AMD CPUs, SoC Rumors and Speculations Temp. thread 2

[8350rocks]

Just for clarification, Polaris is made @ TSMC.

Wrong. TSMC doesn't even have a 14nm node. Polaris is made on 14nm @ Globalfoundries (more concretely the 14LPP node)

http://www.amd.com/en-us/press-releases/Pages/amd-demonstrates-2016jan04.aspx
http://www.tomshardware.com/news/amd-zen-polaris-14nm-finfet,31338.html
http://www.globalfoundries.com/foundry-files-blog/foundry-files/2016/01/07/globalfoundries-to-build-new-14nm-gpus-for-amd

I have honestly paid so little attention to the GPUs this cycle, it did not even register that they were putting these on 14nm. Of course, the APUs will end up at GF because the Zen core is there...guess the APUs will be using an iteration of Polaris and not Vega. Makes sense...low power for APUs.

 

[jdwii]

Guys haven't posted anything on AM4 launching?

https://www.engadget.com/2016/09/05/amd-7th-generation-desktop-processors-ship/

Reminds me of when i bought a 200$ sabertooth AM3+ board to upgrade to bulldozer before seeing the benchmarks first, don't be dumb like me guys. Never forget that was 100% my fault same with people who preorder games and complain when they are not satisfied.

 

[jimmysmitty]

Guys haven't posted anything on AM4 launching?

https://www.engadget.com/2016/09/05/amd-7th-generation-desktop-processors-ship/

Reminds me of when i bought a 200$ sabertooth AM3+ board to upgrade to bulldozer before seeing the benchmarks first, don't be dumb like me guys. Never forget that was 100% my fault same with people who preorder games and complain when they are not satisfied.

I think it is because it is only for APUs and to me they are just finally up to par with what Intel has been running for a few years. That's all good and well but they should have at least made new chipsets to support new features for their mainstream CPUs.

I think Zens launch will merit more fanfare.

 

[ComputerSecurityGuy]

I think the arrival of AM4 and Bristol Ridge is imminent though. From the looks of things, vendors are liquidating Carrizo and Kaveri desktops, and a few HP boxes seem to have discretely appeared. Not scientific speculation, but you have to be suspicious...

 

[ComputerSecurityGuy]

Anand now has a very nice piece on Bristol Ridge.
http://www.anandtech.com/show/10705/amd-7th-gen-bristol-ridge-and-am4-analysis-a12-9800-b350-a320-chipset

Clocks look like 3.8/4.2 for the A12-9800. If it has a 20% higher IPC, that puts it at the equivalent of a ~3.1gHz Skylake chip as AMD claimed. Of course, it is a generalization.

Anyway, quite interesting.

 

[juanrga]

Anand now has a very nice piece on Bristol Ridge.
http://www.anandtech.com/show/10705/amd-7th-gen-bristol-ridge-and-am4-analysis-a12-9800-b350-a320-chipset

Clocks look like 3.8/4.2 for the A12-9800. If it has a 20% higher IPC, that puts it at the equivalent of a ~3.1gHz Skylake chip as AMD claimed. Of course, it is a generalization.

Anyway, quite interesting.

20% higher IPC than what?

 

[ComputerSecurityGuy]

20% higher than Excavator, I should say.

 

[Kulasko]

20% higher than Excavator, I should say.

You know that Bristol Ridge IS excavator? Bristol Ridge is an enhanced Carrizo with DDR4 and the new platform.

 

[ComputerSecurityGuy]

Original Excavator.

I hope I don't have to clarify more.

 

[-Fran-]

Original Excavator.

I hope I don't have to clarify more.

I was wondering if they went with "v2" because they ran out of machinery name, haha.

What other big machinery is commonly used? Dump trucks? That would have been a funny name, but very fitting being the last iteration for the uArch xD

Cheers!

 

[manleysteele]

Dragline? Caterpillar? Truck Crane? Backhoe? Scraper? Motor patrol ( also called Road Grader)? Hay Baler? Disc Harrow? Combine Harvester? Steam Shovel?

 

[juanrga]

20% higher than Excavator, I should say.

Original Excavator.

I hope I don't have to clarify more.

The A12-9800 APU has been tested against the A10-7870K APU.

As for single and multi-threaded performance, the 9800 delivers around 6% single core and 15% multi-core enhancement over Kaveri CPUs. Anoth

The Excavator chip has 4.2GHz turbo. The Steamroller turbo is 4.1GHz. The IPC gain over Steamroller is somewhat around 3%.

The multi-core enhancement could be surely related to the extra BW provided by the improved memory controller on Bristol which supports DDR4-2400 instead the DDR3-2133 of Kaveri. In fact the BW increase 2133-->2400 is 13% and adding the IPC gain of 3% gives a total throughput gain of about 16% which agrees with measured 15% in multicore.

No way Bristol IPC is 20% higher than "Original Excavator", because this is just Excavator with minor tweaks.

 

[Vogner16]

Guys haven't posted anything on AM4 launching?

https://www.engadget.com/2016/09/05/amd-7th-generation-desktop-processors-ship/

Reminds me of when i bought a 200$ sabertooth AM3+ board to upgrade to bulldozer before seeing the benchmarks first, don't be dumb like me guys. Never forget that was 100% my fault same with people who preorder games and complain when they are not satisfied.

Reminds me on when I bought a $280 Crosshair V formula AM3+ board to upgrade my x4 945 to the 8150 before benchmarks. I don't regret it one bit. my 8150 was an OC beast and I had her steady at 4.5. put my old 945 to shame in encoding! Never forget that some people were happy with BD.

I still have the original motherboard and am using it with an 8350 at 4.8 to this day. 2011 motherboard, 2012 cpu, all of the clocks one could want. Do not regret early purchase ONE BIT.

EDIT: Having said that the enthusiast boards are not out yet so I cant get one early even if I wanted too.

 

[ComputerSecurityGuy]

Yeah, I flubbed up a bit with some older assumed info. It looks like it's a bit faster than Excavator but also can hit the high clocks typical of Dozer based architectures. It competes with the Skylake i3 line, but with a stronger iGPU.

I would also imagine that AMD can if needed price these chips very attractively, as TSMC's 28nm fab is nearly empty and has high yields at this point.

 

[Vogner16]

http://wccftech.com/amd-x370-chipset-am4-summit-ridge-processors/

x 370 chipset announced 2 x 16 pcie 3 lanes

 

[jimmysmitty]

http://wccftech.com/amd-x370-chipset-am4-summit-ridge-processors/

x 370 chipset announced 2 x 16 pcie 3 lanes

Looks like they are also limiting features now like Intel. Overclocking is limited to the B350 or higher and the only one to do CF/SLI is the X370. It doesn't say the B350 can't but I would assume it would be limited much like a H170 chipset.

Guess we will see when the reviews come out. At least they have PCIe 3 now.

Watch Intel is going to drop PCIe 4 just for laughs. Not that it matters I doubt PCIe 3 will be bottlenecked for another two generations.

 

[Vogner16]

http://wccftech.com/amd-x370-chipset-am4-summit-ridge-processors/

x 370 chipset announced 2 x 16 pcie 3 lanes

Looks like they are also limiting features now like Intel. Overclocking is limited to the B350 or higher and the only one to do CF/SLI is the X370. It doesn't say the B350 can't but I would assume it would be limited much like a H170 chipset.

Guess we will see when the reviews come out. At least they have PCIe 3 now.

Watch Intel is going to drop PCIe 4 just for laughs. Not that it matters I doubt PCIe 3 will be bottlenecked for another two generations.

As a side note to your points. older AM3 boards all "supported" overclocking so long as that board was not blocked via bios from the OEM. OEM's all blocked overclocking regardless. in regards to system builders the 970 and 990 and 990fx were all the mainstream to enthusiast chipsets, and allowed overclocking, CF, SLI, etc... but less CF as you went down the product stack.

780 and 880 also "allowed" CF but nobody did it on those setups as it brings the question of why CF on a motherboard that is designed for entry users with chipset graphics... same argument could be made here. the A320 chipset is for really cheap systems. AIO PC's HTPC's servers etc... nobody would do a CF with these boards anyway. B350 is for non overclocker regular users who want system to work and do it all but don't care to set records. X370 is all anyone will buy on this forum for personal computers.

to say they are reducing functionallty and features like intel is, is true, BUT when you look at the product stack it really was a pointless feature in the past anyway. The only possible loss of functionallty I see is if they remove overclocking on b350 but as I saw an article on the a12 9800 getting overclocked to 4.8 on a b350 board I will assume this is still possible to do. its a simplification of chipset and allows for better motherboards. I can only see this as a positive thing.

two full PCIe 3.0 lanes and the ablity to have a third party will mean high end boards will have another chip for more GPU's and possibly even m.2 along with CF setups. I fully expect the next motherboards on AM4 with the X370 chipset to be amazing.

 

[jdwii]

I'll only say something if Amd pulls the same stunt Intel pulls with SMT. As its basically free and built in the design anyways.

 

[cdrkf]

I'll only say something if Amd pulls the same stunt Intel pulls with SMT. As its basically free and built in the design anyways.

Well I'm not sure I'd call Intel's policy of offering chips without SMT a 'stunt'... yes it's included in the core- however it *does rely on hardware* to duplicate the threads. Is it not possible that disabling HT could allow them to salvage some parts with poor yields?

Even if not- the way I look at it, a lot of software doesn't benefit from more than 4 threads, and HT threads are weaker than full cores. So the i5 makes a lot of sense- as you're basically getting the full chip for a lot less money. 9 / 10 times that works out as better value as you don't suffer a performance loss because of it.

 

[gamerk316]

I'll only say something if Amd pulls the same stunt Intel pulls with SMT. As its basically free and built in the design anyways.

Well I'm not sure I'd call Intel's policy of offering chips without SMT a 'stunt'... yes it's included in the core- however it *does rely on hardware* to duplicate the threads. Is it not possible that disabling HT could allow them to salvage some parts with poor yields?

Even if not- the way I look at it, a lot of software doesn't benefit from more than 4 threads, and HT threads are weaker than full cores. So the i5 makes a lot of sense- as you're basically getting the full chip for a lot less money. 9 / 10 times that works out as better value as you don't suffer a performance loss because of it.

I will point out performance losses due to SMT is downright rare now; the OS is aware of SMT/HTT and tries to handle thread allocation so that you don't have to threads stall.

 

[cdrkf]

I'll only say something if Amd pulls the same stunt Intel pulls with SMT. As its basically free and built in the design anyways.

Well I'm not sure I'd call Intel's policy of offering chips without SMT a 'stunt'... yes it's included in the core- however it *does rely on hardware* to duplicate the threads. Is it not possible that disabling HT could allow them to salvage some parts with poor yields?

Even if not- the way I look at it, a lot of software doesn't benefit from more than 4 threads, and HT threads are weaker than full cores. So the i5 makes a lot of sense- as you're basically getting the full chip for a lot less money. 9 / 10 times that works out as better value as you don't suffer a performance loss because of it.

I will point out performance losses due to SMT is downright rare now; the OS is aware of SMT/HTT and tries to handle thread allocation so that you don't have to threads stall.

I'm not suggesting having SMT slows anything down anymore- what I'm saying is that once you get over 4 threads on a quad core, the additional performance of the additional threads is quite a bit less than that of a full core.

That means that, even in fully multi threaded scenarios with perfect scaling, an i7 with it's 8 threads is not *twice the speed* of an i5 with 4 threads. It usually depends on the type of code being run but in some situations you may get 10% extra performance, in other you might get 50%. You'll never get double though.

So for a product differentiating exercise, the i5 usually represents good value- as you have the full *hardware resources* at your disposal, just limited to 4 threads. Usually the cost reduction is more than the loss of performance. I agree though, at no point would I expect the i5 to be faster than the i7 if the two are clocked the same (at best the i5 can be expected to match the i7).

What I'm more interested though is if it's possible there is a technical yields based reason for dropping SMT on the i5 (i.e. that allows salvaging of otherwise defective dies) or if it is purely a marketing decision?

 

[-Fran-]

I agree with gamerks point and actually can back it up with some data of my own.

An i5 is a better value if you're not planning to do any heavy threading, but the i7 is the winner hands down on any threaded program. And you're taking the glass half empty with your statement: "you have the full *hardware resources* at your disposal". With an i7 you also have those 4cores at your disposal (obviously), but the implication you're making makes it sound like a bad deal when it's not. I do agree the i5 is a better value, but not because of "having 4 cores and not 4+4"; it's price is right where it justifies *not having* the extra horsepower the i7 line does. Intel made a good job there IMO. The extra grunt for the i7 gets justified perfectly in my eyes when you actually need it.

EDIT: The yield question is a very interesting one, but I'm pretty sure no one knows. If you ask me, that is key marketing and strategic information that Intel will try and keep behind closed doors.

Cheers!

 

[Kulasko]

Afaik the dedicated hardware for SMT makes up such a small part of the core (for intel anyways) and therefore has such a low chance of being defective while the core still functioning that the yield question is a clear no for me. They would be better off (yield-wise) to sell tri-cores with SMT.

Sources? No other than my eyes, serveral die shots and a few block diagrams of the Haswell-architecture. I doubt they changed that much with Broadwell/Skylake.

 

[jdwii]

I'll only say something if Amd pulls the same stunt Intel pulls with SMT. As its basically free and built in the design anyways.

Well I'm not sure I'd call Intel's policy of offering chips without SMT a 'stunt'... yes it's included in the core- however it *does rely on hardware* to duplicate the threads. Is it not possible that disabling HT could allow them to salvage some parts with poor yields?

Even if not- the way I look at it, a lot of software doesn't benefit from more than 4 threads, and HT threads are weaker than full cores. So the i5 makes a lot of sense- as you're basically getting the full chip for a lot less money. 9 / 10 times that works out as better value as you don't suffer a performance loss because of it.

I will point out performance losses due to SMT is downright rare now; the OS is aware of SMT/HTT and tries to handle thread allocation so that you don't have to threads stall.

I'm not suggesting having SMT slows anything down anymore- what I'm saying is that once you get over 4 threads on a quad core, the additional performance of the additional threads is quite a bit less than that of a full core.

That means that, even in fully multi threaded scenarios with perfect scaling, an i7 with it's 8 threads is not *twice the speed* of an i5 with 4 threads. It usually depends on the type of code being run but in some situations you may get 10% extra performance, in other you might get 50%. You'll never get double though.

So for a product differentiating exercise, the i5 usually represents good value- as you have the full *hardware resources* at your disposal, just limited to 4 threads. Usually the cost reduction is more than the loss of performance. I agree though, at no point would I expect the i5 to be faster than the i7 if the two are clocked the same (at best the i5 can be expected to match the i7).

What I'm more interested though is if it's possible there is a technical yields based reason for dropping SMT on the i5 (i.e. that allows salvaging of otherwise defective dies) or if it is purely a marketing decision?

Yes but what about the I3 that CPU is way better then a Pentium over SMT. Just look at Dsogaming/Digital foundry with their benchmarks.

 

[juanrga]

Afaik the dedicated hardware for SMT makes up such a small part of the core (for intel anyways) and therefore has such a low chance of being defective while the core still functioning that the yield question is a clear no for me. They would be better off (yield-wise) to sell tri-cores with SMT.

Sources? No other than my eyes, serveral die shots and a few block diagrams of the Haswell-architecture. I doubt they changed that much with Broadwell/Skylake.

You nailed it!

i3: 2C/4T
i5: 3C/6T
i7: 4C/8T

The minimal granularity of the microarchitecture is two cores. Therefore Intel cannot sell tri-cores with SMT and sells instead four-cores without SMT.

i3: 2C/4T
i5: 4C/4T ~ 3C/6T
i7: 4C/8T

Throughput: 3C + SMT ~ 3 * 1.25 = 3.75 ~ 4C