IDF Spring 2005 Day 1 - Gelsinger Speaks, nForce4 Intel and more
by Anand Lal Shimpi & Derek Wilson on March 2, 2005 3:00 AM EST- Posted in
- Trade Shows
Parallelism, Parallelism, Parallelism
Pat has been on a parallelism kick for a long time now, when AMD was talking about 64-bit and on-die memory controllers, all Pat was interested in was parallelism, more specifically, thread level parallelism. It’s taken a few years but Pat is finally seeing his exciting technology come to fruition given that next year Intel expects 80% of its total shipments to be multi-core enabled.
Pat put together an interesting graph showing the performance boost we’ve seen during the lifetime of the Pentium 4 from inception to present day, showing that the current generation Pentium 4 is about 3x the speed of the first Pentium 4:
Assuming heavy adoption of multithreaded software and workloads, this is where Pat sees performance going in the future due to multi-core:
It’s a very optimistic chart, but Pat Gelsinger isn’t one to exaggerate. Given that the entire industry is behind multi-core CPUs, it wouldn’t be too far fetched to see significant performance gains due to multi-core between now and the end of the decade.
We mentioned before that Intel predicted 8-thread CPUs by the end of the decade, but now we have some more information about that prediction thanks to Pat’s presentation:
In mainstream client PCs, Intel will be shipping CPUs capable of executing two simultaneous threads in 2005; these will be the Pentium D (dual core) and the regular Pentium 4 with Hyper Threading (2 logical processors). Note that Intel isn’t including the Pentium Extreme Edition in this figure which is able to execute 4 threads simultaneously, as it is not a mainstream client processor. On the server side Intel is targeting 8 threads per CPU in 2005/2006.
Towards the end of the decade, Intel is pushing for 8 threads on clients - most likely meaning that mainstream PCs will have four independent cores, each with Hyper Threading technology (or 8 simpler cores instead). And on the server, we’re looking at 32 threads per chip by the end of the decade.
Facebook
Twitter
RSS
19 Comments
View All Comments
sprockkets - Thursday, March 3, 2005 - link
FDimms, the way we get around having each processor have its own memory controller but still have a lot of slots.Of course the right time for you to enter 64 bit for the desktop is now, cause you managed to kludge it on your processors oh BFD.
Sorry Patty, but the AMD/Linux crowd didn't need to wait on the Wintel crowd for 64 bit.
Viditor - Thursday, March 3, 2005 - link
mickyb - "How is this any different than multi CPU SMP? It isn't, except for compressing them to a smaller space"I agree, at least for the Intel model. The AMD dual core is quite different however, in that the 2 cores are connected locally, and the MOESI protocol AMD uses allows for easy cache snooping.
mickyb - Wednesday, March 2, 2005 - link
#16 My statement was in context of the performance issue, not cost. The graphs imply that multi-core is the savior, when we have been experiencing what multi-cpu will do. Any graph that implies "exponential" growth for multi-core vs. single core is just a lie. I have been creating tools that analyze system performance for a while. It is far from the truth. It is still SMP.JarredWalton - Wednesday, March 2, 2005 - link
#9: "How is this any different than multi CPU SMP? It isn't, except for compressing them to a smaller space. SMP has its problems as well and the number of CPUs does not create an exponential graph like Intel is implying."Actually, there is one major difference: one socket is sufficient. Designing motherboards with two CPU sockets increases costs a lot, and so the SMP market in the desktop space is extremely small. There are so many applications that *could* use multiple threads that don't, mostly because programmers would end up spending tons of effort in improving performance on a small percentage of systems.
Just like the move to 64-bits will have more benefits in the future rather than in the short term, multi-core is looking to the future rather than the present. Software will have to be coded properly, but once that is done, multi-core will start to give us a lot of improvements in performance. Now that programmers have an incentive to support threading (probably almost all CPUs sold by late 2007 are going to be SMP), they will spend the time.
Viditor - Wednesday, March 2, 2005 - link
Cygni - "Im also wondering whether Intel will allow the Nforce4 to use the 1066fsb?"Good question...
I would ASSUME that the answer to this is yes..."In for a penny, in for a pound" as they say. Considering all the trouble Intel has been having with 3rd party developers lately, I would assume that they will support Nvidia completely (if not, why support them at all?).
Cygni - Wednesday, March 2, 2005 - link
I thought the Pressler info was pretty shocking too, Viditor... especially considering its slated to replace the single die PD's. Wouldnt two phsyical cores over FSB be far slower? Really strange.Im also wondering whether Intel will allow the Nforce4 to use the 1066fsb?
johnsonx - Wednesday, March 2, 2005 - link
We all said it, and I think recent events and comments at IDF prove it beyond a reasonable doubt: Intel and Microsoft were in cahoots on XP 64-bit all along. XP x64 wasn't ever going to be released before Intel had 64-bit capable P4's and even Celerons ready to go.Viditor - Wednesday, March 2, 2005 - link
I hadn't realised how very different Intel's dual-core is to AMD's until now...Just this one line:
"The two cores in Pressler are totally independent, meaning that they must communicate with each other over the external front side bus and not over any internal bus"
That will be a HUGE latency problem when compared to AMD's dual-core! It takes 1 clock for AMD to communicate from one core to another because of Direct Connect (the on-die "switcher" between cores), I would guess that Intel will require at least 3-5 clocks (based on it taking 6-8 clocks for their SMP)...
skiboysteve - Wednesday, March 2, 2005 - link
even if intel did implement x86-64 at the "right time" at the launch of Win64, how is this the right time? Anyone who wants to use Win64 that intel now harks as the future, will have to buy a new chip, while A64 users already have everything they needed.I dont even have an athlon64 but its just silly to say you are at the forefront of some new technology (64bit) when everyone has to go buy a new CPU to use it, amd had it right by equiping users before the "right time" so they were prepared.
Anyway. Yeah FBDIMMs look tight, so does rambus, the DDRx shit is so lame. by the time they get DDR2 up and running they find 100 ways to do it better and no one wants DDR2 anymore, its stupid.
and BTX is a joke
sphinx - Wednesday, March 2, 2005 - link
I agree with mickyb. I am more interested in the FBDIMM than NF4 for Intel and dual cores.