopefully HBM fixes some of them.

 

CL 9 @ 2400 MHz Absolute latency = 9/(2400/2) = 7.5 nanoseconds

CL 11                                               = 11/(2400/2) = 9.1666... nanoseconds

 

That's 3 cycles wasted on the 4790k at 4.4 GHz.

 

As per the fusion hub, it doesn't matter much outside development platforms, tablets, or embedded systems. It's a good way for AMD to try breaking into those realms via ARM, but I don't see it working out well on the x86 side.

 

If Intel decided it wanted to and dumped 800 million more transistors into graphics (to close the gap in transistor counts between them and AMD), you'd have ~1950 SPs on one processor by some rough estimations. Now add HDL onto that for both the CPU and GPU side, yielding 30% more density... Intel is just warming up honestly. The graphics engineers testify they've always been given the scraps when die space remains. What if Intel told them to have a hayday? I think Zen will be the deciding moment. Either Intel will actually be forced to compete, or the market will permanently stagnate. 2016 will be interesting.

The thing is you won't see any difference in performance from cas latency because total memory latency will remain roughly around 40 ns. Which really doesn't serve any benefit for the iGPU. Keep in mind that is what we are talking about here. Otherwise Intel favors high frequency over tight timings CPU side anyways. You aren't going to notice any difference between 2400 MHz CL9 vs 2400MHz CL11 as the bandwidth will be exactly the same. The only difference is how often the DRAM refreshes its banks. Throughput is the only important part when it comes to graphical computing when you're working with limited bandwidth to begin with. I worked the numbers and each jump in memory frequency is roughly a 12% increase in memory bandwidth. So with delta compression we could see a performance difference equal to anywhere between a 260 MHz and a 866 MHz boost on existing clock frequency's. Broken down at it's best 1866 MHz memory would show performance numbers similar to using 2666 MHz memory at full scale. Which breaks down to around a 5-15 FPS gain with delta compression at the same frequency. Which is quite impressive seeing as how you need change nothing in your system. As I've stated a dropped in Carrizo would immediately show performance gains out of the gate. AMD has their own ARM core coming in 2016 alongside Zen known as K12. As of right now custom ARM architectures rule the low powered market and that's not going to change for a long while. K12 will be FinFET and is expected to be roughly 12x faster than the Cortex-A57. So the fusion hub is more than likely for simplifying the mobile market (e.g. laptops and notebooks).