Even if that is an obligatory standard comment for every hardware-related-story, there is a grain of truth to it.
Most chip-design clusters use either Platform Computing's LSF or Sun's Grid Computing Engine to basically make them look like one big entity. Chip design doesn't require cool low-latency inter-process communication, or much shared data (except for disk storage, easily enough handled with NFS), so Beowulf'ing them doesn't really help. But from a user's perspective they don't really need to know that node0025395 has load of 0.0, so quick get your job over there -- they just submit their job to a cluster scheduler and it finds the low-loaded nodes.
Gigabit Ethernet (and 10GigE) is called "Ethernet" basically because it keeps the same frame size restrictions (64-1518 bytes), CRC algorithm, packet structure (DA, SA, etc), and extensions (802.3x flow control, 802.1p prioritization, 802.1q VLAN, etc).
Ethernet is really a layer 2 protocol, with many Layer 1 implementations (i.e. 10Base5, 10BaseT, 100BaseT, 100VG, 100BaseFL, 1000BaseT, etc) -- the physical signalling for almost all of those is different based on the characteristics of the signalling medium. Unfortunately the term "Ethernet" is often confused with the original Layer 1 implementation of 10 meg Ethernet.
However OC-192 or OC-48 will give you a lower ping time since the packet-transmission delay (and therefore store-and-forward time) is dramatically reduced (if your previous backbone was mostly OC-12s, for example). You are correct in asserting that multiple parallel connections will not improve the latency.
that's a bridge-based way to look at it. imagine if all the routers were configured to route all the packets to the registration machine? Non-trivial with a typical hardware-based router feature set (except for very high-end Cisco), but for a software-based device like this, pretty easy to categorize web traffic and forward/rewrite all internal http packets (or even just all IP traffic, pretty much stonewalling you in with little CPU effort expended). Then when you register on the page it sends some lovely token enabling routing from your IP and briding from your MAC.
Damn I wish I had the points to mod the parent up.
To be honest, I wasn't even concerned with the article, just more MPAA-posturing, until I got to the part where MSNBC started to quote the logic-impaired John Fithian.
First of all, I suspect that people who watch R-rated movies (a large number of whom are are adults, by design), happen to not have enough time to go to the theater (gasp! horror!). Instead they either a) wait 3 months and rent it on satellite/cable PPV, or b) wait 6-9 months and rent it from Blockbuster or Netflix. Of course, the movie industry wanted to kill the VCR when it first came out because it would damage their primary market (the theater). Now rentals are the primary market (from a revenue standpoint).
Second, the R-rating inherently limits the potential audience of a film. For Fithian to claim that "R-rated pictures do not sell" is bizarre. In absolute terms, he's probably right, PG- or G-rated films have higher total admission figures. But if you looked at the ratio of viewers to total possible viewers, would R-rated movies be better or worse than G/PG? What about if you removed over-17's from the "total possible viewers" for G-rated movies. Then what do the numbers look like? I would bet R-rated movies do just fine, thanks.
If the movie industry (and its partners like the theater owners) want to focus on movies-as-business, then please don't come asking for the protection of movies-as-art (copyright extensions, etc). Fundamentally the ratings system is a business-oriented policy that results in some chilling effect on what kinds of movies are made, and that is always bad. Remember that because of the influence of the MPAA, unrated movies often cannot be advertised in many suburban (and even urban) newspapers (for anti-porn purposes, supposedly). So the indie director who doesn't want to submit to the MPAA has no choice, if she wants to get her message out to more than a handful of viewers.
Sigh. Please support your local arthouse/indie films. And don't choose blockbuster, they only further the ratings problem by not stocking NC-17 rentals and requesting alternate/edited versions of even some R-rated films. Hooray for maturity!
> Hmmm... 3MB is astounding, but 8MB is
> unremarkable... Well, I'd have to agree. I
> haven't bought a server with less than 4MB of
> cache in years. Oops, the SunBlade is only a
> workstation... Kinda makes you wonder.
Uh, the 8MB L2 is off-chip on US3 (ever wonder why USparc processors come as those big modules?). the 3MB L3 (and of course also the L2) is on-chip on McKinley. That's why it's a BFD. Try to read a little bit about the processors before you go off on them.
Sun hardware is not fast (enough) if you need really high speed. Price/performance is also horrible.
Linux reliability is good, but not as good as Sun, at least not when you have maintenance windows like you are quoting. However, you ought to just be able to buy 2x redundant Linux/PC everything to cover your multi-million-dollar failure cases and just swap over in the event of failure.
I use almost all Linux for computation servers at work (in the chip-design space, actually), we have a few Suns kicking around just in case an app isn't supported. The Sun HW is almost always half the speed. Our linux boxes can run 90+ days with 100% cpu load, no worry (of course, so can the suns). But, we can buy 2 or 3 times as many fully equipped Linux servers per Sun.
My bad. I confused two magazine stories. The one I was thinking of was a sort of FM waveform synthesis (i.e. think of your old Yamaha DX-7 keyboard) called Sondius-XG.
Not worth nearly as much, sadly as the PCR patent would have been.
Stanford does have some DNA-related patents/technologies in the area of gene splicing which brought in ~$40M in revenue in 1998 (according to the OTL). Goes to show that I should shut up about biotech in detail though and stick to microprocessors.
Here is the general report for 1999-2000 from the OTL.
Disclaimer: I'm a Stanford alum and sometime-studier of silicon valley lore.
MIT's spinoff list is all well and good but if you add up Stanford's it'd eat them alive:
Hewlett-Packard, SGI, Sun, Cisco (regardless of what you believe from this story), Yahoo, Intuit, IDEO just for starters. You could even probably find a way to argue Intel (as an offspring of Fairchild, which settled in SV because of Bob Noyce, and probably indirectly because of Fred Terman and David Starr Jordan), but it'd be a stretch.
Fred Terman was the Dean of the Engineering school when Hewlett and Packard were at Stanford, he was their mentor and encouraged them strongly to start a company. It's well-documented.
Stanford has been highly entwined with the venture capital community in silicon valley since the 70's (perhaps even earlier, I can't be definitive). If you wander around the campus you can see the synergy just in the names on the buildings (Gates, Allen, Hewlett, Packard, Clark), much less the buildings built with money from the Stanford Engineering Venture Fund, which is run by any number of top-tier VCs.
That quote in the story is from Tom Reindfleisch if I recall, in an internal university memo. Like they wouldn't have axes to grind internally, or want to influence future policy. Always, always remember the context of quotes, especially in media-driven stories like this one.
Stanford these days takes equity holdings or cash. The Office of Technology Licensing happens to hold the licensing rights for the DNA polymerase chain reaction (PCR) which is the basis for most biotech. They make millions/billions of dollars for the university endowment every year. I wouldn't fault them for guessing wrong 20 years ago about the potential of two wackos who were (apparently at the time) bilking the university for some intellectual property.
There is a simple reason this technology is not used: it makes the devices virtually untestable. You have to go to great pains to test the asynchronous sections of the device, and all the currently used methodologies (think multi-billion-dollar industry from TSMC/UMC/LSILogic, software from MentorGraphics/Synopsys/etc, and the test harware from Schlumberger/HP/Tek/whoever) would have to be turned on their ears. How would you like the process of buying a CPU to be more like buying a lottery ticket (er, well more like blackjack, I guess). The yields on initial lots of CPUs are pretty poor, sometimes well less than 10% good
devices.
Lots of texts, papers, and symposiums are devoted to clockless/asynchronous circuits. There are a few, very small cases where the work makes sense. Designing a whole CPU, well that would basically mean starting over from von Neumann's conception of a computer. We have countless man-years of work invested in an infrastructure that has produced continuously exponentially improving performance for the better part of 4 decades. Don't knock it just because the research-flavor-o-the-month comes along and gets some headlines.
Last point: just because it doesn't have a clock doesn't mean there isn't a frequency attached to it. You just find the longest path through the device, measure the time (say 333 picoseconds) and you can convert that to a frequency equivalent (3 GHz in that case). This is how lots of I/O interfaces are characterized anyway, so it's not like voodoo math.
I wouldn't blame it on AMD. I'd blame it on the nature of an exponential/logarithmic yield curve. The PIV is a big mother of a chip right now, and would have been nearly unproduceable had they included all the stuff they originally planned for.
I suspect the PIV is a chip waiting for a process shrink, then you will see what it is all about. Remember the ugly, nasty Pentium 60MHz in 0.6 micron? Remember how much less heat a P66 in 0.5 put out? Remember that 0.5 micron chips later hit 90/100 MHz? That will happen here too. The Intel forecast of 2GHz by 4Q01 is probably too conservative, or they are playing coy.
I think the article does have a couple of interesting points though -- Intel relied on the trace cache too much, and probably didn't notice how much of a bottleneck it would be for their execution units if they could only issue 3 instructions from the trace cache (I suspect that some design parameter got changed late in the game, because that is a pretty bad mistake that I would expect to be noticed). However, I also think the author didn't realize how hard it is to run any SRAM-type structure at 1.5GHz, especially to scale it up to bigger sizes (his entire rant about 8K versus 64K I found pretty humorous).
The other humorous thing in the article is the comparison of cycle counts. The author spends lots of bold tags on making sure we know that MHz is not the only thing, but then looks at cycle counts. Well, bub, they are representing the same thing. For a given architecture one must consider the quotient of the clock cycle counts and the clock frequency to get a realistic measure of performance. You can implement a given chip with lots of short-fast pipeline stages or fewer, bigger stages. One approach is not "better" than the other -- it's dependent on process technology and what sample set of code you use to benchmark it on. Therefore, like most aspects of processor design, it's a tradeoff.
On thing he does get very right: I certainly wouldn't buy a PIV right now. But, I think that in 12 months everything will look OK. I also don't think that a "bad" (i.e. a little slow) first version of a chip is a reason to discount an entire architectural implementation.
This is purely targetted at routers and switches, with the extra glue logic they are tossing in. It's completely not designed for a laptop/desktop/whatever -- you would never use 3 GigE interfaces simultaneously except in a switching/routing/whatever environment.
The package is all wrong for the PC market as well, an 860 BGA? Probably requires 12-16 layer PCB just to breakout all the rows. But, that's a common stack for networking...
It'll make a real nice embedded web cache controller (or content-switch, or whatever the term-du-jour is).
And those who claim vapor on it -- I suspect this part is vapor, but I would bet dollars-to-donuts that they already have a single-CPU test chip in their lab without all the other cool goodies in this version. TSMC 0.15 micron process has been stable for a little while now, so the test chip could even be in the same process.
Not sure if that is the motivation behind this cancellation, but it would explain some of the slippages in other projects.
Other sources I have read say that intel has lost a lot of their top chip architects over time -- they aren't interested in doing process shrink after shrink after shrink for the different cores, which is why the HP-designed McKinley is going to smoke the Intel-designed Merced. I would suspect the average age of an intel employee is around 28-30 by now, with the 2000-3000 new college hires every year for the last 4 years. Those people are the ones who got to architect the Willamette (Pentium IV) chip which is really the first new architecture since Pentium Pro (Pentium II and III are essentially the same micro-architecture with some fancy memory hacks and some SIMD extensions for the marketroids to use as "differentiators").
Anyone with 4-6 years experience still doesn't have a great grasp on the hurdles that will be hit on a 2 year-long project with many hundreds of people working in different groups. I know this for a fact, because I am one of those people, who could easily be working at Intel right now. Missing deadlines will be the norm at Intel, not the exception, no matter how many warm bodies get thrown at a project, because new designs are much harder than process shrinks, and quality design practices take lots of experience to be learned adequately.
The only form of intellectual property that requires this "enforce it or lose it" is a trademark.
That's not what the GPL is - as you say, it is a contract (or a license, if you like). I believe the reason for the bickering is that the GPL is a very strong, powerful license that has a very distinct "meaning" in software-development circles. To water it down in any way could pose a problem in the future, if (for example) it was tested in court.
I have no idea where these legal urban legends about enforce-it-or-lose-it start, but I'd like to quash them.
While it is nice that the machine has GigE, I can only assume that it is connected to their PCI bus in some manner. Since the machine only has 33MHz PCI, even with 64-bit access you will be blowing through half of the available bandwith (or so, 64-bit x 33MHz = 264MB/sec = 2112 Mb/sec) at line-rate. Hope you weren't planning on doing anything else at the same time...
Too bad they didn't up for 64-bit 66MHz PCI, which makes the BW crunch much less severe.
And you should start to see on-motherboard 10/100/1000baseT in a lot of places pretty soon, at least in high-end systems. I think it adds about 12 bucks to system cost assuming a discrete-component solution. Once intel integrates a new Ethernet MAC into their motherboard silicon that cost will go down considerably.
Chip area is close to free. It's not like they don't have some custom motherboard ASICs in there already with plenty of useable area. What's an audio controller take, about 4 sq mm, being generous? You're adding about 10 cents to die costs, worst-case, and maybe some schedule time, assuming that they didn't have a block like that lying around already (which I wouldn't believe anyway).
The pins and board-level parts make up most of the cost you would be concerned - PCB area and crappy little connectors and discretes, which was exactly what he was complaining about. Apple saves a buck-fifty and gets an attractive base price, Joe Schmoe has to buy a $25 addon to get useful functionality from his shiny new toy.
And I'm sure there's not a single DSP available in the box that they couldn't provide a driver for that you could run a semi-usable sound card off of either.
Leave it to Apple to dream up Yet Another Proprietary Non-Standard Cable. ADB, RS422 (cuz 232 was only good enough for 99.99% of the world!), their own serial connectors, localtalk, Apple AUI (because MII-standard or AUI connectors would be too easy). Good lord.
I'm almost surprised they give you VGA out the back. Otherwise they could have a nice revenue stream from ADCVGA convertor dongles for everyone but Apple's monitors.
On the overall, it looks perty, seems like decent performance/upgradability specs, and probably would make a nice Linux box.:-)
1. Berkeley is not in the "heart of Silicon Valley". More like the fringe/edge. In geographies, at least, you should look to Santa Clara/Sunnyvale for the valley. And Hollywood, well who knows. The HollywoodBerkeley connection is getting thinner and thinner (most people consider Northern and Southern California to behave like separate states anyway...) Maybe if you have the rally at the USC film school it would have your desired effect.
2. It's taking place at Berkeley because there is a nationally-known law professor taking part who happens to be at Boalt. Conveniently, Berkeley is a university with large lecture-hall facilities in which to watch something like this.
3. I don't understand your notion of "Berkeley". It's not like there is some unified front that represents Berkeley -- there is the town, the school, the alums, the CS dept, whatever. Sometimes one or more of these entities over-reacts to some issue, but to dismiss the whole lot of them for the actions of a few, well that's irresponsible.
I don't particularly attribute the conference to Berkeley -- it's the BayFF, who happen to have some friends at Berkeley. It could just as easily happen at Stanford, I'm sure. (no personal bias or anything... =)
A whole 3 months, eh? Welcome to the terrordome...
Most of the people who are complaining about separate categories are the people who want fine-grained control over censoring JonKatz or any other semi-controversial topic that might trouble them from their busy days of composing slashdot/Everything2 articles.
A side note to the complainers: Does anyone remember what the hell this place was like in 1998? Filtering? Are you kidding? It's supposed to be random-ass crap that someone may/may not think is interesting. If you don't like it, feel free to ignore it. It's not up to Rob to provide mechanisms for you to filter the "icky" stuff out of your daily reading. He did (and does) this because he thinks it's fun, and cool, nothing more, nothing less.
On leasing -- you are not the target market for a lease. The target market for a lease is the person who would be trading in the car around the time (or even before) he/she finished making the payments to BUY it.
In that case, what is the difference between leasing and owning? Why bother building equity if you are just going to turn around and trade it in? The interest used to be tax-deductible but that loophole is long-gone.
The people who lease like driving new cars where the maintenance is (mostly) taken care of. The car becomes an appliance that you pay to own, and every so often you get a new one when the first one becomes "obselete".
I don't exactly agree with this mentality (nor would you, I'm guessing), but I know lots of people like this, and I can sort of understand it. These people want the hassle out of their lives. Something breaks -- take it to the dealer, get a loaner, the dealer calls you in 3 days and you get your car back. If you own it, you have to either have a cool mechanic or a nice dealership (yah, right!) to diagnose and farm out the work for you to the appropriate places, get you the loaner, etc.
It's all about convenience. Of course, if you enjoy solving problems, probably that's not for you. That's why you are a developer/engineer/techie...
Slashdot Mirror
Even if that is an obligatory standard comment for every hardware-related-story, there is a grain of truth to it.
Most chip-design clusters use either Platform Computing's LSF or Sun's Grid Computing Engine to basically make them look like one big entity. Chip design doesn't require cool low-latency inter-process communication, or much shared data (except for disk storage, easily enough handled with NFS), so Beowulf'ing them doesn't really help. But from a user's perspective they don't really need to know that node0025395 has load of 0.0, so quick get your job over there -- they just submit their job to a cluster scheduler and it finds the low-loaded nodes.
I remember when "Hercules graphics" aka HGA was a specially supported monochrome graphics card. And having two monitors -- one MDA and one CGA (oooh!)
well, since 2^64 is enough to individually address every atom in the universe -- I'd hope so.
Erm, wrong. You are thinking of trademarks ("enforce it or lose it"). Patents do not work that way.
Gigabit Ethernet (and 10GigE) is called "Ethernet" basically because it keeps the same frame size restrictions (64-1518 bytes), CRC algorithm, packet structure (DA, SA, etc), and extensions (802.3x flow control, 802.1p prioritization, 802.1q VLAN, etc).
Ethernet is really a layer 2 protocol, with many Layer 1 implementations (i.e. 10Base5, 10BaseT, 100BaseT, 100VG, 100BaseFL, 1000BaseT, etc) -- the physical signalling for almost all of those is different based on the characteristics of the signalling medium. Unfortunately the term "Ethernet" is often confused with the original Layer 1 implementation of 10 meg Ethernet.
However OC-192 or OC-48 will give you a lower ping time since the packet-transmission delay (and therefore store-and-forward time) is dramatically reduced (if your previous backbone was mostly OC-12s, for example). You are correct in asserting that multiple parallel connections will not improve the latency.
that's a bridge-based way to look at it. imagine if all the routers were configured to route all the packets to the registration machine? Non-trivial with a typical hardware-based router
feature set (except for very high-end Cisco), but for a software-based device like this, pretty easy to categorize web traffic and forward/rewrite all internal http packets (or even just all IP traffic, pretty much stonewalling you in with little CPU effort expended). Then when you register on the page it sends some lovely token enabling routing from your IP and briding from your MAC.
Damn I wish I had the points to mod the parent up.
To be honest, I wasn't even concerned with the article, just more MPAA-posturing, until I got to the part where MSNBC started to quote the logic-impaired John Fithian.
First of all, I suspect that people who watch R-rated movies (a large number of whom are are adults, by design), happen to not have enough time to go to the theater (gasp! horror!). Instead they either a) wait 3 months and rent it on satellite/cable PPV, or b) wait 6-9 months and rent it from Blockbuster or Netflix. Of course, the movie industry wanted to kill the VCR when it first came out because it would damage their primary market (the theater). Now rentals are the primary market (from a revenue standpoint).
Second, the R-rating inherently limits the potential audience of a film. For Fithian to claim that "R-rated pictures do not sell" is bizarre. In absolute terms, he's probably right, PG- or G-rated films have higher total admission figures. But if you looked at the ratio of viewers to total possible viewers, would R-rated movies be better or worse than G/PG? What about if you removed over-17's from the "total possible viewers" for G-rated movies. Then what do the numbers look like? I would bet R-rated movies do just fine, thanks.
If the movie industry (and its partners like the theater owners) want to focus on movies-as-business, then please don't come asking for the protection of movies-as-art (copyright extensions, etc). Fundamentally the ratings system is a business-oriented policy that results in some chilling effect on what kinds of movies are made, and that is always bad. Remember that because of the influence of the MPAA, unrated movies often cannot be advertised in many suburban (and even urban) newspapers (for anti-porn purposes, supposedly). So the indie director who doesn't want to submit to the MPAA has no choice, if she wants to get her message out to more than a handful of viewers.
Sigh. Please support your local arthouse/indie films. And don't choose blockbuster, they only further the ratings problem by not stocking NC-17 rentals and requesting alternate/edited versions of even some R-rated films. Hooray for maturity!
uh, apple hires IDEO and/or Frog
all the other companies are cheap-a** bastards who won't pay for the best... except Palm, who hired IDEO.
> Hmmm... 3MB is astounding, but 8MB is
> unremarkable... Well, I'd have to agree. I
> haven't bought a server with less than 4MB of
> cache in years. Oops, the SunBlade is only a
> workstation... Kinda makes you wonder.
Uh, the 8MB L2 is off-chip on US3 (ever wonder why USparc processors come as those big modules?). the 3MB L3 (and of course also the L2) is on-chip on McKinley. That's why it's a BFD. Try to read a little bit about the processors before you go off on them.
Sun hardware is not fast (enough) if you need really high speed. Price/performance is also horrible.
Linux reliability is good, but not as good as Sun, at least not when you have maintenance windows like you are quoting. However, you ought to just be able to buy 2x redundant Linux/PC everything to cover your multi-million-dollar failure cases and just swap over in the event of failure.
I use almost all Linux for computation servers at work (in the chip-design space, actually), we have a few Suns kicking around just in case an app isn't supported. The Sun HW is almost always half the speed. Our linux boxes can run 90+ days with 100% cpu load, no worry (of course, so can the suns). But, we can buy 2 or 3 times as many fully equipped Linux servers per Sun.
Not worth nearly as much, sadly as the PCR patent would have been.
Stanford does have some DNA-related patents/technologies in the area of gene splicing which brought in ~$40M in revenue in 1998 (according to the OTL). Goes to show that I should shut up about biotech in detail though and stick to microprocessors.
Here is the general report for 1999-2000 from the OTL.
Oh hardly.
Disclaimer: I'm a Stanford alum and sometime-studier of silicon valley lore.
MIT's spinoff list is all well and good but if you add up Stanford's it'd eat them alive:
Hewlett-Packard, SGI, Sun, Cisco (regardless of what you believe from this story), Yahoo, Intuit, IDEO just for starters. You could even probably find a way to argue Intel (as an offspring of Fairchild, which settled in SV because of Bob Noyce, and probably indirectly because of Fred Terman and David Starr Jordan), but it'd be a stretch.
Fred Terman was the Dean of the Engineering school when Hewlett and Packard were at Stanford, he was their mentor and encouraged them strongly to start a company. It's well-documented.
Stanford has been highly entwined with the venture capital community in silicon valley since the 70's (perhaps even earlier, I can't be definitive). If you wander around the campus you can see the synergy just in the names on the buildings (Gates, Allen, Hewlett, Packard, Clark), much less the buildings built with money from the Stanford Engineering Venture Fund, which is run by any number of top-tier VCs.
That quote in the story is from Tom Reindfleisch if I recall, in an internal university memo. Like they wouldn't have axes to grind internally, or want to influence future policy. Always, always remember the context of quotes, especially in media-driven stories like this one.
Stanford these days takes equity holdings or cash. The Office of Technology Licensing happens to hold the licensing rights for the DNA polymerase chain reaction (PCR) which is the basis for most biotech. They make millions/billions of dollars for the university endowment every year. I wouldn't fault them for guessing wrong 20 years ago about the potential of two wackos who were (apparently at the time) bilking the university for some intellectual property.
There is a simple reason this technology is not used: it makes the devices virtually untestable. You have to go to great pains to test the asynchronous sections of the device, and all the currently used methodologies (think multi-billion-dollar industry from TSMC/UMC/LSILogic, software from MentorGraphics/Synopsys/etc, and the test harware from Schlumberger/HP/Tek/whoever) would have to be turned on their ears. How would you like the process of buying a CPU to be more like buying a lottery ticket (er, well more like blackjack, I guess). The yields on initial lots of CPUs are pretty poor, sometimes well less than 10% good
devices.
Lots of texts, papers, and symposiums are devoted to clockless/asynchronous circuits. There are a few, very small cases where the work makes sense. Designing a whole CPU, well that would basically mean starting over from von Neumann's conception of a computer. We have countless man-years of work invested in an infrastructure that has produced continuously exponentially improving performance for the better part of 4 decades. Don't knock it just because the research-flavor-o-the-month comes along and gets some headlines.
Last point: just because it doesn't have a clock doesn't mean there isn't a frequency attached to it. You just find the longest path through the device, measure the time (say 333 picoseconds) and you can convert that to a frequency equivalent (3 GHz in that case). This is how lots of I/O interfaces are characterized anyway, so it's not like voodoo math.
I wouldn't blame it on AMD. I'd blame it on the nature of an exponential/logarithmic yield curve. The PIV is a big mother of a chip right now, and would have been nearly unproduceable had they included all the stuff they originally planned for.
I suspect the PIV is a chip waiting for a process shrink, then you will see what it is all about. Remember the ugly, nasty Pentium 60MHz in 0.6 micron? Remember how much less heat a P66 in 0.5 put out? Remember that 0.5 micron chips later hit 90/100 MHz? That will happen here too. The Intel forecast of 2GHz by 4Q01 is probably too conservative, or they are playing coy.
I think the article does have a couple of interesting points though -- Intel relied on the trace cache too much, and probably didn't notice how much of a bottleneck it would be for their execution units if they could only issue 3 instructions from the trace cache (I suspect that some design parameter got changed late in the game, because that is a pretty bad mistake that I would expect to be noticed). However, I also think the author didn't realize how hard it is to run any SRAM-type structure at 1.5GHz, especially to scale it up to bigger sizes (his entire rant about 8K versus 64K I found pretty humorous).
The other humorous thing in the article is the comparison of cycle counts. The author spends lots of bold tags on making sure we know that MHz is not the only thing, but then looks at cycle counts. Well, bub, they are representing the same thing. For a given architecture one must consider the quotient of the clock cycle counts and the clock frequency to get a realistic measure of performance. You can implement a given chip with lots of short-fast pipeline stages or fewer, bigger stages. One approach is not "better" than the other -- it's dependent on process technology and what sample set of code you use to benchmark it on. Therefore, like most aspects of processor design, it's a tradeoff.
On thing he does get very right: I certainly wouldn't buy a PIV right now. But, I think that in 12 months everything will look OK. I also don't think that a "bad" (i.e. a little slow) first version of a chip is a reason to discount an entire architectural implementation.
This is purely targetted at routers and switches, with the extra glue logic they are tossing in. It's completely not designed for a laptop/desktop/whatever -- you would never use 3 GigE interfaces simultaneously except in a switching/routing/whatever environment.
The package is all wrong for the PC market as well, an 860 BGA? Probably requires 12-16 layer PCB just to breakout all the rows. But, that's a common stack for networking...
It'll make a real nice embedded web cache controller (or content-switch, or whatever the term-du-jour is).
And those who claim vapor on it -- I suspect this part is vapor, but I would bet dollars-to-donuts that they already have a single-CPU test chip in their lab without all the other cool goodies in this version. TSMC 0.15 micron process has been stable for a little while now, so the test chip could even be in the same process.
Other sources I have read say that intel has lost a lot of their top chip architects over time -- they aren't interested in doing process shrink after shrink after shrink for the different cores, which is why the HP-designed McKinley is going to smoke the Intel-designed Merced. I would suspect the average age of an intel employee is around 28-30 by now, with the 2000-3000 new college hires every year for the last 4 years. Those people are the ones who got to architect the Willamette (Pentium IV) chip which is really the first new architecture since Pentium Pro (Pentium II and III are essentially the same micro-architecture with some fancy memory hacks and some SIMD extensions for the marketroids to use as "differentiators").
Anyone with 4-6 years experience still doesn't have a great grasp on the hurdles that will be hit on a 2 year-long project with many hundreds of people working in different groups. I know this for a fact, because I am one of those people, who could easily be working at Intel right now. Missing deadlines will be the norm at Intel, not the exception, no matter how many warm bodies get thrown at a project, because new designs are much harder than process shrinks, and quality design practices take lots of experience to be learned adequately.
Comments?
Fighting fire with fire, can we have a little spelling check? It's grammar, not grammer.
"Beavis, grammar flames rule, huh-huh, huh-huh"
Sigh.
The only form of intellectual property that requires this "enforce it or lose it" is a trademark.
That's not what the GPL is - as you say, it is a contract (or a license, if you like). I believe the reason for the bickering is that the GPL is a very strong, powerful license that has a very distinct "meaning" in software-development circles. To water it down in any way could pose a problem in the future, if (for example) it was tested in court.
I have no idea where these legal urban legends about enforce-it-or-lose-it start, but I'd like to quash them.
Larry McVoy was reported to be working on some uber-version-control setup to help out Linus
(BitMover or something like that).
Did anything ever come out of that?
Re: gigabit ethernet
While it is nice that the machine has GigE, I can only assume that it is connected to their PCI bus in some manner. Since the machine only has 33MHz PCI, even with 64-bit access you will be blowing through half of the available bandwith (or so, 64-bit x 33MHz = 264MB/sec = 2112 Mb/sec) at line-rate. Hope you weren't planning on doing anything else at the same time...
Too bad they didn't up for 64-bit 66MHz PCI, which
makes the BW crunch much less severe.
And you should start to see on-motherboard 10/100/1000baseT in a lot of places pretty soon, at least in high-end systems. I think it adds about 12 bucks to system cost assuming a discrete-component solution. Once intel integrates a new Ethernet MAC into their motherboard silicon that cost will go down considerably.
Here's a reason to piss on apple then:
Chip area is close to free. It's not like they don't have some custom motherboard ASICs in there already with plenty of useable area. What's an audio controller take, about 4 sq mm, being generous? You're adding about 10 cents to die costs, worst-case, and maybe some schedule time, assuming that they didn't have a block like that lying around already (which I wouldn't believe anyway).
The pins and board-level parts make up most of the cost you would be concerned - PCB area and crappy little connectors and discretes, which was exactly what he was complaining about. Apple saves a buck-fifty and gets an attractive base price, Joe Schmoe has to buy a $25 addon to get useful functionality from his shiny new toy.
And I'm sure there's not a single DSP available in the box that they couldn't provide a driver for that you could run a semi-usable sound card off of either.
Apple apologists, fire away...
Leave it to Apple to dream up Yet Another Proprietary Non-Standard Cable. ADB, RS422 (cuz 232 was only good enough for 99.99% of the world!), their own serial connectors, localtalk, Apple AUI (because MII-standard or AUI connectors would be too easy). Good lord.
:-)
I'm almost surprised they give you VGA out the back. Otherwise they could have a nice revenue stream from ADCVGA convertor dongles for everyone but Apple's monitors.
On the overall, it looks perty, seems like decent performance/upgradability specs, and probably would make a nice Linux box.
Uhh...
1. Berkeley is not in the "heart of Silicon Valley". More like the fringe/edge. In geographies, at least, you should look to Santa Clara/Sunnyvale for the valley. And Hollywood, well who knows. The HollywoodBerkeley connection is getting thinner and thinner (most people consider Northern and Southern California to behave like separate states anyway...) Maybe if you have the rally at the USC film school it would have your desired effect.
2. It's taking place at Berkeley because there is a nationally-known law professor taking part who happens to be at Boalt. Conveniently, Berkeley is a university with large lecture-hall facilities in which to watch something like this.
3. I don't understand your notion of "Berkeley". It's not like there is some unified front that represents Berkeley -- there is the town, the school, the alums, the CS dept, whatever. Sometimes one or more of these entities over-reacts to some issue, but to dismiss the whole lot of them for the actions of a few, well that's irresponsible.
I don't particularly attribute the conference to Berkeley -- it's the BayFF, who happen to have some friends at Berkeley. It could just as easily happen at Stanford, I'm sure. (no personal bias or anything... =)
*chuckle*
A whole 3 months, eh? Welcome to the terrordome...
Most of the people who are complaining about separate categories are the people who want fine-grained control over censoring JonKatz or any other semi-controversial topic that might trouble them from their busy days of composing slashdot/Everything2 articles.
A side note to the complainers: Does anyone remember what the hell this place was like in 1998? Filtering? Are you kidding? It's supposed to be random-ass crap that someone may/may not think is interesting. If you don't like it, feel free to ignore it. It's not up to Rob to provide mechanisms for you to filter the "icky" stuff out of your daily reading. He did (and does) this because he thinks it's fun, and cool, nothing more, nothing less.
[end of love-it-or-leave-it rant]
On leasing -- you are not the target market for a lease. The target market for a lease is the person who would be trading in the car around the time (or even before) he/she finished making the payments to BUY it.
In that case, what is the difference between leasing and owning? Why bother building equity if you are just going to turn around and trade it in? The interest used to be tax-deductible but that loophole is long-gone.
The people who lease like driving new cars where the maintenance is (mostly) taken care of. The car becomes an appliance that you pay to own, and every so often you get a new one when the first one becomes "obselete".
I don't exactly agree with this mentality (nor would you, I'm guessing), but I know lots of people like this, and I can sort of understand it. These people want the hassle out of their lives. Something breaks -- take it to the dealer, get a loaner, the dealer calls you in 3 days and you get your car back. If you own it, you have to either have a cool mechanic or a nice dealership (yah, right!) to diagnose and farm out the work for you to the appropriate places, get you the loaner, etc.
It's all about convenience. Of course, if you enjoy solving problems, probably that's not for you. That's why you are a developer/engineer/techie...