esrever :
They want in on the valve steamOS push
I would go beyond: Nvidia want to be inside the SteamBox.
8350rocks :
palladin9479 :
In order to scale up you need to start adding dedicated things to the die that aren't directly related to processing power. Cache / instruction schedulers / branch predictors / ect. Those things consume power and space and thus your cost per performance starts to erode rapidly. ARM doesn't scale up nearly as well as it's proponents say. It's a horizontal architecture, you get high performance by connection dozens of ARM processors to each other and executing a ton of distributed code. Try to close an ARM CPU very high and you slam into the same problems everyone else ran into two decades ago.
As for RISC CPU's being high powered, their have been high performance RISC CPU's for a very long time. RISC is actually much easier to make *fast* due to it's incredibly predictable instruction execution time. SUN even tried making a "Desktop" SPARC machine (Sunblade 100/150), didn't work out too well. Apple was using PPC which worked out better then the Sunblade's but even they switched over to x86 eventually. x86 is just cheaper to design around and already has a huge ecosystem to draw revenue from.
So the entire ARM vs x86 is just some wishful thinking. ARM is amazing at low power (electrical usage) devices that do just enough to be useful. It doesn't scale up nearly as well as the other design's but it's ability to scale down is unmatched.
+1
I have been trying to tell them ARM doesn't scale upwardly nearly as well as they think...it just wasn't sinking in, I guess. It scales downward ridiculously well, however, to scale it upward...you're really going into larger and larger clusters of ARM CPUs to get the performance. Otherwise the PPW angle will vanish, and the hardware will have to have more and more added complexities which x86 has been integrating for 10 years now.
ARM is on the precipice of 64 bit processing, but they're not anywhere near the complexity and capability of the x86-64 ISA at this point. It doesn't mean ARM isn't viable, by any stretch, for what it's intended use is; however, it doesn't make sense from a DT OS perspective, or even a high performance server angle either.
ARM will not defeat x86-64 anytime soon. AMD will get into that market to try and grab a toe hold in ARM micro servers because of the margins. Other than that, ARM will remain
THE gold standard for mobile tablets, phones, etc. It will not replace HEDT anytime soon either, as that segment of x86 DTs is clearly on the rise...by a fair margin I might add.
ARM is an architecture optimized for
efficiency. x86 (either AMD or Intel) can provide the same performance (at higher power consumption) or the same power consumption (at lower performance levels) but not both at once.
ARM has a several generations history of adding performance
whereas improving efficiency over the previous generations. ARM has made this again with the new A57 core, which is more faster than the A15 but maintaining the same power consumption. This history destroys the 'argument' that ARM doesn't scale well upwards.
History also shows that x86 doesn't scale well downwards. That is because x86 was not designed for efficiency but for raw performance. x86 was designed mainly for desktops, where power consumption was irrelevant. It was teawked for laptops/servers where efficiency is a factor, but is very far from being optimal in efficiency. Intel has tried to scale downwards to tablet/phones but has systematically failed. The interesting part is that now efficiency start to play a mayor role in fields as HPC. It is easy to see take current fastest supercomputer (x86 based) and scale it upwards 1000x. It is impossible to provide 1000x the current energies. That is why ARM supercomputers are in the target. The goal of the Mont Blanc project is to achieve a supercomputer that was 1000x faster than the fastest x86 today, but consuming less energy!
About raw performance, the ARM 32-bit A15 provides about one half of the IPC of an i7 SB. The new A57 is up to 40-50% faster than the A15, but providing higher efficiency
This is only for older 32 bit code and only 2n nm. The A57 is still faster (> 50%) running on native 64 bit mode. Also A57 cores at 16nm FinFET are going to appear.
Custom ARM64 cores will surpass the A57 raw performance whereas maintaining efficiency. It seems that Apple new core already obtains SB i7 IPC level, but where I expect real breakthroughs is from Nvidia and AMD custom cores.
Before someone say me that SB is old, I will add that Haswell offers only about a 10% more IPC than Sandy.
Manufacturing a high-performance ARM core for a supercomputer is more expensive than manufacturing a basic phone core, but continues being less expensive than manufacturing a high-performance x86 core.
Of course this is all for a single core. The new Seattle CPU already offers essentially the performance of some of fastest x86 chips (Xeons) but only a percentage of their power consumption.
ARM has defeated x86 in phones/tablets for years, despite Intel has fab advantage and spend billions on research. Now ARM has started to defeat x86 in servers.
gamerk316 :
Historically, RISC does not scale as well per clock as X86, and RISC architectures (notably PPC and MIPS) have had issues getting clocks above 3GHz. Using those two arches as a baseline, you can guesstimate performance for ARM.
I think you missed important dates of the history. IBM already released 5Ghz PPC chips years before AMD announced its 5GHz FX (base freq. is 4.7GHz). IBM is already selling 5.5GHz PPC chips. Moreover, you cannot compare PPC/MIPS to ARM. ARM64 is a modern optimized uArch unlike MIPS.
Facebook
Twitter
Instagram