AMD CPU speculation... and expert conjecture

[de5_Roy]

i don't know if anyone has noticed... but ##### is back. this is not a personally aimed remark. it is based on previous experiences.

think of this this like uh.. a badly done horror film (or a sequel of such), right before the ending/last scene, all sorts of creatures start coming out and try to scare the viewers, then the ending comes and after the credit rolls the viewers realize that there will be a sequel.

as the launch gets closer, more and more trolls, shills and clueless fanboys will come out of the woodwork to artificially inflate the hype balloon to overhype proportions. (right before last scene)
beware of false propaganda and baseless fearmongering.
take everything with a pacific rim (jaeger) sized grain of NaCl, even with majority of official statements.
reading between the line helps, up to some extent. but read into too much, you'll drown in misinformation.

but fear not, they will go away right after independent reviews come out (ending), people start buying and posting their real, personal experiences and more in-depth analysis come out (credit rolls). they never stick around to be proven they're wrong.. except the truly blockheaded, shameless ones. they'll be back before the next launch's due date (hint of sequel and repeat) possibly with new forum ids, same schtick. cycle continues.

"Do Not Feed The Trolls (et al): 2013 Edition"

 

[juanrga]

noob2222, good to know you maintain your stupid belief about Kaveri performance and now add your belief that a 200-250W rated processor needs LN2 for cooling. SSShhh! Don't say that to FX-9000 owners.

jdwii, Do you want me to explain you the difference between an average and a maximum value?

 

[juanrga]

http://www-03.ibm.com/press/us/en/pressrelease/41684.wss

I guess we will see if IBM and Nvidia show up with a Power APU or if they get it working across dCPU and dGPU. From what I'm getting out of this is that they want to basically copy HSA, but instead of AMD CPU + AMD GPU, it'd be IBM CPU + Nvidia GPU. Correct me if I'm wrong, but it looks to me like IBM and Nvidia want to go after exactly what I'm describing (as well as HPC) in the same exact way I'm talking about it. So at this point the only one really left out of this type of HSA ecosystem is Intel, who at best as Xeon which is a system on a board (with full Linux OS).

Goals and target of the OpenPower consortium are different to HSA. They are more interested in high-performance systems such as supercomputers and big servers, whereas the HSA foundation target phones and tablets as well.

The original partnership between IBM and Nvidia only considers CPU+dGPU supercomputers/servers and related software (e.g. IBM proprietary datacenter applications will be CUDA accelerated). The new Power8 CPU includes the new CAPI (Coherence Attach Processor Interface) expansion bus to integrate external co-processors such as Nvidia GPUs.

IBM and Nvidia could develop an APU. However, Nvidia is developing its own CPU cores for its all-Nvidia ultra-high-performance (20TFLOP) APU.

Intel is not out... at least not still. They have developed their own approach, which they name "neo-heterogeneity". Instead GPUs they use Xeon Phis. In fact, Intel will be the first to move to socket with a version of the next Xeon Phi

which will offer 3TFLOP of performance. I.e. about double today performance.

 

[gamerk316]

APU solves the problem of HSA having to have high bandwidth communication between CPU and GPU, but it's only short term because it's not viable on higher end platforms. Whether you think AMD wants to solve the problem of fast interconnects between GPUs and CPUs depends on if you think AMD is going to be content with mobile, ARM, and low-end gaming PCs or if they want to be viable in server, workstation, and high end gaming PCs again.

But APU introduces the problem of needing significant amounts of slow system RAM. dGPU's just stick 2GB or so on the cards to get around the PCI latency as much as they can, but APU's simply are NEVER going to have that amount. Likewise, using GDDR as the main system RAM is simply too expensive to consider doing. So it all averages out in the end: dGPU's have worse latency, APU's have worse bandwidth.

 

[gamerk316]

Hey guys.

Sorry for jumping in again, but I thought this would be a good topic to discuss.

Before we find a true victor, lets talk overclocking.

I'll give you an example. Right now I have my i5-3570k @ 4.7 Ghz on 1.750 v-core (which I dont think is too bad).

With this overclock, my scores on the benchies have been exceeding those with AMD's "best processors."

Even at 4.7, this processor seems to have no giving up. With 4 instances of prime 95 running, the highest it has gotten with my h100i on this is 72 ºC, and on the previous 4.5 Ghz and 1.105 v-core, it maxed at 68ºC. With the barrier at (~110ºC I think it is), there is totally room to go since I'm getting a 3770k soon.

So those people that usually max out at 4.8 Ghz, and then the AMD people who say that their 8350 can go to 5.5-6 GHz, think again. Some of those wafers out there are the chosen ones. ^_^

If you can keep the 8350fx cool and give it enough voltage threw the board i'm sure it will go that high. You said 1.750V??? Really that high? Pretty sure a 8350fx can get to 4.8Ghz with just 1.5V.

But here's the thing: Intel has more OC headroom then AMD, and has the higher base IPC, so Intel gains more then AMD overall when OCing to max frequency. If you get into comparing Max OC results, AMD looks even worse then it normally would.

Piledriver v Haswell and even Ivy, AMD DESTROYS Intel in OC headroom.. A H60 can take a 4670K to 4.3, a 8350 4.6... Hell, an H100 or D14 can get em up to 5 pretty easily..

Not really, and you just proved my point:

4670k: Base clock: 3.4GHz. Average OC: 4.3GHz. Average OC Headroom: ~900MHz
FX-8350: Base Clock: 4.0GHz. Average OC: 4.8GHz. Average OC Headroom: ~800MHz

So even at the lower clock, Intel has more room to OC. Which makes sense, when you consider the voltages necessary to drive a high clockspeed.

 

[gamerk316]

has mantle been released yet for battlefield 4 ?

Nope. Hell, DICE has stopped all other work until the issues with BF4 are fixed, so that upgrade will likely be delayed for a bit.

 

[tracker45]

What iGPU will Excavator have in it ? next year ? Could they fit a 7850 on a cpu ?

They already have put a 7850+ on an APU. It's the chip in the PS4.

really, that seems very powerful for a ps4.

 

[gamerk316]

What iGPU will Excavator have in it ? next year ? Could they fit a 7850 on a cpu ?

They already have put a 7850+ on an APU. It's the chip in the PS4.

really, that seems very powerful for a ps4.

Remember it has to last 10 years. Should fare better then the X2900XT/7800GTX variants used in the last generation of consoles at least. CPU wise though, the current consoles are NOT going to age well, in my opinion.

 

[Cazalan]

Remember it has to last 10 years. Should fare better then the X2900XT/7800GTX variants used in the last generation of consoles at least. CPU wise though, the current consoles are NOT going to age well, in my opinion.

A pipe dream if they think that. Tech is still advancing quite rapidly. The 3D memories (HBM/HMC) coming out will greatly increase the bandwidth available to APUs and it will mature nicely to line up with 10nm or so process within 4 years. A SteamBox in 4 years with a 2k res Oculus Rift will be quite a beast. There was a hiccup between 32nm and 28nm production but things are accelerating for 20/14nm.

 

[gamerk316]

^^^7850 ? I keep hearing it but the only evidence is the cu count and speculated just under 2 tflops bandwidth, If true that would make jaguar igfx 2x faster than richland gpu. Even kaveri is not at 7850 level yet, why do you believe jaguar is?

Custom chip. Performance is guesstimated to be a 7850 with the 7890's memory size, based on speed/CU's/amount of RAM.

 

[juanrga]

But APU introduces the problem of needing significant amounts of slow system RAM. dGPU's just stick 2GB or so on the cards to get around the PCI latency as much as they can, but APU's simply are NEVER going to have that amount. Likewise, using GDDR as the main system RAM is simply too expensive to consider doing. So it all averages out in the end: dGPU's have worse latency, APU's have worse bandwidth.

Never say NEVER...

The next 200W Intel APU includes 8-16GB of stacked DRAM on die with a bandwidth >500GB/s. For the sake of comparison the Nvidia Titan uses GDDR5 to achieve only 288GB/s. The Intel APU is complemented with a six-channel DDR4-2400 expansible up to 384GB. This gives more than 3x the memory bandwidth of Richland APU, for instance.

The 200-250W AMD APU includes 32-64GB of stacked DRAM on die and it is complemented with a four-channel of unspecified DDR memory. Bandwidths are not disclosed but they have to be superior to Intel design, because the iGPU has about 3x more performance than Intel.

The 300W Nvidia APU includes 256GB of unified stacked DRAM on die with a bandwidth of 1400-1600GB/s.

 

[-Fran-]

http://www.tomshardware.com/news/google-server-chips-intel-arm,25409.html

I'll go ARM-off-topic on this one, because I think it is interesting for various reasons I thought today.

When there was this "RISC vs CISC/X86" battle back in the 60s and 70s, I think I know why CISC might have won the battle (among other reasons). At the time, compiling code and testing it was expensive. And not 1-hour expensive. Sometimes 1 or 2 days expensive for relative "simple" code. Point here is, when you have an ISA that can do more complex work in tandem using less code to do so, leaves less code (implying less time) to be debugged as long as you stick to the ISA. So, this time around, the software side (in particular, compilers) can produce VERY good code to match the theoretical performance for a CPU in acceptable time frames, meaning in Googles case, choosing ARM over X86 makes a LOT of sense (other than X86 being Intel locked, lol) just because they know they'll be able to squeeze every last bit of performance out of an ARM uArch, making its perf/watt very good. Specially since they can go nuts creating a custom uArch using the ARMv8 ISA.

Whenever Google steps into a market, things shake. Let's see how Intel takes this one.

Cheers!

 

[juanrga]

^^^7850 ? I keep hearing it but the only evidence is the cu count and speculated just under 2 tflops bandwidth, If true that would make jaguar igfx 2x faster than richland gpu. Even kaveri is not at 7850 level yet, why do you believe jaguar is?

Nope. The PS4 iGPU has more CUs, bandwidth, and FLOP than a 7850. Moreover, the PS4 iGPU is a custom design, with several improvements over the ordinary Radeon 7000/8000 line. For instance, each GCN core in the PS4 has 32x more sources for compute commands and there is a second bus to the iGPU that allows it to read directly from system memory or write directly to system memory, bypassing its own L1 and L2 caches.

 

[juanrga]

Sweclockers has just confirmed the leaked roadmap that I posted in my twitter account

https://twitter.com/juanrga/status/406390600804466689

They also got access to the private presentation with the confidential 2015 desktop roadmap

http://www.sweclockers.com/nyhet/17995-amd-forbereder-kabini-for-stationara-datorer-med-sockel-fs1b

 

[juanrga]

http://www.tomshardware.com/news/google-server-chips-intel-arm,25409.html

I'll go ARM-off-topic on this one, because I think it is interesting for various reasons I thought today.

When there was this "RISC vs CISC/X86" battle back in the 60s and 70s, I think I know why CISC might have won the battle (among other reasons). At the time, compiling code and testing it was expensive. And not 1-hour expensive. Sometimes 1 or 2 days expensive for relative "simple" code. Point here is, when you have an ISA that can do more complex work in tandem using less code to do so, leaves less code (implying less time) to be debugged as long as you stick to the ISA. So, this time around, the software side (in particular, compilers) can produce VERY good code to match the theoretical performance for a CPU in acceptable time frames, meaning in Googles case, choosing ARM over X86 makes a LOT of sense (other than X86 being Intel locked, lol) just because they know they'll be able to squeeze every last bit of performance out of an ARM uArch, making its perf/watt very good. Specially since they can go nuts creating a custom uArch using the ARMv8 ISA.

Whenever Google steps into a market, things shake. Let's see how Intel takes this one.

Cheers!

There are also rumors Google is preparing a gaming console based in ARM.

Regarding reasons to choose ARM over x86, the role of compilers was stated many pages ago in this same thread. I repeat:

For several years, makers of high-end computing platforms have had no choice about instruction-set architecture. The only option was the x86 instruction set with variable-length instructions, a small register set, and other features that interfered with modern compiler optimizations, required a larger area for instruction decoding, and substantially reduced energy efficiency.

Denver provides a choice. System builders can now choose a high-performance processor based on a RISC instruction set with modern features such as fixed-width instructions, predication, and a large general register file. These features enable advanced compiler techniques and simplify implementation, ultimately leading to higher performance and a more energy-efficient processor.

Do you remember the rumor AMD was preparing an ARM replacement for tablets? It was confirmed a pair of days ago when I tweeted this BSN Exclusive:

http://www.brightsideofnews.com/news/2013/12/9/amd-working-on-arm-socs-for-consumer-products.aspx

 

[tracker45]

Why can't intel make a good APU ?

 

[szatkus]

Why can't intel make a good APU ?

Lack of expierience. ATI/AMD creates highend GPUs for almost 20 years. Intel just few years ago decided to design non-crappy graphics.

 

[jdwii]

noob2222, good to know you maintain your stupid belief about Kaveri performance and now add your belief that a 200-250W rated processor needs LN2 for cooling. SSShhh! Don't say that to FX-9000 owners.

jdwii, Do you want me to explain you the difference between an average and a maximum value?

Previously on your own website you would just add 30%(magic i guess) to every benchmark saying look. Instead of talk about averages(real world)

 

[griptwister]

Anyone see this yet? :lol: It's just comedy. I personally am a AMD fan, so don't take this as offence.

http://www.youtube.com/watch?v=EIPggCgYK38&list=UUNJP0oF6k62xA_qhCLfwI-Q#t=85

 

[The Q6660 Inside]

Piledriver v Haswell and even Ivy, AMD DESTROYS Intel in OC headroom.. A H60 can take a 4670K to 4.3, a 8350 4.6... Hell, an H100 or D14 can get em up to 5 pretty easily..

Not really... What Intel is doing is making bad calls on OTHER things but the way they're shrinking the transistors. Their BULK is state of the art IMO and GF is just getting into their footsteps.

This is some proof:

I don't recall having 1.4v at 4.4Ghz, but I haven't had the time to fine tune the OC (fat lies again; just laziness).

Anyway, point is credits due where credits owed: Intel's BULK is top of the crop. I'm really skeptical GF in BULK will reach Intel's 32nm in 28nm even. Much less 22nm. I'd even say TSMC's 28nm will be better than GFs (if I got all the BULK thing for Kaveri right, lol).

Cheers!

EDIT: Right... I'm on Air.

I was referring to Ivy and Haswell, which was a dog compared to SB, you could do 5.1 on an H80 with a 2500K for crying out loud.. @Gamerk I seem to be corroded with the idea of "it runs on Turbo all the time".

 

[jdwii]

I was just talking about Intel's latest chips vs Amd's. Haswell vs FX. Typically Amd gets a 10% higher clocks but yes gamer is right Intel still has more overclocking headroom since the 8350fx has such a high clock speed out of the box. To bad it didn't have the performance per clock performance and the high end clock.

 

[Embra]

Anyone see this yet? :lol: It's just comedy. I personally am a AMD fan, so don't take this as offence.

http://www.youtube.com/watch?v=EIPggCgYK38&list=UUNJP0oF6k62xA_qhCLfwI-Q#t=85

Ha Ha... thank you for sharing that, funny!

 

[griptwister]

ATTENTION: On another note about Comedy... I got a email from AMD saying, "Hey AMD fan, Be sure to tune into Late Night with Jimmy Fallon tonight (Friday Dec 13th) for a special surprise from AMD!"

I look forward to hearing what ever they announce.

@The Q6660 Inside: I will forever remember Sandy-Bridge i5 as the cream of the crop as far as Intel CPUs go.

@Embra: You're welcome good sir!

 

[de5_Roy]

TSMC 16nm FinFET in risk production
http://www.fudzilla.com/home/item/33402-tsmc-16nm-finfet-in-risk-production

 

[noob2222]

@juan

Is that cosmology ES kaveri?

Spoiler

If you believe it actually is kaveri then you are the one who believes kaveri is 37% slower than BD

Fx-9000 cpus ship with a water cooler by design and is 32nm. What node is that 250w future apu? Might want to look up the term "thermal density" and "thermal dissipation" because we are at the upper limits for copper and plain old "air".