Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
MacBidouille Rumors Followup
- Thread starter MacRumors
- Start date
- Sort by reaction score
Re: Bus speed?
1) The FSB of the 970 is a 900MHz bus. That bus is 32-bits in each direction (most modern buses are 64-bits bi-directional). I do not believe that this bus is double- or quad-pumped, but I may be wrong about that (IBM's 970 PDF introduction doesn't say if the bus is single-, double-, or quad-pumped). It has an 800MHz effective data rate after overhead (6.4GB/s). I can't see any reasonable way to describe this bus as 200MHz, even if it were a quad-pumped 220MHz(?) bus at heart. 200MHz bus can only refer to the memory bandwidth, matching with a DDR-400 memory implementation.
2) Double-pumped (or double data rate) means that essentially twice as much data is pushed through the bus in one cycle as in a single-pumped bus. However, latency is as high in the "400MHz" DDR bus as in the 200MHz SDR bus (and higher than in a true 400MHz SDR bus). In other words, you're still waiting 1/200M of a second for your data, but you get twice as much of it. This makes a DDR bus twice as effective for burst data as the SDR bus but offers no increase in one-off data reads/writes. Fact is, you're usually writing more than just 32 or 64 bits of data at once, so the latency issue tends to even out and DDR is about twice as fast as SDR.
3) Quad-pumped is an obvious extension of DDR (four times as much data as SDR instead of twice). There is also triple-pumped, although that's not done very often.
1) The FSB of the 970 is a 900MHz bus. That bus is 32-bits in each direction (most modern buses are 64-bits bi-directional). I do not believe that this bus is double- or quad-pumped, but I may be wrong about that (IBM's 970 PDF introduction doesn't say if the bus is single-, double-, or quad-pumped). It has an 800MHz effective data rate after overhead (6.4GB/s). I can't see any reasonable way to describe this bus as 200MHz, even if it were a quad-pumped 220MHz(?) bus at heart. 200MHz bus can only refer to the memory bandwidth, matching with a DDR-400 memory implementation.
2) Double-pumped (or double data rate) means that essentially twice as much data is pushed through the bus in one cycle as in a single-pumped bus. However, latency is as high in the "400MHz" DDR bus as in the 200MHz SDR bus (and higher than in a true 400MHz SDR bus). In other words, you're still waiting 1/200M of a second for your data, but you get twice as much of it. This makes a DDR bus twice as effective for burst data as the SDR bus but offers no increase in one-off data reads/writes. Fact is, you're usually writing more than just 32 or 64 bits of data at once, so the latency issue tends to even out and DDR is about twice as fast as SDR.
3) Quad-pumped is an obvious extension of DDR (four times as much data as SDR instead of twice). There is also triple-pumped, although that's not done very often.
Re: Re: Re: fpnc:
Perhaps you should the news yourself before you post, Itanium 3 has already been introduced, at 1.5 ghz, SpecInt Base is 1250 and SpecFP Base is 2150. Tanglewood, the Itanium being worked on by alpha and intel teams is estimated to be 10x as powerful as Itanium 3 and will feature both Alpha EV8/EV9 and Itanium technologies. It will be out in around 2 years. As I've said before, Alpha is already dead, Compaq disbanded the Alpha design team a while back, Intel has acquired all but one member of the original Alpha team.
SPECint2000:
Power4+ @ 1.7 Ghz: 1113 (base or peak?)
Power4+ @ 1.4 Ghz: 884/910
Itanium 2 @ 1 Ghz: 807/807
Itanium 3 @ 1.5 Ghz: 1250/1250
SPECfp2000:
Power4+ @ 1.7 Ghz: 1699 (base or peak?)
Power4+ @ 1.4 GHZ: 1221/1295
Itanium 2 @ 1.0 Ghz: 1431/1431
Itanium 3 @ 1.5 Ghz: 2150/2150
Perhaps you should the news yourself before you post, Itanium 3 has already been introduced, at 1.5 ghz, SpecInt Base is 1250 and SpecFP Base is 2150. Tanglewood, the Itanium being worked on by alpha and intel teams is estimated to be 10x as powerful as Itanium 3 and will feature both Alpha EV8/EV9 and Itanium technologies. It will be out in around 2 years. As I've said before, Alpha is already dead, Compaq disbanded the Alpha design team a while back, Intel has acquired all but one member of the original Alpha team.
SPECint2000:
Power4+ @ 1.7 Ghz: 1113 (base or peak?)
Power4+ @ 1.4 Ghz: 884/910
Itanium 2 @ 1 Ghz: 807/807
Itanium 3 @ 1.5 Ghz: 1250/1250
SPECfp2000:
Power4+ @ 1.7 Ghz: 1699 (base or peak?)
Power4+ @ 1.4 GHZ: 1221/1295
Itanium 2 @ 1.0 Ghz: 1431/1431
Itanium 3 @ 1.5 Ghz: 2150/2150
Re: Re: Re: Re: fpnc:
Really? Well, you had better run and inform the Intel Corp. of this, because they seem to have forgotten to post information about this "already introduced" on their website. See http://www.intel.com/eBusiness/products/server/processor/index.htm. I guess they just don't want to announce it because they actually might sell some if they did announce it, huh? I mean, geez, if they actually posted information, they might sell as many as 1,000 of these chips or something. That would be about 100 times the total number of Itanic 2 and Itanic 1 processors that they have sold to date, so I don't know if they would be able to ramp up the Itanic production lines fast enough. It could be a real problem - no wonder they don't want to announce that it's already available!
In contrast, the 1.7 Ghz Power4+ has actually been announced and can actually be viewed on IBM's website today (May 9, 2003). See http://www-132.ibm.com/content/home...eServer/pSeries/high_end/pSeries_highend.html
But hey, if we are just making up processors, then I bet you didn't know that the Power17+++ at 158 Ghz was just introduced an hour ago on a .01 micron process! It has 128 cores per chip, simultaneous multithreading that allows for virtual 16 processors per core (i.e. 2048 virtual processors per chip), and still consumes under 100 watts. Each core is rated at a SPECint2000 of 758,231 and SPECfp2000 of 1,255,539. Apparently a single Power17+++ based p690 high end server has more processing power than all the computers in the world combined did as of yesterday.
Look, seriously, I have no doubt that Madison (the 1.5 Ghz Itanic 2 that you are strangely dubbing "Itanium 3", even though Intel hasn't assigned it that name and HP is still referring to it as Itanic 2) will ship some day, presumably before the end of the year. But that day has not yet come.
And as I said earlier, I wouldn't be surprised if Itanic 2 @ 1.5 Ghz had a small but measurable (maybe a little over 10% or 15%) advantage over Power4+ @ 1.7 Ghz in SPECfp2000. However, since Madison is still only single core, whereas Power4+ is dual core, Power4+ will unquestionably be a far more powerful server chip.
Not quite. You are probably thinking of Montecito, the first dual core Itanic. It is the chip that is supposed to ship in 2 years; see http://www.theinquirer.net/?article=7729 The Tanglewood chip will not ship for at least another 3 to 4 years; see http://www.theregister.co.uk/content/61/30411.html And it had better have AT LEAST 10 times the performance of Madison, otherwise it is not going to be even remotely competitive with Power6!
Incidentally, if you want to see some REAL benchmarks (both in the sense that they are not made up numbers and in the sense that it is a real server benchmark, not a single core benchmark like SPECint/SPECfp) on machines that should be shipping concurrently later in the year, go to http://www.tpc.org/tpcc/results/tpcc_perf_results.asp?resulttype=noncluster&version=5. HP has submitted TPC numbers their 64-way 1.5 Ghz Madison system (what you are calling "Itanium 3") that they are planning to release at towards the end of the year. You can see that it puts in a very respectable showing, but is edged out by the Power4+ based p690. The key thing to note, however, as I have told you before, is that the p690 has only 32 processors (64 cores), i.e. HALF the number of the HP Itanic 2 Superdome! Once again, the Power4 packs twice the processing power of Itanic on a per chip basis. This is also apparent in the fact that IBM can ship 8 Power4+ processors (16 cores) in the 4U size p655+ server, whereas HP can only manage to fit 4 Itanic 2 processors (4 cores) in the same 4U space (i.e. two 2U HP rx2600 servers - each holds up to two Itanic 2 processors) . In other words, Power4 based rack mounted servers pack FOUR TIMES the processing power into one rack that Itanic 2 based rack servers do. That, quite frankly, is d*mn impressive.
Now, clearly Itanic has made good strides since the pathetic performance of Itanic 1. However, it is also clear that it has a ways to go before it can catch, let alone surpasses, the POWER architecture in the server space. Furthermore, as an entirely new architecture, there's very little software available for it. Put those two facts together, and you will quickly realize why the number Itanic units that ship each quarter is so tiny.
Really? Well, you had better run and inform the Intel Corp. of this, because they seem to have forgotten to post information about this "already introduced" on their website. See http://www.intel.com/eBusiness/products/server/processor/index.htm. I guess they just don't want to announce it because they actually might sell some if they did announce it, huh? I mean, geez, if they actually posted information, they might sell as many as 1,000 of these chips or something. That would be about 100 times the total number of Itanic 2 and Itanic 1 processors that they have sold to date, so I don't know if they would be able to ramp up the Itanic production lines fast enough. It could be a real problem - no wonder they don't want to announce that it's already available!
In contrast, the 1.7 Ghz Power4+ has actually been announced and can actually be viewed on IBM's website today (May 9, 2003). See http://www-132.ibm.com/content/home...eServer/pSeries/high_end/pSeries_highend.html
But hey, if we are just making up processors, then I bet you didn't know that the Power17+++ at 158 Ghz was just introduced an hour ago on a .01 micron process! It has 128 cores per chip, simultaneous multithreading that allows for virtual 16 processors per core (i.e. 2048 virtual processors per chip), and still consumes under 100 watts. Each core is rated at a SPECint2000 of 758,231 and SPECfp2000 of 1,255,539. Apparently a single Power17+++ based p690 high end server has more processing power than all the computers in the world combined did as of yesterday.
Look, seriously, I have no doubt that Madison (the 1.5 Ghz Itanic 2 that you are strangely dubbing "Itanium 3", even though Intel hasn't assigned it that name and HP is still referring to it as Itanic 2) will ship some day, presumably before the end of the year. But that day has not yet come.
And as I said earlier, I wouldn't be surprised if Itanic 2 @ 1.5 Ghz had a small but measurable (maybe a little over 10% or 15%) advantage over Power4+ @ 1.7 Ghz in SPECfp2000. However, since Madison is still only single core, whereas Power4+ is dual core, Power4+ will unquestionably be a far more powerful server chip.
Not quite. You are probably thinking of Montecito, the first dual core Itanic. It is the chip that is supposed to ship in 2 years; see http://www.theinquirer.net/?article=7729 The Tanglewood chip will not ship for at least another 3 to 4 years; see http://www.theregister.co.uk/content/61/30411.html And it had better have AT LEAST 10 times the performance of Madison, otherwise it is not going to be even remotely competitive with Power6!
Incidentally, if you want to see some REAL benchmarks (both in the sense that they are not made up numbers and in the sense that it is a real server benchmark, not a single core benchmark like SPECint/SPECfp) on machines that should be shipping concurrently later in the year, go to http://www.tpc.org/tpcc/results/tpcc_perf_results.asp?resulttype=noncluster&version=5. HP has submitted TPC numbers their 64-way 1.5 Ghz Madison system (what you are calling "Itanium 3") that they are planning to release at towards the end of the year. You can see that it puts in a very respectable showing, but is edged out by the Power4+ based p690. The key thing to note, however, as I have told you before, is that the p690 has only 32 processors (64 cores), i.e. HALF the number of the HP Itanic 2 Superdome! Once again, the Power4 packs twice the processing power of Itanic on a per chip basis. This is also apparent in the fact that IBM can ship 8 Power4+ processors (16 cores) in the 4U size p655+ server, whereas HP can only manage to fit 4 Itanic 2 processors (4 cores) in the same 4U space (i.e. two 2U HP rx2600 servers - each holds up to two Itanic 2 processors) . In other words, Power4 based rack mounted servers pack FOUR TIMES the processing power into one rack that Itanic 2 based rack servers do. That, quite frankly, is d*mn impressive.
Now, clearly Itanic has made good strides since the pathetic performance of Itanic 1. However, it is also clear that it has a ways to go before it can catch, let alone surpasses, the POWER architecture in the server space. Furthermore, as an entirely new architecture, there's very little software available for it. Put those two facts together, and you will quickly realize why the number Itanic units that ship each quarter is so tiny.
Re: Re: Bus speed?
It is 450 MHz double pumped. I don't remember the exact references, but I have read that in several places (real articles, not rumor articles). Anyway, it would have to be double pumped in order to get 6.4 GB/s of bandwidth (keep in mind that it is 64 bits wide...actually, 2x32 bits, as it is 32 bits each direction).
It is 450 MHz double pumped. I don't remember the exact references, but I have read that in several places (real articles, not rumor articles). Anyway, it would have to be double pumped in order to get 6.4 GB/s of bandwidth (keep in mind that it is 64 bits wide...actually, 2x32 bits, as it is 32 bits each direction).
Re: ok kai....you da man! so there!
http://www.intel.com/products/serve..._mp/index.htm?iid=ipp_srvr_proc_xeon+xeon_mp&
And some one said it was not possible to build a dual Pentium4 system. In fact you need to take Xeon CPUs, but there is virtually no difference between a 3.06 GHz Pentium 4 and a Xeon (excepted price):
http://www.intel.com/products/serve...ex.htm?iid=ipp_srvr_proc_xeonmp+info_xeon512&
Some could argue that running a dual processor PowerMac against a single processor Pentium 4 is not fair. A dual Xeon 3.06 GHz XP box would cost more than $4,000, but dual 2.8 or 2.66 GHz systems could land in the same ballpark as high-end PowerMacs...
Intel Xeon MPs support up to 2 MiB of L3 cache (this is a server oriented CPU):
http://www.intel.com/products/serve..._mp/index.htm?iid=ipp_srvr_proc_xeon+xeon_mp&
And some one said it was not possible to build a dual Pentium4 system. In fact you need to take Xeon CPUs, but there is virtually no difference between a 3.06 GHz Pentium 4 and a Xeon (excepted price):
http://www.intel.com/products/serve...ex.htm?iid=ipp_srvr_proc_xeonmp+info_xeon512&
Some could argue that running a dual processor PowerMac against a single processor Pentium 4 is not fair. A dual Xeon 3.06 GHz XP box would cost more than $4,000, but dual 2.8 or 2.66 GHz systems could land in the same ballpark as high-end PowerMacs...
Re: Re: Re: fpnc:
It certainly will, it's not available yet. IBM actually only announced those new servers, they will ship by the end of July.
Re: Bus speed?
Ars Technica said it would be a 450 MHz DDR bus.
Have you noticed that the IBM document states "up to 900 MHz bit rate"; Apple could stay with a more conservative 800 MHz (400 MHz DDR), especially if they intend to introduce 1.4, 1.6 and 1.8 GHz machines (respectively x3.5, x4 and x4.5 a 400 MHz bus speed).
http://www-3.ibm.com/chips/techlib/...2AE087256C5200611780/$file/PPC970_MPF2002.pdf
Ars Technica said it would be a 450 MHz DDR bus.
Have you noticed that the IBM document states "up to 900 MHz bit rate"; Apple could stay with a more conservative 800 MHz (400 MHz DDR), especially if they intend to introduce 1.4, 1.6 and 1.8 GHz machines (respectively x3.5, x4 and x4.5 a 400 MHz bus speed).
Busses Confusion
Earlier today jettredmont explained to me that the memory bus (between the CPU and the DDR400 RAM) is 200MHz, per the MacBidouille rumor. The FSB (between the CPU and the system controller) has been announced at 900MHz for the 1.8GHz CPU, and is the one that provides the 6.4GBps (not 6.4Gbps, right?) throughput.
That being true, I think we need to be careful about which we are addressing, or those of us who are trying to understand, will get truly lost.
Earlier today jettredmont explained to me that the memory bus (between the CPU and the DDR400 RAM) is 200MHz, per the MacBidouille rumor. The FSB (between the CPU and the system controller) has been announced at 900MHz for the 1.8GHz CPU, and is the one that provides the 6.4GBps (not 6.4Gbps, right?) throughput.
That being true, I think we need to be careful about which we are addressing, or those of us who are trying to understand, will get truly lost.
Re: Busses Confusion
The front buses lead to the system controller (what the PC users call the Northbridge), and one of the features of this controller is to handle memory access. The traditional AGP/PCI bus could be replaced by an HyperTransport link with AGP and PCI-X tunnels.
Take a look at the illustration made for the Xserve:
http://www.apple.com/xserve/performance.html
The main difference is that two PowerPC 970 will not share the same bus (since its a point to point link), and there are no backside L3 caches.
The slow 64 bits 167 MHz bidirectional bus from the G4 machines should become two 32 bits 800/900 MHz unidirectional buses on the new Mach64 mobo, on a dual 970 system you would have four of these buses !
Not exactly, since the PowerPC 970 has no build-in memory controller (compared to the AMD Opteron), it does all it's input/output through its front buses.
The front buses lead to the system controller (what the PC users call the Northbridge), and one of the features of this controller is to handle memory access. The traditional AGP/PCI bus could be replaced by an HyperTransport link with AGP and PCI-X tunnels.
Take a look at the illustration made for the Xserve:
http://www.apple.com/xserve/performance.html
The main difference is that two PowerPC 970 will not share the same bus (since its a point to point link), and there are no backside L3 caches.
The slow 64 bits 167 MHz bidirectional bus from the G4 machines should become two 32 bits 800/900 MHz unidirectional buses on the new Mach64 mobo, on a dual 970 system you would have four of these buses !
Code:
PPC970 PPC970
|| 400 / 450 ||
\\ MHz DDR //
\\ //
Controller
/ | \
200 / | HT \
MHz DDR | AGP |
RAM PCI-X Ethernet, ATA...Re: Re: Re: Re: Re: fpnc:
Just moved from Wisconsin to New Jersey so I was unable to make a timely response to this post, as I've clearly stated before, read the news before you post, HP already HAS a 64 way 1.5 ghz Itanium server out, now tell me honestly, if a 1.5 ghz Itanium wasn't even introduced than exactly HOW can it already be in a HP server. And by the way it IS Itanium 3, a full upgrade cycle has passed since Itanium 2 and the naming schemes used by intel hasn't changed. Also, in case you haven't noticed Power4 has been pretty much been on the behind Alpha and Itanium solutions for much of the past year, now that I think about it, ever since the Alpha 21364 was introduced. The numbers are their, until now, Power architecture was quite clearly surpassed by Alpha and Itanium architecture.
However, you are correct, that with the release of dual core Power4's, IBM has the most powerful server chip, sorry my information is a bit old and rusty, until Intel releases dual core Itaniums, it looks like IBM will remain uncontested.
Just moved from Wisconsin to New Jersey so I was unable to make a timely response to this post, as I've clearly stated before, read the news before you post, HP already HAS a 64 way 1.5 ghz Itanium server out, now tell me honestly, if a 1.5 ghz Itanium wasn't even introduced than exactly HOW can it already be in a HP server. And by the way it IS Itanium 3, a full upgrade cycle has passed since Itanium 2 and the naming schemes used by intel hasn't changed. Also, in case you haven't noticed Power4 has been pretty much been on the behind Alpha and Itanium solutions for much of the past year, now that I think about it, ever since the Alpha 21364 was introduced. The numbers are their, until now, Power architecture was quite clearly surpassed by Alpha and Itanium architecture.
However, you are correct, that with the release of dual core Power4's, IBM has the most powerful server chip, sorry my information is a bit old and rusty, until Intel releases dual core Itaniums, it looks like IBM will remain uncontested.
Another thing
Another thing to note generally is the price, the IBM P-690 turbo (Power4 1.7 ghz) server costs 7.6 million dollars and performs 680,613 transactions per minute, the HP Superdome (Itanium 3 1.5 ghz) server costs 6.4 million dollars and performs 658,277 transactions per minute. Your paying nearly 1.2 million dollars more for a 3% increase in performance.
Another thing to note generally is the price, the IBM P-690 turbo (Power4 1.7 ghz) server costs 7.6 million dollars and performs 680,613 transactions per minute, the HP Superdome (Itanium 3 1.5 ghz) server costs 6.4 million dollars and performs 658,277 transactions per minute. Your paying nearly 1.2 million dollars more for a 3% increase in performance.
Re: Re: Another thing
Hell, save yourself even more, go buy a Compaq. I mean, what's the performance difference you're paying for? There aren't any other factors that go into a computer's worth, after all
Yeah, no kidding. with that kind of logic just go buy a PC.
Hell, save yourself even more, go buy a Compaq. I mean, what's the performance difference you're paying for? There aren't any other factors that go into a computer's worth, after all
And, more importantly...
you can actually GET a p690 with the tested config.
on the hp side, they are not yet shipping 64way itanium systems (and, UNISYS will be first when it happens (they may already have- HP has not.))- they are only in the testing stage with 64way.
now, the SGI altira (sp?) HAS shipped with 64 itaniums, BUT running linux--- NOT windows. this looks to be an impressive machine-- again, VERY high priced.
you can actually GET a p690 with the tested config.
on the hp side, they are not yet shipping 64way itanium systems (and, UNISYS will be first when it happens (they may already have- HP has not.))- they are only in the testing stage with 64way.
now, the SGI altira (sp?) HAS shipped with 64 itaniums, BUT running linux--- NOT windows. this looks to be an impressive machine-- again, VERY high priced.
Re: Re: Another thing
Unfortunately I really don't know, both machines are equipped to be competitive to each other, they both have 512 GB of memory, comparable disk controllers, storage, terminals, server and client software, the Power4 clients used by the P690 ARE expensive but they helped create the TPC score and don't change the overall picture. Remember, we're talking about multimillion dollar servers meant to run 24 hours a day, not desktops, quality and stability are not problems here. If someone can ascertain the price of a 1.7 ghz Power4 I might be able to solve the price difference. Also, Anonmac, what exactly is the cost of the SGI Altix?
Unfortunately I really don't know, both machines are equipped to be competitive to each other, they both have 512 GB of memory, comparable disk controllers, storage, terminals, server and client software, the Power4 clients used by the P690 ARE expensive but they helped create the TPC score and don't change the overall picture. Remember, we're talking about multimillion dollar servers meant to run 24 hours a day, not desktops, quality and stability are not problems here. If someone can ascertain the price of a 1.7 ghz Power4 I might be able to solve the price difference. Also, Anonmac, what exactly is the cost of the SGI Altix?
According to Macbidouille today, Apple have officially confirmed to some customers that the G4 will be around for another three months.....
New TPC-C submission
it seems that fine-tuning was all that was needed for HP's superdome to reclaim the TPC-C crown, the tuned Superdome performs 48,825 more transactions/minute than the original Superdome and 26,489 more transaction/min than the IBM P690 turbo.
Tuned Superdome (1.5 ghz Madison)
unit price: $6,453,433
transactions/minute: 707,102
price/transaction: $9.13/transaction
IBM P-690 Turbo (1.7 ghz Power4++)
unit price: $7,574,961
transactions/minute: 680,613
price/transaction: $11.13/transaction
Original Superdome (1.5 ghz Madison)
unit price: $6,453,432
transactions/minute: 648,277
price/transaction: $9.80/ transaction
it seems that fine-tuning was all that was needed for HP's superdome to reclaim the TPC-C crown, the tuned Superdome performs 48,825 more transactions/minute than the original Superdome and 26,489 more transaction/min than the IBM P690 turbo.
Tuned Superdome (1.5 ghz Madison)
unit price: $6,453,433
transactions/minute: 707,102
price/transaction: $9.13/transaction
IBM P-690 Turbo (1.7 ghz Power4++)
unit price: $7,574,961
transactions/minute: 680,613
price/transaction: $11.13/transaction
Original Superdome (1.5 ghz Madison)
unit price: $6,453,432
transactions/minute: 648,277
price/transaction: $9.80/ transaction
Re: tpc code...
Hmm, did I say that, no, since we were on the subject of servers, and the top TPC scores in particular, I brought up the newest tpc submission just to provide the most up to date information. Now if you for some strange reason don't like this news, than just don't read the post.
Hmm, did I say that, no, since we were on the subject of servers, and the top TPC scores in particular, I brought up the newest tpc submission just to provide the most up to date information. Now if you for some strange reason don't like this news, than just don't read the post.