[Eug]

HUGE article. See here.

"By way of conclusion, I should enumerate the important information that I don't yet have access to:

Confirmation of the Altivec unit latencies
Confirmation of the identities of the new pipeline stages in the vector execution phase and in the common fetch/decode phase.
L1 and L2 cache latencies

Of the above three items, I'm fairly certain that my speculation on the first two is correct. As for the last item, the cache latencies, this is important information that we'll all just have to wait for.

Regarding the future of the 970, much will depend on how aggressive IBM is in ramping up the clock speed and adding new features. When will we see an on-die memory controller? Will we see a version with more cache? Will they add support for an L3 cache? All of these are questions that only time will answer.

I do hold out some hope that we'll see the 970 scale just north of 2GHz fairly soon, based on the fact that IBM has just announced that its top-end Power4 server is now at 1.7GHz. In my first article, I mentioned the fact that the Power4 is built for reliability, not speed; hence it has thicker gate oxides that spell lower clock speeds. The 970 is meant to be faster and less reliable, so just based on the difference in the gate oxide thickness alone I would expect the top-end 970 to bump up against 2GHz. And when you factor in the 970's much smaller die size, even more MHz can be added to the top-end estimate.

Finally, as I've mentioned a number of times in this article and in other places, the 970 is made for SMP designs. So I look for the top-end 970 machines from both Apple and IBM to be dual-processor right out of the gate. And in fact, I would expect that machines with four processors and more are in the works from both companies.

On the IBM side, I think the 970 will make a great Linux platform, especially for content creation companies that are currently replacing their expensive 64-bit *NIX workstations and rendering machines with commodity x86-based Linux systems. If IBM can keep the prices of its 970-based Linux boxes competitive, they could see their systems gain traction in many of the places that Linux-based PC systems are starting to enter.

Finally, turning once again to Apple's use of the 970, I believe that Apple is poised for a huge overhaul of its hardware line based on this processor and a renewed relationship with IBM. I'm finally convinced that Apple's days of wandering in the wilderness with Motorola are over, and that personal computer users will be able to see the Mac as a real option again in terms of desktop, and not just portable, performance.

As with all things Apple, though, the big question is price. Will Apple drop its margins drastically and sell these machines at a competitive price point in order to increase market share, or will it continue to price itself into the increasingly non-existent luxury/lifestyle computing niche? I'm hoping for the former, because I'd love to give my TiBook and iPod some company with a 970-based PowerMac. If the answer is the latter, though, my Apple products will likely find themselves interoperating with an x86-64 Windows box."

In any case, Hannibal has stated that he believes that the PPC 970 won't be hugely faster than the current G4 clock for clock, despite what many others think. Fortunately, it seems the PPC 970 was designed to ramp up much higher clock speeds.

[JB72]

CNet hqas picked up on the story: http://news.com.com/2100-1023-211925.html?tag=fd_day

[JB72]

Originally posted by Eug:
In any case, Hannibal has stated that he believes that the PPC 970 won't be hugely faster than the current G4 clock for clock, despite what many others think. Fortunately, it seems the PPC 970 was designed to ramp up much higher clock speeds.

It seems this 970 will be even more Altivec heavy than the G4s. The real benefit will be faster bus speeds.

I wonder how hot these suckers will run.

[Eug]

Originally posted by JB72:
I wonder how hot these suckers will run.

42 W typical for 1.8 GHz at 0.13 u. 0.09 u will be less of course. Dunno by how much. 19 W typical for 1.2 GHz. The latter is in 1 GHz G4 territory for heat dissipation. 1.4 GHz 15" AluBook anyone?

BTW, here is Part I of the article.

Originally posted by JB72:
CNet hqas picked up on the story: http://news.com.com/2100-1023-211925.html?tag=fd_day

Heh. "Apple Computer intends to use the 200-MHz PowerPC microprocessors in systems available before the end of 1996."

[Eug]

Hmmm... He keeps saying that Apple is to blame for a bad memory bus design in their mobos.

AFAIK (which isn't much), the blame is with the G4 chip, not Apple.

[Mark Tungston]

Originally posted by Eug:

In any case, Hannibal has stated that he believes that the PPC 970 won't be hugely faster than the current G4 clock for clock, despite what many others think. Fortunately, it seems the PPC 970 was designed to ramp up much higher clock speeds.

i think he is not seeing up to 30-50% speed gains but i think even a 15-20% bump at the same clock speeds with faster and more predictable future roadmap makes this a keeper.

[JB72]

Originally posted by Eug:
Heh. "Apple Computer intends to use the 200-MHz PowerPC microprocessors in systems available before the end of 1996."

Hehe

I'm pessimistic. So as far as the 970 goes, I'll just have to believe it when I see it. If it's good, I may jump onboard when it gets to the .09 micron era.

[Staatkunde]

I think we all have to read this with a bit of pessimism. Motorola has ruined it for any Mac fan. Hopefully Apple did switch to IBM and the 970 is real.

Speed tests versus a G4 do not mean much. Apple needs to start releasing machines in the 2ghz speed and above. Intel has spent millions to convince everyone that MHZ matters. It is hard to overcome that perception. If Apple wants people to "switch" they need to give them a reason. MHZ is something people "think" they understand, thanks to Intel advertising.
-Mark

[Eug]

Even the more educated PC people know that while MHz matters, it only matters to a point. These are the people that are going to buy expensive high-end dual-processor hardware. Also important are the mid-end crowd who want multi-media support, and I'm sure Apple will go after them with an intense marketing campaign. (The low end group with 1.8 GHz Celerons doing just surfing and email is less of a focus for Apple for now I'm thinking, but first things first.)

I agree there is a certain psychology to MHz numbers, but indeed, any significant performance boost in Macs, regardless of the MHz, will have an enormous impact on Apple's bottom line. One should also remember that P4s are not SMP capable at all. I fully expect that Apple will continue to market dual-processor PPC 970 machines, to compete with Intel's single-processor P4s. Two 1.6 GHz PPC 970 chips for instance is still 3.2 GHz. Now couple that with the fact that 1.6 GHz already performs well on its own with regards to SPEC numbers, and Apple (and IBM) has got a huge winner. With SMP-able apps, not only does Apple have comparable MHz, but each MHz runs faster. Not to mention the fact that IBM has stated that a conservative estimate is that the initial iteration of the PPC 970 chips will be available up to 1.8 GHz. A 3.6 GHz (2 x 1.8) PowerMac would be computing monster. And of course, speeds will increase with PPC 970 chips, just like they will with x86 chips.

Furthermore, the MHz-alone-doesn't-matter-as-much-as-before awareness is only going to increase, with Intel's Centrino processors performing so much better than their own P4 Mobility, MHz vs. MHz, while at the same time increasing battery life. Centrino has caused great excitement in the PC laptop world, despite its significantly lower MHz numbers. In addition, AMD 64-bit chips are coming out too, and they will also have much lower MHz numbers, but again, perform admirably.

BTW, IBM unveils new-chip details. Not really new information, but now at least it's straight from the horse's mouth.

[wallinbl]

Originally posted by Eug:
(The low end group with 1.8 GHz Celerons doing just surfing and email is less of a focus for Apple for now I'm thinking, but first things first.)

Umm, my two PCs are 650 & 733MHz. Both are running SQL Server and MySQL. One is running Exchange and serving as a webserver (only a few thousand pages a day - not an extreme load). I develop using .NET on both machines. I run Photoshop on both machines. I have no complaints about performance.

The biggest thing is that I don't play games. What do I need the speed for? I'm about to get a Mac because I now have a DV camcorder and I want to start messing around with making DVDs. If it weren't for that, I'd probably have these two PCs for years to come.

The MHz issue is generally restricted to teenagers and college students wanting to brag about synthetic benchmark scores (honestly, why do you care?). A small number of people actually need the speed for regular work (video, etc).

[Eug]

Originally posted by wallinbl:
Umm, my two PCs are 650 & 733MHz. Both are running SQL Server and MySQL. One is running Exchange and serving as a webserver (only a few thousand pages a day - not an extreme load). I develop using .NET on both machines. I run Photoshop on both machines. I have no complaints about performance.

The biggest thing is that I don't play games. What do I need the speed for? I'm about to get a Mac because I now have a DV camcorder and I want to start messing around with making DVDs. If it weren't for that, I'd probably have these two PCs for years to come.

The MHz issue is generally restricted to teenagers and college students wanting to brag about synthetic benchmark scores (honestly, why do you care?). A small number of people actually need the speed for regular work (video, etc).

I'm talking about people buying new ultra-low end computers.

You cannot buy a new computer (at least around here) at 650 MHz. Low-end computers START at well over 1 GHz, and generally, ultra low-end computers are in the 1.7+ GHz Celeron range. Currently, and in the near future, Apple cannot compete in this ultra-low end market because Apple computers are priced much higher.

Interestingly, mid-end PC computers don't have a hugely higher MHz than the low end ones. They really just have much better quality parts. Apple computers are spec'd (hardware - excluding the CPU - and software) more like mid-end to high-end PCs. Hence, this is where Apple will compete.

[Leonard]

Originally posted by Eug:
In any case, Hannibal has stated that he believes that the PPC 970 won't be hugely faster than the current G4 clock for clock, despite what many others think. Fortunately, it seems the PPC 970 was designed to ramp up much higher clock speeds.

He doesn't say anything of the kind. You're being a little too general in that statement. He says the PowerPC 970 will excel (be better than a G4) at Floating Point and Vector (Altivec) operations but not that great at Integer Operations (probably be better than the G4, but not twice as good). He also states that there are alot of unknowns that could affect the results of his findings. So in my opinion, we won't know what the performance of the 970 is like until machines with it are benchmarked or much more info on the chip is published.

[Superchicken]

I'm kinda surprised this thing won't be comming with a huge cache...

[Leonard]

I was looking for where Hannibal actually quantifies his findings in numbers. Here's where he does it:

"In order to know how the two processors will stack up against each other, you have to know how fast the 970's clock will be when the processor launches and how fast the P4's clock will be during that same timeframe. If the 970 can stay at 2/3 of the P4's clock speed, I would expect it to be very competitive for floating-point and vector code. The ratio will need to be even higher for it to match the P4 on integer code. Ultimately, though, I think we can count on the P4's clock rate to scale much faster than that of the 970 for the following reasons:"

So in otherwords a 970 at 2Ghz will be just as fast as a P4 at 3Ghz in FPU and Altivec operations, in Integer ops it'll be a bit slower. Based on what he knows, but what he doesn't know can change this stats.

[Eug]

Originally posted by Leonard:
He doesn't say anything of the kind. You're being a little too general in that statement.

Actually he did, but not in that article. In fact, I was being rather generous for his assessment of the PPC 970. He actually said the 970 would be slower clock for clock overall (but of course the 970 would have much higher clock speeds, and a superfast bus). He made those comments in a thread over at /. about the Macbidouille.com PPC 970 benches. See here:

"We've had some discussion of these in the Ars Mac forum, and the consensus is that they're bogus. I'm currently wrapping up part II of my 970 article, and I'm pretty certain that these numbers are made up.

Here's how it will break down clock-for-clock:

Floating-point: the 970 will spank the G4e
Integer: The G4e will spank the 970
Vector: it's a tie, even though the 970's Altivec hardware is inferior to that of the G4e. What gives the 970 a boost is Dual-channel DDR400 and a real FSB. If you were to put the G4e in a similar system, it would out perform the 970 clock-for-clock pretty handily.

Anyway, I could elaborate more, but I'd rather work on my article."

But yes, he also did say there are many variables he can't predict.

[Leonard]

While I'm on the subject he made a statement about the 970 and Intel being on the 90nm technology that seems a bit off. He stated "Process: Intel will be the first to bring its 90nm fabs online, and I expect that by the time the 970 comes out the P4 will have migrated to the new process. Once on a 90nm process, the P4's clock speed will continue to climb. The 970, which will debut on a 13um process, is expected to transition to 90nm a bit later."

First, I'm sure the 970 is in production. So as the 970 is out, is the p4 already on 90nm? Also, if I'm not mistaken, IBM already has 90nm fabs online too. They may not be producing 970's on 90nm, but they are producing chips on 90nm. Heck, IBM may even be producing a few 970s on 90nm tech to test out the process. Will Intel actually get the P4 to 90nm technology before IBM gets the 970 there?

[Eug]

Originally posted by Leonard:
While I'm on the subject he made a statement about the 970 and Intel being on the 90nm technology that seems a bit off. He stated "Process: Intel will be the first to bring its 90nm fabs online, and I expect that by the time the 970 comes out the P4 will have migrated to the new process. Once on a 90nm process, the P4's clock speed will continue to climb. The 970, which will debut on a 13um process, is expected to transition to 90nm a bit later."

First, I'm sure the 970 is in production. So as the 970 is out, is the p4 already on 90nm? Also, if I'm not mistaken, IBM already has 90nm fabs online too. They may not be producing 970's on 90nm, but they are producing chips on 90nm. Heck, IBM may even be producing a few 970s on 90nm tech to test out the process. Will Intel actually get the P4 to 90nm technology before IBM gets the 970 there?

IBM has stated that high volume 0.13 u production won't be until the second half of 2003. Whether that's true or not is a different matter though.

Also, I'm no expert, but it seems to me that producing some chips at 90 nm is far from being the same thing as producing a PPC 970 at 90 nm. What chips are being made at 90 nm (if any)?

[Leonard]

Originally posted by Eug:
IBM has stated that high volume 0.13 u production won't be until the second half of 2003. Whether that's true or not is a different matter though.

Also, I'm no expert, but it seems to me that producing some chips at 90 nm is far from being the same thing as producing a PPC 970 at 90 nm. What chips are being made at 90 nm (if any)?

Okay, I'll accept that, we don't have any proof the 970 is in production now, but still, the second half of 2003 is a month and a half away, is Intel producing P4's on 90nm technology before the end of June? Are they producing them now? I don't keep track of the Intel world, is there a P4 on 90nm tech?

[Leonard]

Originally posted by Eug:
Also, I'm no expert, but it seems to me that producing some chips at 90 nm is far from being the same thing as producing a PPC 970 at 90 nm. What chips are being made at 90 nm (if any)?

Oh, yeah, on that chip that's being produced by IBM on 90nm tech. It's only a very small general purpose chip, probably 1000 times less complex than a PowerPC 970. Here's a link to the press release from IBM on that chip http://www-3.ibm.com/chips/news/2002/1216_xilinx.html So yes, it's no PowerPC,
but still, has Intel produced anything on 90nm technology? Have they taped out a P4 yet or something smaller? Anything? I mean I'm not surprised that they would be producing a P4 on 90nm tech, but have they done it yet or are they doing it soon?

[Eug]

Originally posted by Leonard:
Okay, I'll accept that, we don't have any proof the 970 is in production now, but still, the second half of 2003 is a month and a half away, is Intel producing P4's on 90nm technology before the end of June? Are they producing them now? I don't keep track of the Intel world, is there a P4 on 90nm tech?

I think the P4 is slated for Q4 2003 or may be delayed to some time in 2004.

I guess Hannibal feels that PPC 970 won't be shipping until later than many hope.

Perhaps his prediction is simply off, esp. considering that the PPC 970 is supposedly already being sampled to some of IBM's customers.

[TimmyDee51]

Originally posted by Superchic[k]en:
I'm kinda surprised this thing won't be comming with a huge cache...

Where did you get that? If you read the article, Hannibal says that he doesn't know what size of cache it will have. That, and the cache latencies, will determine how the 970 will really stack up.

[Catfish_Man]

Originally posted by Superchic[k]en:
I'm kinda surprised this thing won't be comming with a huge cache...

It's got 1.5X as much L1 cache as the G4, and 2X as much L2 cache. Shipping it with more than that would drastically increase the price.

[jcadam]

Originally posted by Catfish_Man:
It's got 1.5X as much L1 cache as the G4, and 2X as much L2 cache. Shipping it with more than that would drastically increase the price.

right, cache size is another major difference between the Power4 and the 970.

Sometimes I wonder what a computer with an all-SRAM memory would cost.

[DrBoar]

Benchmarks only tell so much. Remember the big difference in scores between G3 with 512K and 1024K caches while the difference IRL was way smaller?
That being said the "weakness" of the integer performance of the 970 is so bad I think. While a G4 has to run at 6 GHz to get the same SPEC200 fp score as a 1.8 Ghz 970 the corresponding G4 speed for the integer parity is a slow poke 3.2 GHz, that is almost as slow as a P4 that only need 2.8 GHz for the same score. So sure the integer will show less of an improvement than the fp...

I am so depressed soon I can replace my G4/400 with something that only correspond to a dual G4 at 3 to 6 GHz, I want my THz CPU now, nobody understand my pain, how much it Herz

[JB72]

Integer: The G4e will spank the 970

That makes me a bit nervous.

[Pismo2001]

Originally posted by Leonard:
Oh, yeah, on that chip that's being produced by IBM on 90nm tech. It's only a very small general purpose chip, probably 1000 times less complex than a PowerPC 970.

It's not a small chip at all, it's an FPGA in the 1 million gate range, which I believe means around 40 million transistors, perhaps a bit less than a 970, but not an order of magnitude less.

FPGA have a very regular architecture which makes much easier to test and locate faults than a complex microprocessor, and are therefore ideal as a testbed for new processes. This would be very good news if they claimed that they had produced some, as long as it is only a tape-out, who knows what it means?

[kupan787]

Floating-point: the 970 will spank the G4e

True.

Integer: The G4e will spank the 970

I don't get this. The 970 scores way (read 3 to 4 times) better than a G4 in specint. Even if you had a therotical G4 at 1.8GHz, the 970 would still score better. I think someone is missing something here...

Vector: it's a tie, even though the 970's Altivec hardware is inferior to that of the G4e. What gives the 970 a boost is Dual-channel DDR400 and a real FSB. If you were to put the G4e in a similar system, it would out perform the 970 clock-for-clock pretty handily.

Hmm... if you were able to stick a G4 in a system with a 900 MHz bus, and dual DDR-400, it woudln't run, so I don't think it would beat a 970

The problem is that G4 could [b]not be made to have dual chanel DDR-400 on a 900MHz bus[b]. Its just not possible. The 7457 (the supposed) was to have an upgraded 200MHz bus, and have true DDR. But thats still nothing comaperd to the 970. So to say "if I had this magical G4 setup that just worked, it would win" is ridiculous.

[EDIT] A few notes I thought I would cross over here from AppleInsdier (incase you don't read over there).

Programmer: Its an interesting article, but remember it is heavily based on opinion and supposition. There are big unknowns in the article and he fully admits not having that information, and that that information could dramatically effect performance (in either direction).

Outsider: Especially regarding L1 & L2 cache latency and VMX performance in which he knows nothing about apart from the very basics.

[JB72]

Originally posted by kupan787:
I don't get this. The 970 scores way (read 3 to 4 times) better than a G4 in specint.

That's true huh? Sometimes Hannibal makes no sense at all.