UNIX gets the boot! In favor if Lintel Machines
 

http://zdnet.com.com/2100-1103-983898.html

=)

Server Application Error

The server has reached the maximum recovery limit for the application during the processing of your request. Please contact the server administrator for assistance.

From: http://www.racksaver.com/products/clusters.asp

:D

SPARC is not really competative performance wise unless you need massive IO bandwidth, which graphics rendering doesn't. So it's not really a surprise as the Intel chips are faster at this. Seeming as I can't see what the servers they're now using are I can't really comment, but I know Sun's blade servers offer pretty high compute density, 16 650MHz SPARC chips / 16x (2.5GHz??) Intel chips per 3 rack units. I don't think the intel blades are shipping yet though.

SPARC isn't going to get competative until the dual core chips arrive and even then we're still going to be waiting for the 8 core chips before things get really interesting.

pixar looked into offers from Sun, and AMD as well but they chose Intel. :) I think they bought 1000+ Intel Xeons@2.8GHz with 512k L2.

With that many systems I'm surprised N1 didn't tempt them, obviously it's just compute power that was the factor here. Reliability and service wise you can't really fault Sun too much. They need N1 to come to fruition and better processors to win back this kind of business.

Intel has always been #1 in price/performance in these areas. :)

I have serious doubts

Actually the SPARCs are much faster than any x86 chip; the Ultrasparc III, even if it runs at 1-1.1 GHz is still faster than a 3 GHz x86

You mentioned 512 kb L2; this is not everything; L1 is very important, too

Unfortunately, to read L1 you need some 3 clock cycles on an x86; theoretically it should be 1 clock cycle and most design is based on this assumption, but this never happens; Sun worked hard and brought this down to 2 clock cycles;

The UltraSPARC III has 32 bk L1 for instructions and 64 kb for data, both functioning at 2 clock cycles read speed; also, because RISC instruction set needs fewer bytes, it will hold in those 32 kb L1 for instructions, double the amount of instructions than an x86 can hold

Also, if L1 or L2 increases in sise. the velocity of that cache will drop, at a critical point so dramatically, that it wouldn't bring any benefit; currently L1 needs on x86 3 clock cycles to be read, L2 needs 100 clock cycles

As processors get faster, the cache will have to get smaller to have a velocity benefit, so storing more instructions will be difficult for x86, as the mean is 2-3 bytes per instruction for x86

Graphic editing servers used to render those movies need however some explanation: if you're doing rendering for something important like building, plane, shuttle you need exact calculations; if you're doing movies, you can do some aproximations and avoid an expensive computer

However, real application still need those powerfull machines

Now the question of the price: a Sun runs, an Intel will clog a couple of times, their clustering technology is not so advanced as N1, you will need more personnel, you will need more tweaking, more downtime; it won't be as competitive as it seems in the beginning

Kind regards,

discoleo

intel's not good in price/performance. they have the highest-end processors, but AMD will get you "more bang for your buck", still. they just can't compete at all with the really high-end chips.

You are trying to compare a 32-bit chip with a 64-bit chip. The SPARC and the x86 aren't even comparable because they execute instructions so differently wouldn't you agree? It's pointless to keep arguing that "RISC" is better than "CISC" or better than "MIPS".
The only accurate way to compare an UltraSPARC III to say, a Xeon, is with price/performace tests am I correct? According to your beloved SUN Microsystems and Oracle Magazine, Intel on Linux is the most affordable to date just like I said. Don't beleve me? This is all comming from SUN's mouth: http://www.sun.com/smi/Press/sunflas...0030805.5.html SUN would disagree with you right here. ;)

High-end Opterons running in 32-bit mode is being beat by Xeons easily and is way cheaper. But I beleve that Opterons will mop the floor with xeons if they are in 64-bit mode. Athlon 64 sounds even more promising.

Not really. Opteron is an Althlon 64 that can be used in upto 8 processor systems. An Athlon can only be used in dual processor systems (I think it's 2P, less than 8 anyhow), but that's the only difference. SPARC can run on 106P systems with ease and top end chips run at 1.2GHz with 8Mb cache. So if they'd gone for a few big SMI machines from Sun, maybe a cluster of fully loaded Sun fire 15k's they'd be better off.

I can't see a Lintel solution being as easy to manage, reliable or as expandable - cheaper yes, and seeming as I still can't get anything from racksaver.com it seems to me they'll have a lot of trouble.

Heh, I just found it funny that on the article about Pixar buying a bunch of Intel machines there was an ad for AMD Opterons.

I find the Crusoe processor to be very interesting, however. If only some other companies would adopt the method that Transmeta Technologies is using...

Yah, I knew that. I was saying it is promising because it's the Prescott killer. Xbitlabs.com says that it's going to be a close fight because both companies have corrected every single little flaw in their upcoming chips. Athlon 64 has now adopted almost all of the instructions in the Intel chips. Xbitlabs says that they will be evenly priced too. And it makes sense because AMD's current 64-bit chip is god damn expensive!

Oh some wonder for linux!

I was also irritated a little bit about the title: well, linux is still some flavour of unix, so I don't think that unix was bitten out

Also, Linux benefited much from Sun, including from Sun Solaris and Linux clustering support is a basic version of Sun's clustering; well, probably it does not support all the advanced features, but its core clustering is surely much derived from unix <aka sun>

So, I really wondered why they choose Intel + Linux; in a 1000 P cluster, a Sun would be more appropriate; and overall it would be more cheaper: I shall explain this

The price contains not only the machine prize, but the operation costs, downtime cost through lost businesses, maintenance costs and so on

The major problems with high clock cycles is their enormous power consumption and heating: I shall explain this in detail

A basic logic contains 3 instructions: NOT, AND and FANOUT
NOT negates: 1 bit input->1 bit output
AND: 2 bit input->1 bit output
FANOUT: 1 bit input->2 bit output

Any other instruction can be constructed with these 3 basic instructions:

x OR y = NOT <<NOT x> AND <NOT y>>

Now lets see a 1 bit addition:
x+y=<NOT<FANOUT<FANOUT<x> AND FANOUT<y>>>> AND <x OR y>, where one FANOUT<x> generates the x used in the OR operation, FANOUT<y> generates the y used in the OR operation and FANOUT<©©©> generates a bit that can indicate us if the addition generates an overflow or we can use that bit in a multibit addition©

What I wanted to say is, that all instructions are generated with this basic 3 instructions©

What is important:
1© AND converts 2 bits into 1 bit, so 1 bit must be lost as heat
2© FANOUT converts 1 bit into 2 bits, so you must add energy to the system

You see that all complex instructions contain a lot of ANDs and FANOUTs, so they need much energy <current> and convert it to heat© <There is a method of avoiding this through Toffoli operator, but no current chip uses it! A Toffoli has input 3 bits and outputs 3 bits, no net gain or loss, but difficult to track 3 bits>

The consequence will be a powerfull heating, which will increase with clock speed; This was a problem in multi CPU computers, so the need for tremendous cooling systems

Most interseting, Sun developed a smaller ink/manufacturing chip technology, but choose NOT to increase clock cycle, but to reduce power consumtion and heat generation

When I think of 1000 3 GHz CPUs, I firstly think at extream heat, powerfull cooling systems, extrem energy consumption through CPUs and through cooling system©

Is such a system more reiable then something with less heat and less power consumption? I strongly suggest it is NOT the case: more power fluctuations, more damage if a part of cooling system malfunctions

So prize will add as you have to pay for energy, for potential losses and maintenance of this system

Kind regards,

discoleo

You get a few racks of anything and it'll produce a lot of heat. These things are designed with cooling in mind and you also put them in rooms with air-con and very clean isolated power. So all your saying is the fater a chip goes the hotter it gets and you just need better cooling. Great. So they wanted fast processors which they got. They want service and support, they'll undoubtably have that which is probably comparably to what you'll get from Sun only most likly cheaper.

Just like Philly Baby said, they have the cooling covered. Have you seen massive clusters in action? I used to work at 3DO before they went under, I visited their massive racks and they have cooling covered. They've got an air conditioned room, large fans blowing at the systems. I think Pixar has cooling covered.

As for downtime and maintence, that souldn't be a concern because according to Oracle, in some instances, Linux is faster and more reliable than Unix. And when large companies buy massive Linux or unix machines, they have custom coded software written specifically for their application.

And for Xeon chips being unreliable, there is no way to prove it. Intel chips are just as reliable, if not more reliable: http://news.com.com/2100-1001-255296.html?legacy=cnet

And do you know why Intel would be more cost effective vurses AMD and SUN? Hyper-Threading. A dual Xeon with Hyper-Threading on, emulates a 4-way system rather effectively. In tests, the dual Xeon achieved scores almost nearing that of a 4-way Opteron. If you want, I can post the link.

So by spending all that money on 1000 Xeons, they can effectively emulate 2000 CPUs. The way I see it, they saved a hell of a lot of money. And that money that they saved can pay for other expenses, like say, maintenence?

I haven't yet seen a 1000P cluster with 3 GHz CPUs

The power consumption and heat generated must be enormous, well over anything to date :D

Well, it surely will be 2-3 times higher than a Sun and it could be far higher than that

Electricity price will steadily increase in the next 30-50 years, so I'm intersted to know the effect on costs; it probably will force businesses to think more on such facts in the future

What are you thinking about the latest power shortages in the US :D

About the ZDNet Story: no 1 GHz UltrSPARC III is affected by that problem; also important, all UltraSPARC III have a 2 clock cycle L1 cache ¥all other chips have a 3 clock cycles L1, including Sun's older chips; Intel's, AMD's and others; and most research and projects are done assuming a 1 clock cycle L1 cache; how wrong :D ¤

Kind regards,

discoleo

The 600-1500MHz UltraSPARC III is rated at 1.7V Core, and the 2000-3060MHz Xeon is rated at 1.525V Core. So the UltraSPARC consumes more energy. ;) I can post links at your request.

Well, it seems you were not very attentive in your physics class.

My old comp with 5 V CPU consumes some 60 W, well below my currecnt desktop PC, with 3.3 V and 1.5 V consuming 300-400 W.:p

Take it easy, 3 GHz at 1.5 V still uses 2x than anything with 1 GHz at 1.7 V.

Chips in the future will function at even higher frequencies and consequently consume more. It has nothing to do with voltage (or very little, for continous current P=UI, for chips this is a little bit different). Why do you think that even modern desktops consume 5x-6x that of older computers?

If you noticed, I mentioned smaller ink/gate sizes for the latest UltraSPARCs in a previous post, mentioning that Sun decided to reduce power consumption AND NOT INCREASE clock speed.

Now, some economic facts:

1. 70% of US electricity comes from oil/coal
2. 30% of US electricity is produced in nuclear plants (and there are 100-70-30=almost 0 alternatives (like renewable energies in Europe))
3. oil prizes should be between 40-50 cents/barrel and are kept artificially low; when we think that Texas reserves are dissapearing, European reserves in the North Sea are dissapearing, there were no massive investitions in Iraq or in Saudi Arabia (where some 70% of worlds oil is coming), this is becoming even more dramatic
4. electricity prizes are kept artificially low in the US; should be twice current prize and will surely increase if electricity grid will be updated;
5. nuclear energy is NOT competitive; should be at least 10x more expensive, as the waste is paid by public tax, and current waste will still be paid in 10,000 years; when considering all waste that will be produced (even in only, say 100 years), it would ruin any economy (and it surely does) :D

Kind regards,

discoleo.

Well, it seems you were not very attentive in your physics class.

My old comp with 5 V CPU consumes some 60 W, well below my currecnt desktop PC, with 3.3 V and 1.5 V consuming 300-400 W.:p

Take it easy, 3 GHz at 1.5 V still uses 2x than anything with 1 GHz at 1.7 V.

Chips in the future will function at even higher frequencies and consequently consume more. It has nothing to do with voltage (or very little, for continous current P=UI, for chips this is a little bit different). Why do you think that even modern desktops consume 5x-6x that of older computers?

If you noticed, I mentioned smaller ink/gate sizes for the latest UltraSPARCs in a previous post, mentioning that Sun decided to reduce power consumption AND NOT INCREASE clock speed.

Now, some economic facts:

1. 70% of US electricity comes from oil/coal
2. 30% of US electricity is produced in nuclear plants (and there are 100-70-30=almost 0 alternatives (like renewable energies in Europe))
3. oil prizes should be between 40-50 cents/barrel and are kept artificially low; when we think that Texas reserves are dissapearing, European reserves in the North Sea are dissapearing, there were no massive investitions in Iraq or in Saudi Arabia (where some 70% of worlds oil is coming), this is becoming even more dramatic
4. electricity prizes are kept artificially low in the US; should be twice current prize and will surely increase if electricity grid will be updated;
5. nuclear energy is NOT competitive; should be at least 10x more expensive, as the waste is paid by public tax, and current waste will still be paid in 10,000 years; when considering all waste that will be produced (even in only, say 100 years), it would ruin any economy (and it surely does) :D

Kind regards,

discoleo.

Oh, and I forgot to mention something very important.

When clock speed increases, error rate increases, too.

Go to public.itrs.net to view the ITRS 2002 update. Later this year, the ITRS 2003 will become available.

What is ITRS?

It describes the evolution of computers in the next 15 years (and all major chip manufacturers and researchers contribute to the ITRS). It is the ultimate roadmap to miniaturization.

Well, the current update is a 190 pages pdf, very technical, but somwhere it states that future chips, by getting faster, inevitably will increase error rate. I mentioned the Toffoli operator to avoid heat problems. There is one exception where you will not use it: you want that the chip discards any errors and this can occur only through heat. So, even ideal chips will generte errors (and at high speeds a lot of them, because less matter/electrons/photons/whatever are used to store 1 bit) and consequently heat and as faster they get, more heat will be produced: not only through velocity, but through error correction as well.

I know, I got a little bit into detail, but how many manufacturers, resellers discuss such issues with their customers? And my opinion is that this will have dramatic effects in the future, more important than actual speed.

Kind regards,

discoleo.

Nice try SUN, this is why you're far behind in performance. I bet they regret it now, because their top sellers are high clocked x86's. Plus it's not much of a concern even in the future, as long as there's adiquate cooling.CPUs are going to keep getting hotter and hotter. its normal. remember when a heatsink used to be enough? and having a fan was considered excessive? well, times change with the technology. Power consumption for Intel's 3GHz is in the same ball park as SUN's UltraSPARC III. SUN's US3 is rated at 70 watts and Intel's 3GHz is rated at 68 watts (80 max). I can post links at your request.

It is a concern for the future. See the ITRS 2002 Update on //public.itrs.net. Faster chips and electronics are inevitably prone to fail more often (yes, ALL major manufacturer and researchers agree on this, but nothing mentioned in the common press; even current devices fail more often than devices some 2-3 years ago :igor: )

And clock speed is the single most significant parameter influencing heat generation. x86 has other bottlenecks as well. U III is still more faster than the 3 GHz x86 and if considering serial operations it is so much faster, that any x86 won't ever be competitive (even future versions).

Adequate cooling?
1. more energy for cooling process
2. there is some optimal temperature for CPUs; higher or lower temp negatively influences performance
3. likely great fluctuations of temp in time and space, if more heat generated (even if excellent cooling)
4. newer chips (more fast) will be more likely influenced by such temp fluctuations; will also fail more readily (see ITRS)

This is not normal. Performance rapidly decreases and error rate rapidly increases above a critical temperature. This has already been reached for silicon gates. Even future materials won't benefit from excessive heat. And again, components will fail sooner when hotter. Thermal fluctuations become so high above 70-80 centigrade, that most chips will behave eratically above this temperature. HD with 10.000-15.000 rpm tend to heat to 60-70 centigrades and function almost at the limit, too. Everything functions damn close to the limit in this days.

Kind regards,

discoleo.

I fail to see what your saying apart from things are getting hotter and things fail when they get hot. This has always been the case for years and years and years. Heck 33MHz chips used to have heat sinks. Also the optimal temperature that electricity moves at is about -240C.

You're also completly kidding yourself if you think the 1.2Gig USIII's are any match to 3gig P4's. In pure processing terms they are a lot lot slower which is a major reason why Sun struggles at the moment. Why go for a chip that's a lot slower in a system that I'm not going to ever put more than 4 gig of memory in anyway. Even if you really need 5 gig of memory you're still going to find people more willing to go for the faster processor 4 gig ofmemory and faster memory speeds at 800mhz, over a processor that can give them the extra gig they ideally want.

if cooling progresses, this temperature will not be reached - that's the theory that manufacturers seem to be going for, anyway. smaller chip processes and higher clock speeds lead to more heat production - but it seems they're trying to counter it with more effective cooling, to keep a lower, or at least stable, temperature.

Won't even be competitive? Won't even be competitive?? Then tell me, why isn't SUN in the top spot in the top 10 TPC Performance? http://tpc.org/tpcc/results/tpcc_perf_results.asp Sun comes in 9th place barely even making it on the chart and at almost 2 times the price of the number 1 slot! Hell, you can buy twice the processing power for the cost of 1 SUN. And with how many CPUs? 128 SPARCs getting smashed by 64 Itaniums at half the price? You've got to be joking yourself. SUN can't even make it on the price/performance chart, It's all owned by Intel: http://tpc.org/tpcc/results/tpcc_price_perf_results.asp
So is this why SUN has stooped so low as to sell chips made by another company? Heck, AMD is having one hard ass time staying alive but you don't see them selling other company's chips. tpc.org and spec.org is the only place that you can accurately compare 64bits vs 32-bits or company vs company. Show us proof of a US3 smashing Intel.Fritction sounds normal to me. Even US3s have heatsinks on them. I have a US IIi (Ultra 10), massive ass-ed heatsink on it placed directly behind the front case fan. Everyone counters heat with cooling. The IRTS can complain all they want. SUN is the one that can't compete. If you can't beat 'em, join 'em. And that's what SUN, IBM, Compaq, VIA, (Apple thought about joining but pulled out at the last moment.) did.

If You noticed, the Sun computer was from 2001 (running Solaris 8, that's really antics and was not a cluster), that is 2 years old and it still competes with Intel's computers from 2003. The other older computers were all clusters.

I don't know why there aren't any newer Sun computers compared. I didn't do the benchmarks. :D

Still, SPARC is faster than x86. I shall explain this.

In this lesson I will explain:
1. Basic concepts: processor design and the von Neumann machine
2. Pipelining: SPARC vs x86
3. Parallel computing vs seriall computing
4. Some basic SPARC assembler: common vs uncommon
5. Conclusions

Some basic concepts

All modern processors contain a hardware interpreter, interpreting the assembler instructions: the assember instructions are not directly executed. A C-equivalent would be:

while (! HALT){
load instruction;
decode instruction;
load operands;
execute instruction;
save any data;
}

This is called the von Neumann principle or machine (it is really hardware done in the CPU).

Most instructions consume many clock cycles: an x86 usually 10 to 20 clock cycles. So it would be a waste of time to run only 1 instruction at a time.

Pipelines: SPARC vs x86

Beginning with the Pentium, Intel introduced a second pipeline, but this pipeline can not execute or terminate the instruction before the first pipeline is complete. So it virtually does only decode and try to load operands.

The first RISC processors had 3 independent pipelines, which could get out of order. The US III can simultaneusly execute some 4 to 6 different operations (can terminate before the other pipelines, too). SO, an x86 still needs at least 10 clock cycles per instruction, while the US III approaches 1 clock cycle/instruction.

What does this mean?

A 3 GHz x86 executes only 3 million dividet to 10= some (at best) 300,000 to 500,000 instructions per second. A 1.1 GHz US III approaches 1 million instructions (actually early this year a Fujitsu-built SPARC chip broke the 1 million/second barrier).

You see, that a 1.1 GHz US III is still 2 times faster that the 3 GHz x86.

Paralel vs serial computing

Then, one may argue that having 2x the CPUs would resolve the problem.

If we suppose complete paralell computing, than doubling CPUs indeed halfes computing time. If, on the other hand, we have a completely serial process, where every data depends on its predecessors, than 1000 CPUs or 1 million CPUs are as fast as 1 CPU.

In graphic computing, some pixels can be computed in paralel, other depend on previous frames, so there is a serial component, too. This means, that 1000 x86's behave like 500 US III's, but the USs still would be 1.5 times faster than the x86 equivalent.

SPARC assembler

In SPARC assembler there are 32 registers per window (I will later explain what a window is): 8 locals (%l0-%l7), 8 globals (%g0-%g7), 8 input (%i0-%i7) and 8 output (%o0-%o7); the %l, %i and %o are grouped into so called windows, with 8 windows available making a total of 128 registers (outputs from window x overlap with input of window x++, so the number of 128; makes sense, because when you swith program execution like calling a subroutine, the output registers are becoming input registers). By having so many registers, saving registers to stack and then popping it back is reduced to a minimum and execution time is very fast. Also, when calling a subroutine, you don't kneed to push caling arguments and then the routine to read them, you just switch window (is done automatically) and output registers become input registers.

Commonplace

This is where commonplace is taken into account. With 128 registers you are most of the time right (and you don't have those ugly registers of x86). With 8 windows you're most of the time on the win (if there is need for more, the OS is notified and one window is stored on the stack, emptying it; this is a costly approach, but most of the time 8 windows are sufficient; it is a bad coding practice to nest more then 8 subroutines). This is why SPARC has a reduced instruction set: make the common fast; don't let rare things compromise the common ones.

I prefer to call x86 as slowing all things in favour of the slowest parts. That's why x86 needs 10-20 clock cycles per instruction.


Conclusions

1. You can make a CPU faster by increasing clock speed, but in a very inefficient way.
2. A far better approach is to decrease clock cycles per instruction, and this is done through better design. Increasing clock speed is done through smaller gates and better litographic techniques. It has nothing to do with design (so Intel is a very poor chip designer, while principally it is possible to implement the SPARC on 2-3 GHz without significant design workaround; some small workaround is practically always needed).
3. When you have 10 clock cycles per instruction, 3 clock cycles for reading L1 cache are slowing down the system, but not as significant as when you are working close to 1 clock cycle per instructon; when the chip is close to 1 cc/instr, having a 3 cc/L1 would render it actually a 2cc/instr (instead of close to 1cc/instr). So a better L1 (at least 2 cc L1) is needed for performance in this group.
4. In parallel computing, more CPUs will reduce computing time. When a serial component is added, while it still lowers the paralell component, it will never reach the time limit set by a fast CPU (instructions per second relevant for fast) (for purely serial computing, an infinity of CPUs is still not faster than 1 CPU). With 500 US III you can do the job of 1000 3 GHz x86 still some 1.5 times faster.
5. Considering Moore's law, that doubling velocity occurs every 18 month, the x86 still needs 2 years to break the current SPARC limit (now you probably better understand, why the Sun in the benchmark, while 2 years old, was still competitive).

With the 10.000 rpm HDs I meant that they function at the limit, any increase of temperature would damage the stored information. Currently, everything function at this limit. Some other remark: odd numbers in US mean major new design (like US I, US III). Even numbers mean some minor improvements and lithographic advances (like IIi, IVi probably coming).

Kind regards,

discoleo.

Sun's computer isn's clustered. No new entries means they can't compete.Too busy marketing x86 chips over their own.Having a hard time in 9th place, and at almost 6 MILLION dollars more than the top spot. This is obviously not competitave.Theoretical.At what cost? Test setup benchmarks? Need more proof. I know it's damn expensive too.Nope. If SUN thinks that the US3 is that fast, they wouldn't be marketing Intel and AMD products and would definately produce machines that showed the US3 with a clear advantage over x86 like they have always done. And they would definately show proof. But now they can't. They are losing market share at an extremely rapid rate ALL because of Intel vs AMD, and Unix vs Linux.You know as well as I do that this is NOT an accurate comparision. Instruction execution are completely different with CISC/RISC/MIPS. It's like this: You are trying to climb a hill with a bike, you can pedal faster on high gear or you can pedal slower on low gear. The difference is, how much effort is put in? Who has more will to win? AMD is racing Intel because they both want to win the trophy, but SUN knows that it doesn't have to win because he gets money from sponsorship anyways. My point is, competition between Intel and AMD has led to very fast x86 chips. Faster than anything RISC based. (With Proof) The ONLY accurate way of comparison is Price/Performance. Would you even be climbing that hill with a bike if for the same price, you can get a motorcycle?

For single CPU performance (CINT2000 [36]), SUN's SPARC64@1350MHz scored a 776/905 (base/peak), and a Xeon@3GHz got 1242/1294. Itanium 2@1.5GHz is the TOP performer at 1322/1322. See the HUGE performance gap just for a single CPU? The results are all updated every quarter so these are all new. I can post the link at request.

Poor enough to have SUN selling them?You still haven't shown proof, because there isn't any. Even Intel's 32-bit x86's are out performing SPARCs. The proof has been shown, and everyone seems to agree:So the folks at these companies who reviewed and tested The Intel and SUN workstations say that they cost 20% less, and is 3x faster. Links for these quotes at request.

Ugh! I downgraded all chips to my old 3 MHz C64. Well, to err is human. ;)

OK, all the instructions per second must be 3 orders of magnitude higher.

Some other news:
Well, SPARC notebooks; it sounds good! Here is the Tadpole webpage.

And this is from august, too:

Here are the TPC results. Comparisons aren't easy. But, please note the price differences (Price/QphH). Also relevant is clustered vs non-clustered. See the 300, 1000 GB and 3000 GB databases, too.

Some corrections to previous posts:

This is NOT correct. Some materials become superconducting at around 2-3 Kelvin (-273 centigrades); other hot superconductors can be still superconducting at 70-80 K (-200 C), but silicon gates function optimally more near room temperature (or slightly above). The slower parts are clearly the silicon gates and the metal/gate interface (so called coupling constant plays here a major role). So you really don't need -200 C (it downgrades this conduction furher); nevertheless hotter temperatures (around 100 C) degrade performance and reliability in the Si-gates.

Read my previous post on architectures. More important than clock frequency is the number of clock cycles per instruction. This nr is huge for x86 (with mean 10-20), while on SPARC it approaches 1 cc/instruction because of multiple pipelining, better branch guess, L1 cache speed, 128 registers, windowed registers, better design of instructions.

The other great thing is that cost for SPARC processors are likely to drop. They did most of the design improvements in the UltraSPARC and in the US III they did also some important design improvements. So, no need for dramatic design modifications. All they need to do is to adapt the US to higher clock cycles when the right technology is here. This is NOT design. This is the lithographic part (it always needs minor design workarounds, but is mostly done through computer calculations and emulations, this processes beeing already established, so no need for great investments).

Not true. :( Processors will grow in size, too. I strongly reccomend reading at least the introduction from the ITRS 2002 Update (the fisrst 20 pages of the pdf document at public.itrs.net). Because of the size increase, there will always be a temperature gradient across the processor, generating huge t fluctuations. You get a bigger CPU with more heat and less capable of dissipating the heat. At this point, cooling plays no substantial role (its in the chip and as I mention, you cannot precool the chip because of performance drop.

The major computing units are based on SPARC, so it is clearly superior to anything else. OK, prize was a concern. But it has dropped and continues to drop. See the TPC prize/performance comparisons on my previous post and you will agree on this. There are even SPARC notebooks and the prize is somewhere around $3000, this is a good prize. :)

Kind regards,

discoleo.

Comparisons are VERY hard to do in DB that is why I didn't post it even though Intel takes lead in most areas and results are totally dependant on system setup. It's not very accurate for comparing CPU power, SPEC.org is better for that. For example, take a look at 1kGB results, Intel has a total of 3,385QphH@$58, in the same area, SUN has 2,240QphH@$104. Better, more equal comparisons can be drawn from SPEC because they compete for pure CPU performance. As for SUN owning the top spot in 3,000GB Database, every company below it has a submission date of 02, so expect SUN to be at the bottom once 03's come in. Intel still takes the lead in all other results except for those two.;) If you're talking about Sun Fire V480 in the 100GB results, HP owns the top spot.

And just because Intel makes fast, cheap, hot x86 chips doesn't mean it can't make fast and low power consuming CPUs like SPARCs. A great example is the Pentium M, designed by some nerds in Israel. The Pentium M is super low power consuming and low heat emissions CPU, (Already out in notebooks and blade servers) at the same time, it's VERY fast. A 900MHz/1MBL2 will out perform a 1.2+GHz P3-M, and the Pentium 3 is actually better than the P4 believe it or not. The 2GHz Pentium M is coming out soon with a whopping 2MB of L2, This thing will easily compete with a 3GHz P4. The P-M's chip design is so good, that they will bring it to desktops pretty soon, it's already in blade servers. And Intel can make some blazing fast RISC chips too. Faster than ANYTHING to date according to spec.

So are you starting to believe me when I say that Intel makes chips that are fast and cheap? ;)

Cheap, maybe. Fast, not really. Good? Well, I think you're kidding.

There are 2 things which hampered computer development: firstly M$ and second Intel. x86 is so much bad, that only a complete new design will change this. But this will be to expensive, so I don't see it happening.

But good to know that cheap SPARC notebooks are produced. :)

Intel chips are fast:

1. You fail to show proof that SUN has a Clear lead, CPU performancewise. I have shown the proof.

2. I have also shown proof that a large majority of the time, Intel-based systems are cheaper and fasterperformancewise.

3. SUN has absolutely no performance advantage over Intel, because if they did, Industries that depended on the sole performance of CPU power would not be Switching. You saw it youself, the professionals who use SUN Workstations say that Intel is cheaper and most importantly, faster.

4. And to say that you need Microsoft for Intel is pure wrong and biased. Even you know that. You can run more OS's and apps on Intel than you can on anything else. Solaris, Unix, Linux, you name it.

5. To say that SUN's chips are faster than current and future Intel chip is completely false. This whole time, i was comparing how much faster and cheaper the 32-bit Xeon was over any of SUN's products by comparing SUN's OWN BENCHMARKS AGAINST INTELs, and Xeons still come out on top. And don't go forgetting that SUN knows this too.

6. If SUN chips are so fast, then why should they or ANYONE market Intel's chips for them? You saw it in SUN's own words in their press release. Tell me why IBM, SUN, SGI, VIA, Alpha/Compaq, is selling Intel chips over their prized PowerX series, SPARC, Cyrix+, Alpha! Now apple buying massive Intel chips too? Have the industry gone insane?

7. According to SPEC, which is the ONLY place to get accurate CPU processing comparision, Intel's RISC Itanium 2@1.5GHz on is the fastest to date. Leading by more that 200% faster than the fastest SPARC alive. And this is nearly equal in MHz.

If you can prove me wrong, with real benchmarks and proof, feel free to do so. But until then, the Industry, meaning the folks who actually buy SUN's stuff, says that I am right, Intel is king in CPU power. ;)

The new Gemini SPARC chip will consume only 32 Watts, that is well below all x86 products. Here is a link.

Also, Sun focuses on more powerfull supercomputing. Read this. Why isn't Intel represented here. Read also this.

Also, to note:That Japanese system used vector chips, fairly popular in Japan supercomputers. Nothing about x86. x86 is just a disaster. Don't believe anything saying x86 is the fastest or Intel has the fastest things. It just prooves it is incredibly false. If you wish more data on Vector chips I can tell you a heck lot more.

I know this has little to do with SPARC, but by combining Java with SGI's OpenGL, Sun establishes this for a broader audience, so that future SPARCs won't be just for fast DB applications. The world is changing and Sun has recognized that the web is the future. Here is an article about java version of databases (JDO).

The European Space Agency (ESA) has also focused on SPARC processors. Its future missions will run with SPARC chips. See the Gaisler Research homepage, describing the LEON2 SPARC-certified chip.

You may ask, why SPARC for extraterestrial missions?

Simply, because it is fast, reliable (it has a very high fault tolerance; here contributes its low clock speed, too) and low power consumption (again, low clock speed is essential for this). This is a mission using a SPARC chip.

Now, think of the future:
1. Web -> SPARC contributes
2. fast computing -> SPARC will contribute (the defense programe)
3. space missions -> SPARC contributes

Nowhere Intel mentioned.

Kind regards,

discoleo.

Still won't grasp the high-end worksttion or desktop areas, that means that these chips are useless in single or dual mode, meaning it won't be able to compete in small numbers and won't have price advantage either. Either way, this is still an attempt get back what they have lost.Still going to sell intel and AMD products.1. Defense/Military - Why no Intel? This one is TOO obvious. The Military have already BEEN using SUN/UNIX platform for years. That means its imossible to change suppliers. AND most importantly, do you think the military as time to re-train all those troops and tech supports to learn how to use all the new stuff?
2. Web - So does every other company including cisco, inel networks, amd, lucent...
3. NASA already using x86. Still are, and continues to purchase Intel. Same goes for Earth Supercomputer, they've been sung SUN for years.
4. What does this mean? Absolutely nothing. If it is impossible to change supliers, why do it?
5. Nowhere Intel mentioned - What does this have to do with anything? Some articles I post don't list SUN. So what?Since Earth Supercomputer is the only computer in its area, it's not comparable because the applications for these comptuters serve different purpouses.Need comparable proof, and specifications. So I was right, they don't change platforms even if it's better. Look at how old that thing is! It's just what I've been saying all along. Space, Military, Supercomputers will not change suppliers. That is the reason why SUN does not compete and struggles to survive. Because they OWN this kind of market, same monopoly tactics as Microsoft. Why do you think SUN sells x86? Because there is no money to be made in SUN's current market.

Xeon consumes 35 Watts, and Pentium M consumes only12 Watts typical. As always, links at your reuest.

Now back on topic, what yo've just posted have absolutely nothing to do with SUN's curent place, and Intel's CPUs. Still no curent benchmarks that prove SUN is faster than Intel in CPU power.

By the way, I forgot to mention that Pentium M's max comsumption is 24 watts. Well below SUN's Gemini. And at 2GHz, with 2MB L2, it's already out performing P4's with a lot more MHz.

Also, you mentioned supercomputing and space exploration, here are some real facts:

x86 still has the best performance/price ratio in the supercomputer area. take a look: http://aggregate.org/KASY0/press.html and with real benchmarks too: http://www.haveland.com/index.htm?povbench/index.phpThis supercomputer broke the barrier for supercomputer performance/price with x86, again. And this is all based on a project started by NASA. NASA started this project because they wanted fast, and at the same time, cost-effective supercomputing performance. I would also like to add that the Intel Itanium at 1GHz is twice as fast as a 1.8GHz Opteron in the top 5 single processor performance for the same benchmark. ;)

The POVRay Benchmark is a nice benchmark. Only drawback, only x86 compared (and those who wish to run it). I am using POVRay, too, nice renderer by the way.


Gemini is dual core, so way faster and with 32 W less than 2x24 W. Please note that on high clock frequencies CPU consumption is one thing, but you have a lot of power loss over cables, wires, other chips, everywhere where impedance changes and so on (by the way your microwaveoven runs on 2 GHz).

You're not serious.;) Everybody nowadays can use a PC. And tech support: well Linux is supported by Sun, too, so that wouldn't be the primary problem.

UltraSparc has been optimized for the Web (more precisely for running Java, and Java has been optimized for running on SPARC; well, when Sun develops Java and its motto is 'the net is the computer' it is logical). And Java is not only for PCs but is used in mobile phones and many more devices.

The Top 500 Site is good proof. Also this. The Earth Simulator still dominates with 35 thousand Gigaflops being 2 times faster than anything else. By the way it has only 5000 P as compared to 8000 that the 2nd comp has (an alpha). It is still 2 times faster as compared with the Itanium (considering speed/ CPUs). The only computers to be able to compete by speed/CPUs are the Crays (positions 112-118) and the SPARCs from Fujitsu (they have only 60-80 processors, so no perfect comparison).



That mission was from early 2000. Hope you noticed it. Consider planning 2-5 years (you need your processor in the planning, too). The new Leon2 SPARC chip is SPARC v8 compliant. And in space are other rules: low power consumption, extreme fault tolerance (I really mean extreme, considering the background radiation and extreme temperature fluctuations -250 C to +200 C), and absolute reliability. By the way, did you now that one Arianne rocket that was destroyed (because got offcourse) happened because of a wrong float to int transformation (however didn't used SPARC; a similar error can be found in all PII :D ; it's not the old div error).

I wish to say a word on fault tolerance: as clock speed increases, the error rate increases, too. However not linearly. If you read the IRTS 2002 Update, you will know that this is a natural process and will increase more dramatically in the future. In a space mission you must rely on your hardware: if a PC $ucks, you reboot; if a P on a mission makes the slightest error, the mission is lost. So x86 is not better; it simply can't be used, so high fault rate, that when implementing corrections would be so slow and even than would produce faults.

Kind regards,

discoleo.

Military equipment, not standard PCs. They purchase SUN products for a custom solution (e.g. custom OS that runs only on memory, so they don't need a HDD, meaning they wont purchase a CD-ROM or anything standard in any SUN computer or PC or even SUN's OS. They don't purchase what you can buy on their web site. And that's what I mean by impossible to switch. Other than that, FBI, CIA, Government offices all purchase PC in massive bulks.Intel doesn't make dual core designs, only dual CPU designs so the only comparison that you can make is with IBM's CPUs, also featuring dual cores. I can see the strategy behind this, they based it on the cheaper, more successful USIIi (I happen to own one ;)), so adding a second core will be a bit cheaper than 2 seperate CPUs, and way cheaper that a dual core US3 because the cost of manufacturing each core is nearly the same.We both know that has nothing to do with anything here. ;) I am aware of top500.org, but this is not comparable to Intel or even other supercomputers in any way because the systems built are made for spicific applications, and the benchmarks are all theoretical and synthetic fp/int numbers. So this is the poorest of poor in comparison.That CPU was from the early 80's. I posted before that they did not need a fast CPU in space that is why they used this 1980's CPU. That is what I mean by not changing platforms even if it's faster. Uh Yes they can. NASA has been using Intel chips in space for some time now.

1. Hubble Space Telescope is based on Intel's 80386/DF-224 and in 1998 was replaced with a faster 80486.
2. One of NASA's Mars probes uses the Intel 80C188
3. Mission and flight control uses 8086 and newer more up-to-date x86 CPUS.
4. Shuttle diagnostics uses 8086 too.

Some quotes:
"The new advanced computer will be based on the Intel 80486 microchip. This new computer will be 20 times faster, and have six times as much memory, as the current computer on Hubble. In a good example of NASA's goal of "faster, cheaper, better," commercially developed, commonly available equipment was used to build this new computer at a fraction of the price it would cost to build a specialized computer designed specifically for the spaceflight environment. NASA performed a number of mechanical, electrical, radiation and thermal tests to guarantee that the computer would survive the trip to orbit, withstand bombardment by cosmic and solar radiation and work flawlessly in the extreme temperatures of space for the rest of Hubble's life.

As a final check, NASA carried the computer to space in the Space Shuttle for 10 days in 1998 during the STS-95 mission. The computer worked perfectly." Link: http://www.shuttlepresskit.com/STS-103/payload50.htm

Just like I said. Not only have they been using Intel chips, NASA likes to point out that it costs less to use Intel chips, and reliable enough to be used in Hubble, Shuttle and Mars spacecrafts. The quote above also shows NASA's concern with performance/price. That must be the reason why NASA started the Beowulf project.

More links at request.

Something interesting: NASA's Personal Satellite Assistant

http://ficworkproducts.arc.nasa.gov/...psa_sla-02.jpg

Floats around, follows a person around. Has an LCD, a few fans to make it move, powered by a low-power Pentium III Processor on Linux.