Intel Core i7 (Nehalem) benchmarks are out

[Uraniun235]

Don't forget that period of time when Pentium 4 chips required RAMBUS... ugh.

Bean, current processor design is about achieving a balance between heat output, power consumption and processing power. Too much clock speed either requires extremely good chip samples, or extremely good cooling systems. It's a toss up.

The other major problem is that if you need a really good cooling system, it means you're consuming a lot of electricity and dumping a lot of heat into the surrounding air. Not so much of a problem for a home enthusiast, but a major problem for server manufacturers whose customers have to worry about electrical and cooling needs in a server room or datacenter.

[Fingolfin_Noldor]

Don't forget that period of time when Pentium 4 chips required RAMBUS... ugh.

I remember that not only were the early motherboards for that chip with RAMBUS was less than satisfactory, the RAMBUS drove the total price of the system up. I mean, which fool would pay more for performance that one could get with a cheaper platform?

[The Kernel]

I suppose they could hope for a colossal Intel fuckup again. Let's be brutally honest: they should never have been competitive. The only reason they became competitive in the first place was a series of Intel missteps: unreasonably high pricing, new CPUs that were only marginally better than their predecessors, a spectacularly failed and buggy motherboard chipset which shook consumer confidence in the brand, and all of this happening at the same time.

Actually probably their biggest blunder was not realizing that electron leakage would become a serious problem at 90nm. Remember that Northwood was actually quite a successful chip and held the performance crown for some time. However, the clockspeed that Intel had planned never panned out; they were expecting to hit 5ghz at 90nm and it never materialized thanks to the aforementioned electron leakage problem. Fundamentally though, Pentium 4 was just not a very smart design even if the manufacturing problem hadn't come up. It was designed for an era where clockspeed was king and no one gave a shit about power consumption, which is not very forward looking. Unfortunately when you have to milk a design for 2-3 years minimum and may go 5 years before a major redesign, you have to be very forward thinking in the chip business.

However, it does appear that Intel has learned their lesson and is not resting on their laurels. The Tick-Tock cadence alone brings a much welcome design refresh cycle to Intel products if they can afford to keep it up.

[phongn]

I suppose they could hope for a colossal Intel fuckup again. Let's be brutally honest: they should never have been competitive. The only reason they became competitive in the first place was a series of Intel missteps: unreasonably high pricing, new CPUs that were only marginally better than their predecessors, a spectacularly failed and buggy motherboard chipset which shook consumer confidence in the brand, and all of this happening at the same time.

AMD also benefited from acquiring a solid CPU design team when they purchased NexGen (who would design the K6) and a bunch of ex-DEC engineers when Compaq killed that company (which provided the EV6 bus design for the K7). Add in the integrated memory controller in the K8, and they really were able to get a good design going for quite awhile. Then they seemed to have run out of steam and overpaid for ATI.

Actually probably their biggest blunder was not realizing that electron leakage would become a serious problem at 90nm. Remember that Northwood was actually quite a successful chip and held the performance crown for some time. However, the clockspeed that Intel had planned never panned out; they were expecting to hit 5ghz at 90nm and it never materialized thanks to the aforementioned electron leakage problem.

To be fair to Intel, the entire semiconductor industry ran right into that wall at 90nm.

[Kitsune]

This might be worthy a new thread but going to post it here....

My fastest computer is an AMD Athlon XP 2800+

If I was to upgrade to current standards, what applications would I notice a performance jump in?
I am trying to find out where we have gained over the last few years?

[phongn]

If I was to upgrade to current standards, what applications would I notice a performance jump in?
I am trying to find out where we have gained over the last few years?

Well, a dual-core processor will make your computer run multiple applications better, and the Core 2 is significantly faster clock-for-clock compared to the old Athlon X2. If you run play games you'll probably get a framerate boost (unless you're GPU-limited, anyways).

[Starglider]

My guess is that Intel have a couple of clock steps in reserve on i7. However 3.6 GHz or above in a production part (as opposed to hand-picked demo chips) probably isn't feasible right now; most motherboard/cooler combos wouldn't be able to handle the power and cooling requirements, but that won't stop idiots putting expensive high-freq CPUs in budget motherboards anyway. The rash of ensuing failures would be really bad PR, with consumers and motherboard manufacturers. Better to keep the rated clocks down and let the overclockers take the blame if they destroy their CPUs.

Thus I think you will see 4 GHz rated Nehalems but not until the next die shrink, and probably the second or third stepping of that, late 2010 maybe. I do agree that 4 to 8 cores per chip is getting excessively hard to parallelise for and that we've probably wrung nearly as much IPC as we're likely to get out of x86 without radical instruction set changes, so a change of focus back to clock speeds would be welcome.

As for AMD, it's only a combination of somewhat underwhelming i7 reviews and stories like this that allow me to hold out hope for them.

[The Kernel]

To be fair to Intel, the entire semiconductor industry ran right into that wall at 90nm.

True, but the P4 design exacerbated the problem with it's deep pipelining and need for massive clockspeed to overcome it's relatively low IPC.

[Arthur_Tuxedo]

Pentium 4 was a disaster for Intel and probably single-handedly allowed AMD to shoot ahead for those years even without Intel's other fuckups. It's no accident that Core 2 is an evolution of the PIII design principles and P4 was left in the scrap heap.

[phongn]

Thus I think you will see 4 GHz rated Nehalems but not until the next die shrink, and probably the second or third stepping of that, late 2010 maybe. I do agree that 4 to 8 cores per chip is getting excessively hard to parallelise for and that we've probably wrung nearly as much IPC as we're likely to get out of x86 without radical instruction set changes, so a change of focus back to clock speeds would be welcome.

I don't think radical ISA changes would really help, either. x86 is terrible, yes, but the cost of cracking it really isn't much given modern transistor budgets.

As for AMD, it's only a combination of somewhat underwhelming i7 reviews and stories like this that allow me to hold out hope for them.

I know you're just using the Inq as an example, but still ... come on, Starglider

To be fair to Intel, the entire semiconductor industry ran right into that wall at 90nm.

True, but the P4 design exacerbated the problem with it's deep pipelining and need for massive clockspeed to overcome it's relatively low IPC.

Well, yes. SMT helped it fill its deep pipeline but certain architectural problems with NetBurst meant that it could have to flush the pipeline and drop performance.

Pentium 4 was a disaster for Intel and probably single-handedly allowed AMD to shoot ahead for those years even without Intel's other fuckups. It's no accident that Core 2 is an evolution of the PIII design principles and P4 was left in the scrap heap.

Honestly it wasn't a huge disaster. It performed well enough once Northwood came out, and at the time it was designed nobody could've forseen the problems at 90nm.

[DaveJB]

Prescott's problems went way beyond the 90nm process. Intel managed to make its performance worse, while somehow breaking it so that it didn't respond to cache increases and went to pieces if you tried to take the FSB above 800MHz. Had they put a shrunk Northwood out against the Athlon 64 it would probably have competed better (until they went dual-core, when the Athlon would probably have really started smacking down the P4 again).

[phongn]

Prescott's problems went way beyond the 90nm process. Intel managed to make its performance worse, while somehow breaking it so that it didn't respond to cache increases and went to pieces if you tried to take the FSB above 800MHz. Had they put a shrunk Northwood out against the Athlon 64 it would probably have competed better (until they went dual-core, when the Athlon would probably have really started smacking down the P4 again).

Prescott was an attempt to continue the "more clockspeed!" philosophy of NetBurst. I'm unsure as to whether a die-shrunk Northwood really would've been able to compete as well either, and it'd hit the same scaling wall but probably at a lower clock point.

[DaveJB]

The original Willamette/Northwood version of Netburst really benefited from increased cache and FSB speeds, and the P4EE managed to outpace the release Athlon 64 in some areas with a big but relatively slow Level 3 cache. I'm not saying it would have outperformed it, but if Intel had managed to shrink Northwood to 90nm, stick a 1MB or 2MB L2 cache on it and increase the FSB to 1066MHz, it would have competed with the Athlon 64 a lot better, at least until dual-cores arrived.

[Starglider]

I don't think radical ISA changes would really help, either. x86 is terrible, yes, but the cost of cracking it really isn't much given modern transistor budgets.

I agree that the cost of decoding x86 into RISC micro-ops isn't a significant problem. It uses a little more power and adds a little more latency but it isn't terribly signficant. What I was saying was that I think the general ideas behind IA64 (increase instruction level parallelism by allowing the compiler more control of the instruction stream processing, both direct and hinted) still have a fair chance of working. Itanium failed horribly of course (the P4 was just mediocre - Itanium was a genuine disaster), but the software landscape looks a lot different now compared to when it was first designed (early 90s). I don't know for sure that EPIC and similar can be made to work as promised, I'm just allowing for the (IMHO reasonable) possibility that they might be.

Other ISA improvements e.g. the various iterations of SSE, 64-bit ints, FMAC and the upcoming 256-bit vectors help too of course but that's a steady incremental process which tends to address various special cases in turn.

[Kitsune]

Well, a dual-core processor will make your computer run multiple applications better, and the Core 2 is significantly faster clock-for-clock compared to the old Athlon X2. If you run play games you'll probably get a framerate boost (unless you're GPU-limited, anyways).

Sounds more or less modest for me to be honest...

Really just working now on slowly bringing all of my computer's memory to maximum.....
It did seem I got a jump from 1 gig to 1.5 gig.

[Arthur_Tuxedo]

Core 2 is probably twice as fast as Athlon XP or more given identical clockspeeds and assuming you're only using one core. AXP 2800+ is 2.1 GHz and you can now get a Core 2 at 2.5 GHz for $90, which would be equivalent to an AXP in the 4-5 GHz range not counting the extra core. I think it's safe to say you'd notice a huge difference. Whether you actually need that extra power enough to pony up the money for the chip, mobo, and DDR2 RAM is the question, however (if you have AXP you're on DDR RAM and that won't be compatible).

I don't think you would see much benefit with more than 1.5 GB RAM on that system, although I could be wrong.

[Kitsune]

I go with the three to four times speed rule, when processors get three to four times faster as far as performance, I usually upgrade.

Also, you kind of underestimate price......
In this case, I would have to buy processor, motherboard, memory, and video card.

[phongn]

I agree that the cost of decoding x86 into RISC micro-ops isn't a significant problem. It uses a little more power and adds a little more latency but it isn't terribly signficant. What I was saying was that I think the general ideas behind IA64 (increase instruction level parallelism by allowing the compiler more control of the instruction stream processing, both direct and hinted) still have a fair chance of working. Itanium failed horribly of course (the P4 was just mediocre - Itanium was a genuine disaster), but the software landscape looks a lot different now compared to when it was first designed (early 90s). I don't know for sure that EPIC and similar can be made to work as promised, I'm just allowing for the (IMHO reasonable) possibility that they might be.

I'm not optimistic on if VLIW architectures can be made to work well for the majority of programs. Now, for some purposes they work extremely well, but most applications aren't the sort of scientific or HPC code that VLIW excels at. Compilers just haven't really developed to the extent that VLIW needs, and I don't see them doing so for the foreseeable future.

Other ISA improvements e.g. the various iterations of SSE, 64-bit ints, FMAC and the upcoming 256-bit vectors help too of course but that's a steady incremental process which tends to address various special cases in turn.

Those are pretty much all special-case solutions, really. Though, people've been pleading for FMAC for years and years ...

I go with the three to four times speed rule, when processors get three to four times faster as far as performance, I usually upgrade.

If you want an interim machine, you could get a 780G-based motherboard and a dual-core Athlon64 for not much money, and it'll be quite a bit faster than your present machine.

[Kitsune]

You need to understand that for the most part I am fine....
I tend to buy older games simply because they are cheap

I was more just curious than anything else, I can only use my own machines as benchmarks as well.

[Starglider]

I'm not optimistic on if VLIW architectures can be made to work well for the majority of programs. Now, for some purposes they work extremely well, but most applications aren't the sort of scientific or HPC code that VLIW excels at. Compilers just haven't really developed to the extent that VLIW needs, and I don't see them doing so for the foreseeable future.

Compilers have advanced quite a lot since 1994, when IA64 was first designed. The best modern compilers are capable of global state analysis on a scale that was impossible 15 years ago. Semantic anlysis and algorithm rewriting is now possible on a local level, though this technology is mostly still in the lab (see start-up efforts such as supercompilation, or various interesting papers on metacompilation). This technology is potentially able to make much more use of fine-grained processor control (VLIW explicit instruction parallelism is only one type of this, there are many other types of hinting, predication and control flow possible). Global semantic analysis is still a pipe dream as far as the IT industry is concerned, but my own start-up is working on very exciting technology that is making breakthroughs in that area. It works using a software model (kind of like an emulator, but with more metadata) of the target (virtual) machine, and I really do think that technology like this could make good use of a target architecture that allowed more control.

So yeah, I'm optimistic that it could work, but after Itanium (and the slow death of all non-x86 architectures, save possibly Power) I'm very pessimistic about anyone putting major investment into such a project in the forseeable future.

[phongn]

Compilers have advanced quite a lot since 1994, when IA64 was first designed. The best modern compilers are capable of global state analysis on a scale that was impossible 15 years ago. Semantic anlysis and algorithm rewriting is now possible on a local level, though this technology is mostly still in the lab (see start-up efforts such as supercompilation, or various interesting papers on metacompilation). This technology is potentially able to make much more use of fine-grained processor control (VLIW explicit instruction parallelism is only one type of this, there are many other types of hinting, predication and control flow possible). Global semantic analysis is still a pipe dream as far as the IT industry is concerned, but my own start-up is working on very exciting technology that is making breakthroughs in that area. It works using a software model (kind of like an emulator, but with more metadata) of the target (virtual) machine, and I really do think that technology like this could make good use of a target architecture that allowed more control.

Well, I'll certainly be excited if such things start becoming more and more common! And I've certainly been looking at some of the other compiler designs like LLVM.

So yeah, I'm optimistic that it could work, but after Itanium (and the slow death of all non-x86 architectures, save possibly Power) I'm very pessimistic about anyone putting major investment into such a project in the forseeable future.

Well, there's a lot of risk in that sort of thing. Though, I think ARM and PowerPC will stick around for quite awhile since they continue to rule the embedded roost and I'm not sure if x86 can scale down that low.

[Starglider]

Though, I think ARM and PowerPC will stick around for quite awhile since they continue to rule the embedded roost and I'm not sure if x86 can scale down that low.

Intel is gunning for ARM with Atom, and if current trends continue I suspect they'll eventually suffocate them. Thumb is a great architecture for resource constrained systems (actually it's a nice architecture full stop), but the notion of cross-platform support is one that has been steadily dying off in IT since the 80s. Developers want the same tools and ideally the same code to work everywhere, regardless of whether the fundamental basis of those tools (x86) is crap. ARM's licensing fees are quite expensive and their power efficiency is constrained by the quality of the manufacturing process used by the licensor. Intel has state of the art manufacturing processes (in many cases previous-gen fabs that just ceased making desktop/laptop x86 cores) and (if history is anything to go by) can and will heavily subsidise Atom pricing to kill off the competitor.

My guess is that ARM will follow MIPS into marginalisation and practical irrelevancy. Power (in its various iterations) is really the last credible x86 competitor, at least as long as the current microprocessor design and manufacturing paradigm continues.

[phongn]

Intel is gunning for ARM with Atom, and if current trends continue I suspect they'll eventually suffocate them. Thumb is a great architecture for resource constrained systems (actually it's a nice architecture full stop), but the notion of cross-platform support is one that has been steadily dying off in IT since the 80s. Developers want the same tools and ideally the same code to work everywhere, regardless of whether the fundamental basis of those tools (x86) is crap. ARM's licensing fees are quite expensive and their power efficiency is constrained by the quality of the manufacturing process used by the licensor. Intel has state of the art manufacturing processes (in many cases previous-gen fabs that just ceased making desktop/laptop x86 cores) and (if history is anything to go by) can and will heavily subsidise Atom pricing to kill off the competitor.

I know about Intel's amazing processes and their enormous investment into getting x86 to scale to places nobody ever thought it'd go, but ARM does have a commanding lead in the embedded world. I'm still a bit skeptical if they can get Atom small enough power-wise to play down low.

My guess is that ARM will follow MIPS into marginalisation and practical irrelevancy. Power (in its various iterations) is really the last credible x86 competitor, at least as long as the current microprocessor design and manufacturing paradigm continues.

PowerPC seems to do well in the higher-performance embedded and HPC world. Plus, Freescale and IBM seem to be putting reasonable amounts of effort into keeping them going.