Chat summary: Pat about SMT, x86-64/I64, HT etc.
Here's my summary (copy & paste) of the most interesting statements by Pat:
pat: Sonix: Well, your employer has stated that hyperthreading will only be enabled in the P4 Xeon MP the first half of 2002. Hello. This is pat! Well, the first hyperthreading implementations will be in server (xeon). It will come into the desktop space later in the year and eventually into Mobile as well.
pat: re: transistor of 2009 -- Yes! isn't it very exciting that all of a sudden everyone is chasing us to prove that Moore's law continues to 2009 and beyond. Just a year ago, everyone questioned if we could continue to 2010 with Moore's law, now they are all saying they will beat us to it. I couldn't be happier.
pat: Sonix: Pat, is Hyperthreading an implementation of SMT? There is much confusion and discussion on this subject on many forums. Could you elaborate on this? We are viewing hyperthreading as the category of execution of multiple simultaneous threads running in a single core (which may in fact be multiple cores). As we begin integrating multiple threads into a single core there will be a time of great innovation regarind SMP, SMT and the types of scheduling of dynmaic, Switch on Event, compiler controlled, OS controlled etc. The next several years of threading research like the early days of pipeline, super-pipe, super scaling, dynamic, out of order etc. A renasaince of uArchitecture!
pat: onestone: K8/IA64 is very serious, to support both optimal you must have two binaries of the os and apps. Let me address overall the question on K8/Sledgehammer vs. Itanium. The bigger issue is -- designing big system which are inclusive of I/O subsystems, big memory hierarchies, large caches, ECC, machine check, manageability etc. These are the big hard problems. Adding 64 bits is aa very tiny piece of the challenge of building big iron. IPF is our strategy to address the "big task" of "big systems". SLedgehammer might add "64bits" but they have barely started to attack any of the big system problems. Its sort of neither fish (desktops don't need it) nor fowl (it doesn't address the big systems issues). IA32 will meet the needs of the desktop/mobile, ipf of big systems.
pat: as: NoName: hobold: Why has the P4 so many transistors? What needs 42 Million? The Athlon has only 37 Million... May be there is Hyperthreading already included ... Interesting question... I can hardly wait till you learn the answer...
pat: re: P4 is faster than Itanium: Well, depends what you are measuring it against. If you are looking at a 16way McKinley next year will blow away any configuration of P4P and Xeon configs next year due to the big memory, caches etc. on Tpc-C, tpc-W, etc.
pat: NoName: The AMD Palomino has SSE extensions. Why has the Northwood no 3DNow! extensions? The difference between leaders and followers -- leaders lead, followers follow and immitate
pat: Prost: We saw the move from bare processore packages to heatsinks and now HSFs, will the move to liquid cooling solutions be inevitable in the future? I don't think it will be cost effective for desktops/mobile any time soon (weight and form factor play a big role). vapor heat pipes seem promising for servers however in high-density configurations.
pat: EasyFrag: Goethe: Did you hear that even AMD voted for 3GIO (Arapahoe from Intel) in the PCI-SIG? Yes, I think this was clear evidence that 3gio is firmly on track to be the PCI replacement and much more. Further, I think that HT while decent technology will end up being niched as a result.
pat: re: effor into processors/compilers which is right (ala Hp Dynamo). I think the big transiiton is to dynamic languages (like C# and Java). The JIT/Run-time/GC will be difficult to dinstinguish where the SW environment stops and the uCode starts and the uCode stops and HW starts. You'll see this be a big issue
pat: re: A-Block multithreaded code -- Yes, we are already working heavily in this area due to our KAI acquisition. We also published some research results on this at miprocessor Forum I believe.
pat: re: lite version of ia-64 -- No. IA32 with continuing improvements/extensions will be the desktop architecture...
pat: re: memory of the future -- right now it looks like the ADT work is merging nicely into DDR2 JEDEC efforts (some might fall into DDR3 and beyond).
pat: re: Rambus dead? no, its great techonlogy but given the legal and industry dynamics, it will be a nich technology
pat: re: jave/c# optimizations: GIven the growing role of dynamic languages, you'll see as focus on uArch/instruction enhancements which do a great job on these lanugages and evironments.
pat: re: 300mm -- we've stated this as Q1 and we're on track for that. The fabs are big, highly automated and very impressive. No one will deliver even a fraction on 300mm of what will we next year. The issue isn't the first, its who gains the cost advantage through volume of the larger wafers.
pat: Marco re: electrical or optical -- our research work is focused on making optical cheap forthe future. However, today we're at 2.5Gbps, I'm certain we can take copper to >10gbps so it has lots of legs left. Optical needs to make dramatic cost reductions before we can use it in interchip work.
pat: re: article for c't -- I guess well need to negotiate the royalty structure first!
pat: re: distributed environment -- I'm a huge fan of distributed computing at every level -- at the course level like p2p, at the focused tera-grid model, clustered computing and smp/smt. There is a very broad spectrum of course to fine grain distributed computing and low-level to software level. Across the spectrum we need research and implementaitons. This is the way of the future.
pat: re: memory latency -- a very hard problem since the access times for memory are pretty stable and relative to the CPU clock rate, the latency is approaching 1000 CPU clocks. Our approaches need to focus on memory hierarchiy (as we have been), threading (which allows us to hide latency of one thread while executing another ) and software systems that minimize memory delays (allow, minimize lock use, semaphores etc) where latency is a killer.
pat: re: chip market of the future -- I'm very bullish. I think the internet is going from adolescence exuberience to a period of long-term sustained growth. I expect the hh market to continue strong growth, PC to rebound, servers to have solid growth and comm, after a few more quarters of struggle to begin to resume.
pat: I further expect that if Intel can posiiton itself in communications and lead the Comm-Puting industry, we're on track to be a $100B company in this decade.
pat: re: internet in wrong direction - the beauty of the net is that no one (including commercial entities) control it. its a technological democracy of the purist form. No one controls, we all vote and the best win
pat: re: wine - andreas will receive it with appropriate pomp and circumstance at IDF.
Andreas
PS: It looks like we might see a chat with Dirk Meyer soon (as: Tobbic: @as is there a possibilty that you will do this with Dirk Meyer also sometimes in the future? That a good idea, I will work on it.) |
