Intel & AMD, Architectural Discussion, How Far Ahead Is Intel ?

[Faa]

You seem to have missed the fact the 8350 had a serious RAM advantage too. 1866 CL 9 vs 1600 CL 9 is a big difference in both bandwidth and absolute latency

9/933 = 9.646 ns

9/800 = 11.25 ns 

 

1/(3.7*10^9) =0.2703 seconds per clock.

 

That's an 8 CPU cycle difference in fetch time after you account for the time to travel from the bus back to the registers.

 

Benches should never have different RAM speeds across different chips unless it's to look at the effect of RAM on performance on 1 or 2 chips.

If he says Sandy bridge has 128bit FP's, thats all we need to know.

[patrickjp93]

If he says Sandy bridge has 128bit FP's, thats all we need to know.

That part is true. The thing is Intel makes up for it with much better pipeline and thread management.

[Faa]

That part is true. The thing is Intel makes up for it with much better pipeline and thread management.

AVX is 256bit. AVX2 just allows the integer units to process 256 bit instructions as well.

[patrickjp93]

AVX is 256bit. AVX2 just allows the integer units to process 256 bit instructions as well.

Yes but the FP units were only 128-bits wide. Two together made the 256-bit unit. It's not far unlike AMD's model except there were 2 128 bit units per core.

[TERAFLOP]

AVX is 256bit. AVX2 just allows the integer units to process 256 bit instructions as well.

Sandy does only have 128bit unit per FPU which combines with a 128bit unit in the integer pipe to process 256bit instructions.

The above slide talks about Haswell not Sandy not sure if you're just clueless or intentionally spreading misinformation.

[GPUXPert]

not sure if you're just clueless or intentionally spreading misinformation.

He's both... actually.

TechFan@tic posted this which should've been pretty much enough to end this pointless argument.

Compared to Sandy Bridge, Bulldozer only has two advantages in FP performance: FMA support and higher 128-bit AVX throughput. There's very little code available today that uses AMD's FMA instruction, while the 128-bit AVX advantage is tangible.

clickity click

 

[GRJohn]

Some details I'd like to see are: what is AMD doing to alleviate its shortcoming in single-thread performance, and what is Intel likely to do in implementing a heterogeneous architecture given it is now working towards unified memory with Skylake. Given Intel already has the superior float performance in its CPU cores, what would it likely do when implementing its newest GPU core architecture? Would it cut down on the CPU float scheduler, or would it develop them both in tandem towards being better?

 

Obviously, if Intel cut down on its CPU core resources, it could potentially fit more cores and increase their clock rates to better equal AMD's and widen the performance gap, though this may also slow down the GPU cores.

Do you have any source on that intel is working towards unified memory in Skylake ? From all the leaks going around seems like Skylake isn't going to do anything like that just a bigger GPU like in Broadwell and a few core tweaks + 512bit AVX.

Jim Keller is working towards a new core that's going to implement low-gate count high frequency design, more focus on floating point performance and higher efficiency.

Also if you follow patents you'll notice that AMD has filed several dozen patents on new cache technologies and decode since Jim Keller joined amd two years ago.

[patrickjp93]

Do you have any source on that intel is working towards unified memory in Skylake ? From all the leaks going around seems like Skylake isn't going to do anything like that just a bigger GPU like in Broadwell and a few core tweaks + 512bit AVX.

Jim Keller is working towards a new core that's going to implement low-gate count high frequency design, more focus on floating point performance and higher efficiency.

Also if you follow patents you'll notice that AMD has filed several dozen patents on new cache technologies and decode since Jim Keller joined amd two years ago.

Which is great considering their cache structure vs. Intel has been like a WWII German U-Boat vs. a U.S. 90s nuclear submarine with stealth.

 

From an authoritative source on scientific computing tech, which cites Intel's own roadmap. 

http://www.scientificcomputing.com/articles/2014/01/mobile-tech-between-rock-and-hard-place

[TechFan@ic]

Do you have any source on that intel is working towards unified memory in Skylake ? From all the leaks going around seems like Skylake isn't going to do anything like that just a bigger GPU like in Broadwell and a few core tweaks + 512bit AVX.

Jim Keller is working towards a new core that's going to implement low-gate count high frequency design, more focus on floating point performance and higher efficiency.

Also if you follow patents you'll notice that AMD has filed several dozen patents on new cache technologies and decode since Jim Keller joined amd two years ago.

Which is great considering their cache structure vs. Intel has been like a WWII German U-Boat vs. a U.S. 90s nuclear submarine with stealth.

 

From an authoritative source on scientific computing tech, which cites Intel's own roadmap. 

http://www.scientificcomputing.com/articles/2014/01/mobile-tech-between-rock-and-hard-place

Unified memory is the only way forward, using dedicated memory pools for each different type of processor will be looked upon in a few years time like it was ancient technology.

It's asinine, cumbersome and simply inefficient.

Heterogenous computing is the only way to continue performance scaling once Moore's law slows down to a halt.

AMD figured this out nearly a decade ago when they started negotiations to acquire ATi. Intel being a much larger comapny was significantly slower to respond but they're moving towards the same goal. HP is developing a fundamentally similar processor that relies on specialized processors for each category or type of workload as well which it calls "The Machine"

[patrickjp93]

Do you have any source on that intel is working towards unified memory in Skylake ? From all the leaks going around seems like Skylake isn't going to do anything like that just a bigger GPU like in Broadwell and a few core tweaks + 512bit AVX.

Jim Keller is working towards a new core that's going to implement low-gate count high frequency design, more focus on floating point performance and higher efficiency.

Also if you follow patents you'll notice that AMD has filed several dozen patents on new cache technologies and decode since Jim Keller joined amd two years ago.

Low gate count, high frequency may sound great to enthusiasts (though who really cares since the 9590 BARELY breaks 5GHz and requires top-end coolers), but for people who want a chip performing great at stock speeds for years, eh, not so much. If all of Broadwell is released at 4GHz + base clock, that means Skylake will release at those speeds with more efficiency gains in addition to the new 512-bit AVX 3.x instructions.

 

I'm excited to see what Keller comes up with, and getting rid of the modular design is music to my ears, though he isn't in charge of the Carrizo project, so that's still going to be a disappointment on the CPU side (though a very neat look at the first truly complete APU with schedulers on the GPU cores). I want AMD to be able to pressure Intel, but it's not looking good unless they really restart their fire in 2015.

 

And with them being held back by global foundries and TSMC, it looks even worse. Intel is reporting the shrink to 10nm is proving less of a struggle than the shrink from 22 to 14. That puts them in a position to be able to react and remain ahead of AMD without having to push new processing nodes on new architectures while bugs are being worked out. AMD is going to have to become a whole new brand of threat and convince the server world to switch, because PC sales are continuing to decline and laptop vendors aren't picking up the Kaveri line.

[TechFan@ic]

I'm excited to see what Keller comes up with, and getting rid of the modular design is music to my ears, though he isn't in charge of the Carrizo project,

He took charge of all projects when he took over, including the Jaguar & Bulldozer development roadmap.

He worked with Puma the successor of Jaguar which already launched so it's much more obvious that he would've worked on Steamroller's successor, Excavator.

 

And with them being held back by global foundries and TSMC, it looks even worse. Intel is reporting the shrink to 10nm is proving less of a struggle than the shrink from 22 to 14. That puts them in a position to be able to react and remain ahead of AMD without having to push new processing nodes on new architectures while bugs are being worked out. AMD is going to have to become a whole new brand of threat and convince the server world to switch, because PC sales are continuing to decline and laptop vendors aren't picking up the Kaveri line.

Surprisingly that really isn't the case. Samsung already began 14LPE production, Glogbalfoundries will begin in Q4 of this year & ramp up products in Q1 2015.

That's only two quarters later than Intel for GF & one quarter later than Intel for Samsung. The majority of process industry experts speculated early on that all major foundries would meet at 14nm and they were seemingly correct.

[patrickjp93]

Unified memory is the only way forward, using dedicated memory pools for each different type of processor will be looked upon in a few years time like it was ancient technology.

It's asinine, cumbersome and simply inefficient.

Heterogenous computing is the only way to continue performance scaling once Moore's law slows down to a halt.

AMD figured this out nearly a decade ago when they started negotiations to acquire ATi. Intel being a much larger comapny was significantly slower to respond but they're moving towards the same goal. HP is developing a fundamentally similar processor that relies on specialized processors for each category or type of workload as well which it calls "The Machine"

Intel was not slower to respond. They waited for AMD to lay the ground work and see if software companies would get behind the idea. It takes Intel far less time to look at what AMD has done, tweak it enough to avoid copyright infringement, and implement it to meet AMD at the summit when HSA would theoretically take off (I predict 3 years, which is in line with Intel's roadmap going to unified memory and then equal scheduling).

 

Intel is a predator and a parasite in a way, but then that's the world of business competition.

[patrickjp93]

He took charge of all projects when he took over, including the Jaguar & Bulldozer development roadmap.

He worked with Puma the successor of Jaguar which already launched so it's much more obvious that he would've worked on Steamroller's successor, Excavator.

That's not what AMD has reported. He's in charge of the Jaguar and Puma projects, but Carrizo belongs to someone else, and I'm pretty sure Keller was also slated to take over the successor to Vishera.

[patrickjp93]

He took charge of all projects when he took over, including the Jaguar & Bulldozer development roadmap.

He worked with Puma the successor of Jaguar which already launched so it's much more obvious that he would've worked on Steamroller's successor, Excavator.

 

Surprisingly that really isn't the case. Samsung already began 14LPE production, Glogbalfoundries will begin in Q4 of this year & ramp up products in Q1 2015.

That's only two quarters later than Intel for GF & one quarter later than Intel for Samsung. The majority of process industry experts speculated early on that all major foundries would meet at 14nm and they were seemingly correct.

Except they're not. Intel's production began at the start of 2014. Everyone else is behind EXCEPT Samsung, who is arguably Asia's Intel on everything electronic BUT CPUs (relatively adoption of ARM excluded).

 

Also, AMD is hobbled by TSMC specifically on the GPU side, and TSMC is having problems even today despite their reports 20nm production has begun. The yields are still lower than average much to Nvidia's chagrin wanting the 880 to cost less to the consumer than the 780.

[TechFan@ic]

That's not what AMD has reported. He's in charge of the Jaguar and Puma projects, but Carrizo belongs to someone else, and I'm pretty sure Keller was also slated to take over the successor to Vishera.

There is one development team for all of AMD's cores including x86 & ARM & it's headed by Jim Keller.

 

Except they're not. Intel's production began at the start of 2014. Everyone else is behind EXCEPT Samsung, who is arguably Asia's Intel on everything electronic BUT CPUs (relatively adoption of ARM excluded).

 

Also, AMD is hobbled by TSMC specifically on the GPU side, and TSMC is having problems even today despite their reports 20nm production has begun. The yields are still lower than average much to Nvidia's chagrin wanting the 880 to cost less to the consumer than the 780.

 

14nm mass production at Intel was delayed from late Q1 2014 to Q2 2014 according to Digitimes, Samsung already began production and again Globalfoundries will begin at Q4 of this year so again they're only 2 quarters behind.

AMD can manufacture at Samsung since both GF & Samsung agreed to use the same 14LPE & LPP using the Copy-Smart methodology in all of their fabs. And Samsung has already caught up to Intel.

http://techreport.com/news/26336/globalfoundries-licenses-samsung-process-tech-grants-amd-access-to-finfets

 

Broadwell isn't coming to the desktop until Q2 2015.

http://www.fudzilla.com/home/item/35154-broadwell-14nm-desktop-comes-late-in-q2-2015

Meaning AMD can have 14nm product to compete with Broadwell in Q4 2015 or even Q3 if they decide to make chips at Samsung's US fab.

Also it's worth noting that AMD has officially announced that it's going to begin GPU production at Globalfoundries & will have GPU products based on GF silicone this year.

http://www.maximumpc.com/amd_tweaks_wafer_agreement_globalfoundries_include_gpus_2014

[patrickjp93]

There is one development team for all of AMD's cores including x86 & ARM & it's headed by Jim Keller.

 

14nm mass production at Intel was delayed from late Q1 2014 to Q2 2014 according to Digitimes, Samsung already began production and again Globalfoundries will begin at Q4 of this year so again they're only 2 quarters behind.

AMD can manufacture at Samsung since both GF & Samsung agreed to use the same 14LPE & LPP using the Copy-Smart methodology in all of their fabs. And Samsung has already caught up to Intel.

http://techreport.com/news/26336/globalfoundries-licenses-samsung-process-tech-grants-amd-access-to-finfets

 

Broadwell isn't coming to the desktop until Q2 2015.

http://www.fudzilla.com/home/item/35154-broadwell-14nm-desktop-comes-late-in-q2-2015

Meaning AMD can have 14nm product to compete with Broadwell in Q4 2015 or even Q3 if they decide to make chips at Samsung's US fab.

Also it's worth noting that AMD has officially announced that it's going to begin GPU production at Globalfoundries & will have GPU products based on GF silicone this year.

http://www.maximumpc.com/amd_tweaks_wafer_agreement_globalfoundries_include_gpus_2014

AMD can't have 14nm as nothing on their roadmap suggests it and Carrizo is already being manufactured at the 20nm process, and that is their Q1 2015 release.

 

Samsung doesn't exactly compete with Intel and nor does it have a dog in the x86 fight. Its focus is on flash for its SSDs and ARM for its phones and refrigerators (and RAM controllers).

 

Also, Fudzilla is demonstrably false. Intel announced just last week Broadwell would be out for Christmas (much to Apple's delight).

 

Broadwell desktop may not be out until Q1 2015, but that is not the same as Broadwell not being out and 14nm not being out.

 

AMD is far behind and has only begun researching the 14nm process, not production.

[TechFan@ic]

AMD can't have 14nm as nothing on their roadmap suggests it and Carrizo is already being manufactured at the 20nm process, and that is their Q1 2015 release.

 

Samsung doesn't exactly compete with Intel and nor does it have a dog in the x86 fight. Its focus is on flash for its SSDs and ARM for its phones and refrigerators (and RAM controllers).

 

Also, Fudzilla is demonstrably false. Intel announced just last week Broadwell would be out for Christmas (much to Apple's delight).

 

Broadwell desktop may not be out until Q1 2015, but that is not the same as Broadwell not being out and 14nm not being out.

 

AMD is far behind and has only begun researching the 14nm process, not production.

#1 AMD already taped out 14nm silicone.

http://www.xbitlabs.com/news/other/display/20131017231002_AMD_to_Tape_Out_First_20nm_14nm_FinFET_Chips_Within_Next_Two_Quarters.html

#2 GF & Samsung use the exact same process technology, 14LPE & LPP (which GF licensed from Samsung) & AMD can choose either to make its chips as part of the deal that GF made with Samsung.

http://www.eetimes.com/document.asp?doc_id=1321974

http://semimd.com/blog/2014/04/17/globalfoundries-and-samsung-join-forces-on-14nm-finfets/

http://www.fudzilla.com/home/item/34537-globalfoundries-and-samsung-team-up-for-14nm-finfet

AMD doesn't conduct any research on developing manufacturing processes. All of their research & patents went to GF as part of the deal to spin-off AMD's manufacturing arm to the government of Abu Dhabi which owns the investing arm ATIC which created Globalfoundries in 2009.

 

[patrickjp93]

#1 AMD already taped out 14nm silicone.

http://www.xbitlabs.com/news/other/display/20131017231002_AMD_to_Tape_Out_First_20nm_14nm_FinFET_Chips_Within_Next_Two_Quarters.html

#2 GF & Samsung use the exact same process technology, 14LPE & LPP (which GF licensed from Samsung) & AMD can choose either to make its chips as part of the deal that GF made with Samsung.

http://www.eetimes.com/document.asp?doc_id=1321974

http://semimd.com/blog/2014/04/17/globalfoundries-and-samsung-join-forces-on-14nm-finfets/

http://www.fudzilla.com/home/item/34537-globalfoundries-and-samsung-team-up-for-14nm-finfet

AMD doesn't conduct any research on developing manufacturing processes. All of their research & patents went to GF as part of the deal to spin-off AMD's manufacturing arm to the government of Abu Dhabi which owns the investing arm ATIC which created Globalfoundries in 2009.

 

I'm aware of that. What I'm saying is AMD has barely begun researching how to implement circuits on the 14nm process, the actual quantum mechanical problems that begin to arise with interconnects that small and close together. It's not the manufacturing which is the problem. It's the intra-chip interference which is, which has been the massive problem Intel dealt with up until recently (though Broadwell will be here by Christmas).

 

And taping out is several stages removed from mass production. AMD is 3 quarters behind, possibly more.

[TechFan@ic]

I'm aware of that. What I'm saying is AMD has barely begun researching how to implement circuits on the 14nm process, the actual quantum mechanical problems that begin to arise with interconnects that small and close together. It's not the manufacturing which is the problem. It's the intra-chip interference which is, which has been the massive problem Intel dealt with up until recently (though Broadwell will be here by Christmas).

 

And taping out is several stages removed from mass production. AMD is 3 quarters behind, possibly more.

Once a tape out is achieved product ramp up can begin as long as the fab is ready to begin that ramp.

Tape-out-to-product timeline depends entirely on the capacity of the fab and not on the designer which is AMD in this case.

The issues with Broadwell have nothing to do with the SOC interconnect, defect density is the issue behind the broadwell delays; basically yield. Which means that Intel's fab is fundamentally not ready to mass produce a 14nm product with a sufficient yield to make a profit.

http://www.tweaktown.com/news/33458/intel-delays-broadwell-production-to-2014-cites-defect-density-issue/index.html

Yield improves with time which is why they've delayed production twice already amounting to three quarters.

So again AMD isn't 3 quarters behind I have no idea why you even say that. They're product has taped out, meaning it's ready for production if they choose GF that means Q4 2014 if they choose Samsung that means they already began production, they might've even chosen to make chips at both manufacturers to optimize time to market & cost.

Which GF themselves say is possible & intended, check out the video in my previous post.

[patrickjp93]

Once a tape out is achieved product ramp up can begin as long as the fab is ready to begin that ramp.

Tape-out-to-product timeline depends entirely on the capacity of the fab and not on the designer which is AMD in this case.

The issues with Broadwell have nothing to do with the SOC interconnect, defect density is the issue behind the broadwell delays; basically yield. Which means that Intel's fab is fundamentally not ready to mass produce a 14nm product with a sufficient yield to make a profit. Yield improves with time which is why they've delayed production twice already amounting to three quarters.

So again AMD isn't 3 quarters behind I have no idea why you even say that. They're product has taped out, meaning it's ready for production if they choose GF that means Q4 2014 if they choose Samsung that means they already began production, they might've even chosen to make chips at both manufacturers to optimize time to market & cost.

Which GF themselves say is possible & intended, check out the video in my previous post.

Intel fixed that yield problem back in march. Also, this would not be the first time GF has said they're ready and found out a month later they weren't. 

[GPUXPert]

Intel fixed that yield problem back in march. Also, this would not be the first time GF has said they're ready and found out a month later they weren't.

You can't "fix" defect density you have to improve the yield gradually until it reaches the target required by Intel, TSMC or whomever. Which is a slow process and depending on your target can take anywhere from three months to  15 months and even 2 years+.

Intel saying we have a defect density issue is just a generic statement that means our yield is still bad we can't mass produce yet without destroying our margins.

[patrickjp93]

You can't "fix" defect density you have to improve the yield gradually until it reaches the target required by Intel, TSMC or whomever. Which is a slow process and depending on your target can take anywhere from three months to  15 months and even 2 years+.

Intel saying we have a defect density issue is just a generic statement that means our yield is still bad we can't mass produce yet without destroying our margins.

When you're Intel, problems get fixed quickly or people get fired, and Intel fixed the problems they were having in Q1 back in March,

[GPUXPert]

When you're Intel, problems get fixed quickly or people get fired, and Intel fixed the problems they were having in Q1 back in March,

Well they obviously didn't otherwise they would've began production in March but they've only recently started production a few months ago, an entire year later.

[TERAFLOP]

Well they obviously didn't otherwise they would've began production in March but they've only recently started production a few months ago, an entire year later.

yeah lol 14nm has been a bitch for intel. Which isn't a bad thing necessarily. Gives others more chance to compete.

[patrickjp93]

Well they obviously didn't otherwise they would've began production in March but they've only recently started production a few months ago, an entire year later.

Production began in April, which is still a massive delay. Testing of the Haswell Extreme series began around the same time, which puts it on its schedule to launch in September with Broadwell in December (as Intel has claimed as recently as the past couple weeks).