AMD CPU speculation... and expert conjecture

[Cazalan]

The final point to note, when this gen came out I remember a comment (I think from AMD) that the approach with these consoles was going to be smaller updates more often. I can see in a couple of years time with a process node shrink + some newer more efficient cores there will be a refresh. Also as the architecture is unlikely to change, they may well keep supporting the current consoles along side the newer ones.

The 20nm shrink has already taped out and the 14/16nm version is in the works (according to linkedin profiles). What remains to be seen is if they'll enable different software profiles for these so they can turn up the details or if its just for cost reduction.

 

[palladin9479]

Current consoles are far better then last gen by a large margin. The problem is people were expecting PC level graphics and with sub 200w total system requirements your not getting that. Consoles need to be cheap and low power, current PC technology favors power hungry CPUs and GPUs. Just think 400w PSU's used to be for "gaming PCs" about a decade ago and the coolers for both used to be much smaller. We've had this constant march towards more powerful devices, but they are also far more power hungry and as a result of EU's politics you now have limited console capabilities.

Also I posted in that other thread about VISC, it's not what many of you are interpreting it as. They are using heavy marketing language to discuss what is basically the same thing as Intels / AMD's front end instruction decoder / scheduler. They merely developed it separately from any specific uArch and so it can be purchased by other companies for implementation. It won't run x86 as it requires the code to be compiled into their own VISC binary language first.

 

[-Fran-]

That explains a lot about the claims.

So, if you need to re-compile, this will be like ARM with Android/iOS; not much of an improvement at first glance. I wonder if this will be another "Itanic".

Also, something I kind of forgot: to decode/interpret X86, you have to ask Intel first and AMD second.

Cheers!

 

[cdrkf]

That explains a lot about the claims.

So, if you need to re-compile, this will be like ARM with Android/iOS; not much of an improvement at first glance. I wonder if this will be another "Itanic".

Also, something I kind of forgot: to decode/interpret X86*, you have to ask Intel first and AMD second.

Cheers!

*Actually you need to ask Intel to decode X86, and AMD to decode X86-64

 

[Embra]

Pandora Machine: A Steam Box Made In Italy [Intel/AMD inside]:
http://www.tomshardware.com/reviews/pandora-steam-machine-valve,3962.html

 

[logainofhades]

Leave it to the EU to screw things up for people.

 

[blackkstar]

There is some major disconnect with what x86 can do as well. As others have mentioned, these games are all compiled for old CPUs on PC. You're lucky to get something like SSE3 in a lot of situations. Skyrim didn't even use SSE at all, it used raw x87.

Kabini/Jaguar has the best instruction set support out of any of the small cores.
http://www.cpu-world.com/Cores/Kabini.html

Compare with Bay Trail:
http://www.cpu-world.com/CPUs/Celeron/Intel-Celeron%20N2805.html

But Bay Trail competes with updated AMD cat core. The Atom that came out when Kabini was available was this:
http://www.cpu-world.com/CPUs/Atom/Intel-Atom%20Z2520.html

Intel solution is missing:

SSE4a, SSE 4.1, SSE 4.2, AES, AVX, BMI1, F16C, amd64. That's a lot of extra performance laying on the table when you can use those instructions. And you don't normally have access to those in Windows.

It is a bummer Kabini has no FMA, which is a huge performance boost. But it has a lot of good stuff besides that. I run Gentoo on my A4-5000 laptop and on my Athlon 5150 router I just built. They are completely different behaving chips in Gentoo and Windows. Everytime I have to use Windows I groan because it feels like I downgraded my CPU. I realize this is anecdotal evidence but I lived off that notebook for months before. It is a big difference.

 

[gamerk316]

There is some major disconnect with what x86 can do as well. As others have mentioned, these games are all compiled for old CPUs on PC. You're lucky to get something like SSE3 in a lot of situations. Skyrim didn't even use SSE at all, it used raw x87.

Kabini/Jaguar has the best instruction set support out of any of the small cores.
http://www.cpu-world.com/Cores/Kabini.html

Compare with Bay Trail:
http://www.cpu-world.com/CPUs/Celeron/Intel-Celeron%20N2805.html

But Bay Trail competes with updated AMD cat core. The Atom that came out when Kabini was available was this:
http://www.cpu-world.com/CPUs/Atom/Intel-Atom%20Z2520.html

Intel solution is missing:

SSE4a, SSE 4.1, SSE 4.2, AES, AVX, BMI1, F16C, amd64. That's a lot of extra performance laying on the table when you can use those instructions. And you don't normally have access to those in Windows.

It is a bummer Kabini has no FMA, which is a huge performance boost. But it has a lot of good stuff besides that. I run Gentoo on my A4-5000 laptop and on my Athlon 5150 router I just built. They are completely different behaving chips in Gentoo and Windows. Everytime I have to use Windows I groan because it feels like I downgraded my CPU. I realize this is anecdotal evidence but I lived off that notebook for months before. It is a big difference.

What you forget to mention is that 99% of all tasks see major diminishing returns after the SSE2 code path.

 

[8350rocks]

I would say SSE4 is about where diminishing returns kicks in full force...actually. Especially considering games tend to run instructions aimed more at FMAC type operations.

 

[gamerk316]

I would say SSE4 is about where diminishing returns kicks in full force...actually. Especially considering games tend to run instructions aimed more at FMAC type operations.

As of the last Compiler benchmarks I've seen, SSE4 actually reduced performance over SSE3. That being said, that was right after SSE4 hit, so possible compiler fixes may have changed that.

Point being, the results showed HUGE jumps going to SSE2, but very small (<5%) improvements above that. Only very specific workloads really benefit from these new instructions.

 

[szatkus]

Who cares about SSE4?

http://forums.evga.com/FindPost/870670

http://forums.pcsx2.net/Thread-SSE4-vs-AVX

 

[gamerk316]

Who cares about SSE4?

http://forums.evga.com/FindPost/870670

http://forums.pcsx2.net/Thread-SSE4-vs-AVX

Specific workloads. Not all tasks benefit.

 

[anxiousinfusion]

2.6Ghz turbo sound about right for Carrizo-L?

http://www.chw.net/2014/10/primeros-benchmarks-de-los-futuros-soc-amd-a-8000-series-carrizo-l/

 

[szatkus]

2.6Ghz turbo sound about right for Carrizo-L?

http://www.chw.net/2014/10/primeros-benchmarks-de-los-futuros-soc-amd-a-8000-series-carrizo-l/

It's an engineering sample. Clocks could be different in finished product.

 

[-Fran-]

2.6Ghz turbo sound about right for Carrizo-L?

http://www.chw.net/2014/10/primeros-benchmarks-de-los-futuros-soc-amd-a-8000-series-carrizo-l/

It's an engineering sample. Clocks could be different in finished product.

The point in that "leak" or whatever it is, is that Kaveri's equivalent is getting the same numbers (GFLOPs) at 3.2Ghz and this engy sample is at 2.6Ghz. In terms of Perf/Watt its very good.

Cheers!

 

[gamerk316]

We have our first THz chip, thanks to good old Federal government R&D:

http://www.extremetech.com/extreme/193343-darpa-creates-first-1thz-computer-chip-earns-guinness-world-record

 

[Reepca]

Why can't we ever have quality English leaks :'{

On the other hand, why can't Google Translate make its translations not sound like they were written by a 6-year-old?

On the other other hand... at least it got all the spelling right. More than most text on the internet can claim.

Now, me being a filthy casual and all, I'm curious - is a decrease in clock speed directly proportional to a decrease in power consumption? The closest abstraction I've heard is "higher voltage enables higher clock speed... and higher temperature" (from reading that "overclocking basics" stuff). So I assume that a lower clock speed would at least cause a shift in power consumption in the same direction, if not of the same proportion.

But hey, regardless, if the same amount of computing is being done at 81% clock speed, doesn't that mean that IPC is up by 23%? Looks promising.

 

[cdrkf]

Why can't we ever have quality English leaks :'{

On the other hand, why can't Google Translate make its translations not sound like they were written by a 6-year-old?

On the other other hand... at least it got all the spelling right. More than most text on the internet can claim.

Now, me being a filthy casual and all, I'm curious - is a decrease in clock speed directly proportional to a decrease in power consumption? The closest abstraction I've heard is "higher voltage enables higher clock speed... and higher temperature" (from reading that "overclocking basics" stuff). So I assume that a lower clock speed would at least cause a shift in power consumption in the same direction, if not of the same proportion.

But hey, regardless, if the same amount of computing is being done at 81% clock speed, doesn't that mean that IPC is up by 23%? Looks promising.

The relationship between clock speed and power consumption *isn't linear*. Most designs have a 'sweet spot', and going above this is where the consumption starts to sky-rocket, a good example is to compare the Kaveri A10 7800 and A10 7850, the former can run fully unlocked at about 3.2 (?) ghz at 45w, the latter hits 95w at 3.9 ghz.... so double the power for not much more actual performance.

 

[Cazalan]

It's the voltage that causes it to be non-linear. Stability at higher frequencies requires higher voltages.

Sisoft includes AVX2 and GPGPU so certainly Carrizo can get a speedup over Kaveri with all things included. If the GPU included the compression feature of Tonga that will help too.

 

[jdwii]

2.6Ghz turbo sound about right for Carrizo-L?

http://www.chw.net/2014/10/primeros-benchmarks-de-los-futuros-soc-amd-a-8000-series-carrizo-l/

Wow if true its not looking good. If it only has a turbo of 2.6Ghz vs 3.2Ghz for kaveri. 23% less clock speed with what 10% at best improvements in IPC. But it is the A8 to and it comes with 512 GCN cores. However it could just be a leaked sample.

 

[esrever]

2.6Ghz turbo sound about right for Carrizo-L?

http://www.chw.net/2014/10/primeros-benchmarks-de-los-futuros-soc-amd-a-8000-series-carrizo-l/

Wow if true its not looking good. If it only has a turbo of 2.6Ghz vs 3.2Ghz for kaveri. 23% less clock speed with what 10% at best improvements in IPC. But it is the A8 to and it comes with 512 GCN cores. However it could just be a leaked sample.

The L means its low power. It will be a 15w part at most.

 

[8350rocks]

2.6Ghz turbo sound about right for Carrizo-L?

http://www.chw.net/2014/10/primeros-benchmarks-de-los-futuros-soc-amd-a-8000-series-carrizo-l/

Wow if true its not looking good. If it only has a turbo of 2.6Ghz vs 3.2Ghz for kaveri. 23% less clock speed with what 10% at best improvements in IPC. But it is the A8 to and it comes with 512 GCN cores. However it could just be a leaked sample.

The L means its low power. It will be a 15w part at most.

This.

That sample is an ES for ULP stuff. Now, Tonga's compression is supposed to be an integrated feature across the GCN family moving forward since the introduction of that chip...

That should actually solve a LOT of the bandwidth issues with RAM for the flagship SKUs with so many cores...

 

[Reepca]

We're already looking at Carrizo-mobile engineering samples and I STILL can't find a laptop with an fx-7600p in it (by the way, if any of you sight one... let me know! :'}). Closest I've seen are rumors of a pointless MSI system with a non-dual-graphics-able discrete card... why do they even make those? That and an fx-7500 model "exclusively at best buy"... but no reasonably priced/made fx-7600p.

Part of me wonders if I should just wait for Carrizo before I try getting a laptop with an APU in it.

 

[Cazalan]

HP is selling a new Hexa-core laptop with AMD APU.

 

[Reepca]

"HP Hexa-Core is equipped with six powerful cores for lightning-fast processing, in addition to four CPU and two graphic cores to power all your daily tasks and entertainment".

Erm... sometimes I wonder if the marketing department hasn't got a clue what they're talking about, or they know something everyone else doesn't. Maybe it's just a classic foreign-sourced marketing department with imperfect English.

If they DO know what they're talking about, that sentence implies that there are 6 *something* cores in addition to 2 GPU cores and 4 CPU cores, for a total of 12. Frankly I'm more led to believe that they're using really cheesy marketing terms (calling what must be one of the lowest-end APUs, to have 4 cpu cores and 2 GPU cores, a "hexa core" feels... wrong) and using poor English while they're at it.

I'm not even sure it's an AMD APU - they never once mention AMD, and only call it an "HP hexa-core processor". I would assume that if it's got GPU and CPU cores, it's SOME kind of APU, but that picture asks way more questions than it answers.

I typed in one of the "WebID"s into google, which took me to a site called futureshop, where it told me the (sarcasm) immensely useful information of "CPU type: HP".

EDIT: apparently the guys over at anandtech identified the processor as an a8-6410... I dunno, I feel like someone has grounds to sue for false advertising. Especially considering the drastic disconnect between the quality of the processor and the price of the laptop (you can get a10-5750m's for that price, on good days).