Intel in the GHz Game Again - Skulltrail Hits 5 GHz

An anonymous reader writes "Intel's Skulltrail dual-socket enthusiast platform has been making the rounds on the web for half a year or so, but we haven't seen many details yet. TG Daily got a close look at an almost complete prototype, which surely sounds almost like a production ready version, judging from the article. Everything that TG Daily describes sounds like Skulltrail PCs will be very limited in availability and insanely expensive. Intel also has said it has developed 'special' Xeon processors with desktop processor attributes just for Skulltrail. These chips are currently running at a stable 5 GHz."

Comments

Traslation

[king-manic]

These chips are currently running at a stable 5 GHz. A practical translation:

It will be 20% faster, 200% hotter, needs a 300% nosier fan, consumes 500% as much power.

Insanely expensive...

[Anonymous Coward]

And will be obsolete in a year. Honestly, who spends thousands of dollars every year for the most advanced stuff? Even if you did have a Skulltrail, the rest of you system would bottleneck it. 3 8800GTX's would be the bottleneck, 8GB's of the fastest DDR3 ram would bottleneck, and your harddrive would bottleneck too. The only thing Skulltrail gives you is bragging rights.

Hertz by themselves are useless

[porkThreeWays]

Measuring computer performance in Hz is like buying a car based on red line RPMs. It only tells you one component that is meaningless by itself. Just like a car needs torque to give rpm's context, processors need how many instructions can be completed per cycle to be compared to the frequency. I've lost faith in the MHz race and generally look at benchmarks closest to the intended purpose of the processor.

MHz wars are over

[geekoid]

Please, it's all about cores.

Look at the history of processors speed. We've been pretty flat, and will stay that way in all practical manner for a while.
Before someone throws the quote like they are smart, Moore's law refers to transistor not speed.

1) Faster chips require better fabs. Fabs are having difficulty producing better platters with a few enough flaws to produce mass quantities. Strides are being made, but know massive breakthroughs.

2) Multiples cores and real parallel processing development is just starting to become expected knowledge for the average application developer. Lets be honest, a lot of developers don't bother to understand multi-threading and avoid it like a plague. Fortunately there are some IDEs that make it easier for developers.

Re:MHz wars are over

[AuMatar]

Cores only help so much- if your problem is not paralelizable, or if it is only minimally so, a billion cores won't help. A word processor is not going to work any faster on a 1000 core machine than on a 1 core machine. Video games might see a small speed up from a multicore, but not that much of one- it doesn't break down into equally weighted threads. For the vast majority of users, 2 cores aren't even really utalized (email and web browsing doesn't use 2 cores). I doubt any home user will see much improvement beyond 2 cores, and absolutely none after 4 even for hardcore multitaskers. Business and scientific apps will see some beyond that, but memory tends to be the bottleneck there- we'd be better off increasing memory bandwidth and latency than clock speed.

Re:MHz wars are over

[TheRaven64]

We have had about 40 years of practice getting one processing unit to pretend to be n, and we're pretty good at it now. We have no good ways (even in theory) of getting n processing units to pretend to be one in the general case. If you have a 5GHz core then you can run two processes on it happily with only a small amount of overhead. If you have two 2.5GHz cores and only one process, you will end up running that process at half of the theoretical speed of your CPU.

Fewer faster cores will always be more flexible than more slower ones. The reason we go with more slower ones is that slower cores use less power (power scales much worse than linearly with speed, so two 1GHz cores will use a lot less power than one 2GHz one). Some workloads are intrinsically parallel (e.g. web serving) and so having lots of cores using less power is a big win. Others are not and so extra cores are just a waste (although you can often consolidate multiple serial tasks onto one machine with lots of cores).

Re:MHz wars are over

[TheRaven64]

You are missing the point. Any problem that can be solved on a parallel machine can be solved on a serial machine of the same computational power in the same time. The converse, that any problem that can be solved on a serial machine can be solved on a parallel machine of the same processing power, is not true. At the abstract level, any nondeterministic finite automaton can be reduced trivially to a deterministic equivalent, but an arbitrary degree of nondeterminism can not be trivially added to a DFA.

If you can solve a problem on an n-core mGHz machine, you can solve it on a n/2-core, 2mGHz machine in (roughly) the same amount of time. Whether you can solve it on a 2n-core, m/2GHz machine in anything like the same amount of time depends heavily on the problem (and also the value of n - see Amdahl's law).

It doesn't matter how much more practice we've had solving problems on parallel systems, because a serial system can trivially emulate a parallel system, while the converse is not true. This means that all of our experience solving problems on parallel systems can be directly applied to serial systems, while our experience solving problems on serial systems can not be applied to parallel systems.

Don't get me wrong; I like designing parallel systems and did quite a bit of it for my PhD (among other things), but at the end of the day I'd rather have serial hardware than parallel, all other things being equal. Of course, in the real world, all other things are not equal. You can't get a general-purpose serial processor with the same IPC as a quad-core Opteron (for example), and individual processors typically have less (total) cache than multiprocessors.

Some of the code I wrote was originally aimed for a 64-processor 600MHz MIPS box and ended up running on a 16-core 2.4GHz Opteron system. It was highly parallel (roughly 1000 lightweight processes, often more), but still ran faster on the 16-processor machine than the 64-processor one, because each of the 16 processors was more than four times the speed of the old MIPS chips (and that's ignoring the latency differences).

Re:Excessive?

[drix]

we all know that games generally speaking are the most intensive software ever run on a PC Not even close. Games, after all, run in realtime. There are many, many applications out there that have no problem pegging top-of-the-line hardware for hours on end: DV editing, raytracing, scientific computing. In fact, the whole reason I'm posting this is because I'm waiting for my PC to solve a big math problem :-)

Re:Excessive?

[Luyseyal]

Not to mention running something like World Community Grid. I love using my idle processor time to tackle AIDS, Cancer, Muscular Dystrophy, Dengue, etc.

-l

Re:what does "desktop processor attributes" mean?

[Kjella]

Servers are about making a lot of people happy at a reasonable speed. Desktops are about making one user happy at an extreme speed. A lot of crap is single-threaded or not suitable for parallelization, and the best solution is to push that single thread at maximum speed. That's the only desktop quality of significance I know of. With that said, I have a quad-core and my biggest annoyance right now is disk thrashing. My CPU is usually almost idle, but having a lot of tasks using the disks at the same time slows everything down. I really look forward to SSDs and near penalty-free random access.

Re:what does "desktop processor attributes" mean?

[forkazoo]

I've seen this before, I've never understood it. What does it mean?

Thad

I'm not intimately familiar with the specifics in this case, but starting with a server chip and "adding desktop processor attributes" would typically entail:

adding the inability to use ECC.
adding a reduction in cache.
adding a lack of fault tolerance or error checking capabilities.
adding the feature of being impossible to use with > 2 sockets.
adding a whizzy new marketing name.

And, the enthusiast desktop parts are often easy to overclock, while server parts assume you'll just buy a faster CPU instead of wasting time fiddling with something that may catch fire.

BTW, hey, I remember you from alt.movies.visual-effects "back in the day" before the death of Usenet. good to see you haven't fallen off the face of the planet. I'm not in the process of working on a compositing demo reel so I can try to jump from straight IT to visual effects in the near future. I blame this career change in part on all your interesting and informative posts getting stuck in my head. :)

Re:Also, Duncan Hill Coffee

[dino2gnt]

Yes, but is it blacker than the blackest black times infinity?

Hertz? I care about Watt much more

[dallaylaen]

I have no need for a machine more powerful than mine. I would rather buy a silent one.

Re:So...

[Bluesman]

My third party observation is that 99% girls will look at the comfortable and stable guys, wonder why THEY can't find a guy like that, and then hop in the M3 with the asshole.

Confucius say, a small dick is still better than an unused one.

Re:But...

[WhiteWolf666]

I'm skeptical about that.

I'd be fairly certain that the NSA uses some kind of off-the-shelf processors, whether that be Power, Itanium, or X86.

What the NSA does different, most likely, is scale. You put 1,000 of these in a supercomputer? They'll put 100,000.

Chip fabs are expensive, as is chip design. There's no reason not to leave that to the experts (AMD/Intel). It's a commodity process, and they'll do it better than the government ever can.

Supercomputer design is something else. That's not commodity; and it's a simple scaling problem. More $$ = Bigger computer.

Why should they bother reinventing the wheel?

Re:Excessive?

[scheme]

Im pretty sure games stress the whole system overall a lot more than any application im aware of. math problems and ray tracing and DV editing if im not mistaken are CPU exclusive operations. Im not an expert but high end graphics cards are more powerful than cpus, even if they are specialized. i cant think of any other application that will stress the CPU, GPU, RAM, HDD and everything else to 100% other than games.

You are mistaken. Take particle physics simulations for example. The system might be downloading a 10GB dataset to do the next simulation while it's working on simulations of a detector which involves working with the current dataset. The download would max out your net connection while the simulation work would max your cpu and require something like 2-3GB of ram. The two activities are probably generating a decent i/o load as well.

Same deal with audio or video processing, if you're streaming a video or audio source or two from an array, processing it and writing it back, well that's pretty much using everything. Given a raw video stream can be about 20MB/s, you're generating about 40MB/s of read/write per stream and you might be working with a few streams if you're trying to overlay two video sources or something. That's significantly more activity than a game will produce.