6800 / AMD / IBM Speculation?

[AIWGuru]

Originally posted by: Cerb

Originally posted by: AIWGuru

You mean 76% more (unless I missed their 500m transistor GPU somewhere...).

No...
P4northwood-55million transistors
NV40-220million transistors
That's 4 times.
Me thinks you need to go back to math class

P4 Prescott: 125m transistors.

most of which is cache. GPUs have barely any cache because of their SIMD nature. Northwood is a better indication of actually working transistors relevant to the kind of work we're talking about.

 

[Cerb]

Originally posted by: AIWGuru

Originally posted by: Cerb

Originally posted by: AIWGuru

You mean 76% more (unless I missed their 500m transistor GPU somewhere...).

No...
P4northwood-55million transistors
NV40-220million transistors
That's 4 times.
Me thinks you need to go back to math class

P4 Prescott: 125m transistors.

most of which is cache. GPUs have barely any cache because of their SIMD nature. Northwood is a better indication of actually working transistors relevant to the kind of work we're talking about.

CPUs need plenty of cache to perform well, due to their nature, with all the multitasking and decision-making (not to mention bloated apps). If it's got to be apples to oranges anyway, they might as well be apples and oranges from the same season.
This GPU is the current and future graphics technology. The Prescott P4 and Hammer are the current and future CPU technology, and will likely not be replace until around the time of the next major change for graphics (not counting sockets--ugh).

 

[AIWGuru]

Originally posted by: Cerb

Originally posted by: AIWGuru

Originally posted by: Cerb

Originally posted by: AIWGuru

You mean 76% more (unless I missed their 500m transistor GPU somewhere...).

No...
P4northwood-55million transistors
NV40-220million transistors
That's 4 times.
Me thinks you need to go back to math class

P4 Prescott: 125m transistors.

most of which is cache. GPUs have barely any cache because of their SIMD nature. Northwood is a better indication of actually working transistors relevant to the kind of work we're talking about.

CPUs need plenty of cache to perform well, due to their nature, with all the multitasking and decision-making (not to mention bloated apps). If it's got to be apples to oranges anyway, they might as well be apples and oranges from the same season.
This GPU is the current and future graphics technology. The Prescott P4 and Hammer are the current and future CPU technology, and will likely not be replace until around the time of the next major change for graphics (not counting sockets--ugh).

Not for SIMD instructions they don't.
That's the nature of Single Instruction Multiple Data.
They only store one tiny little instruction such as filter, encode, whatever and then tons of data passes through them such as an hour of video and is instantly written back to memory without the cache being used.
Cache is more citical in operations where there is a single small piece of data which has any number of instructions executed on it (eg. SETI)
Anyway, this is irrelevant.
Take a look at what the extra SIMD instruction sets which were added to the x86 set have done for SIMD performance such as encoding (MMX, SSE, SSE2.)
Now apply that increase to a processor with more than 4 times the ammount of active working transistors designed from the ground up for this kind of work, and with 16 pipes instead of one general purpose one.
There's not really any question that this GPU should rape any general purpose CPU at encoding.

 

[Cerb]

Originally posted by: AIWGuru

Originally posted by: Cerb

Originally posted by: AIWGuru

Originally posted by: Cerb

Originally posted by: AIWGuru

You mean 76% more (unless I missed their 500m transistor GPU somewhere...).

No...
P4northwood-55million transistors
NV40-220million transistors
That's 4 times.
Me thinks you need to go back to math class

P4 Prescott: 125m transistors.

most of which is cache. GPUs have barely any cache because of their SIMD nature. Northwood is a better indication of actually working transistors relevant to the kind of work we're talking about.

CPUs need plenty of cache to perform well, due to their nature, with all the multitasking and decision-making (not to mention bloated apps). If it's got to be apples to oranges anyway, they might as well be apples and oranges from the same season.
This GPU is the current and future graphics technology. The Prescott P4 and Hammer are the current and future CPU technology, and will likely not be replace until around the time of the next major change for graphics (not counting sockets--ugh).

Not for SIMD instructions they don't.
That's the nature of Single Instruction Multiple Data.
They only store one tiny little instruction such as filter, encode, whatever and then tons of data passes through them such as an hour of video and is instantly written back to memory without the cache being used.
Cache is more citical in operations where there is a single small piece of data which has any number of instructions executed on it (eg. SETI)
Anyway, this is irrelevant.
Take a look at what the extra SIMD instruction sets which were added to the x86 set have done for SIMD performance such as encoding (MMX, SSE, SSE2.)
Now apply that increase to a processor with more than 4 times the ammount of active working transistors designed from the ground up for this kind of work, and with 16 pipes instead of one general purpose one.
There's not really any question that this GPU should rape any general purpose CPU at encoding.

I still disagree with the 4x (northwood->5x00, Prescott->6x00)...but with the qualifier "for this kind of work," I'd be suprised if it were anywhere near 4x.

Note: I'm disagreewing (edit: that's my elmer fudd voice...really, no typo...) with your statements about software (seems like if it were supported fully right now, it would be a good advertising point for the software companies) and comparisons old tech to new tech, not that a video card can't make a quad-Opteron look like a slug for certain kinds of work.

 

[AIWGuru]

I still disagree with the 4x (northwood->5x00, Prescott->6x00)...but with the qualifier "for this kind of work," I'd be suprised if it were only 4x.

Which is why I used the qualifier "more than 4 times..."

 

[Cerb]

Originally posted by: AIWGuru

I still disagree with the 4x (northwood->5x00, Prescott->6x00)...but with the qualifier "for this kind of work," I'd be suprised if it were only 4x.

Which is why I used the qualifier "more than 4 times..."

And you replied before I got done revising (I can't read the gray in the preview window that well, so don't go into that).

 

[Eug]

There has been news that all of IBM capacity for 90 nm chips is being used to assist AMD?

IBM doesn't make AMD chips. IBM helps AMD develop the process. IBM's 90 nm CPUs (G5 PowerPC 970FX) are almost all going to Apple, which is a problem since it's not enough for Apple, and certainly not enough for Apple + IBM's JS20 eServers (for 64 bit clusters) + 3rd parties.

So yeah, Intel is having 90 nm problems, but so is IBM.

If IBM ever wanted to get back into the chip wars they could buy AMD and buy a big stake in Nvidia...

IBM's high end is POWER, and they've been pushing development of PowerPC for the low end. In fact they've opened up the development of the POWER/PowerPC architecture to 3rd parties. For PowerPC they are pushing 970FX for the low end, but right not heavily partially because they've been having problems ramping the chip, and also because they've got a lot invested in Opteron machines for the low end.

BTW, IBM is collaborating with ATI for future consoles.

IBM has positioned its future in services not hardware.

Hmmm... I guess we'd better tell the IBM hardware guys to pack their bags then...

 

[AIWGuru]

Originally posted by: Eug

There has been news that all of IBM capacity for 90 nm chips is being used to assist AMD?

IBM doesn't make AMD chips. IBM helps AMD develop the process. IBM's 90 nm CPUs (G5 PowerPC 970FX) are almost all going to Apple, which is a problem since it's not enough for Apple, and certainly not enough for Apple + IBM + 3rd parties.

If IBM ever wanted to get back into the chip wars they could buy AMD and buy a big stake in Nvidia...

IBM's high end is POWER, and they've been pushing development of PowerPC for the low end. In fact they've opened up the development of the POWER/PowerPC architecture to 3rd parties. For PowerPC they are pushing 970FX for the upper low end, but right not heavily partially because they've been having problems ramping the chip, and any chips they do make goes directly to Apple, with only a few chips trickling out to their own JS20 eServers (for 64-bit clusters) and to 3rd parties..

IBM has positioned its future in services not hardware.

Hmmm... I guess we'd better tell the IBM hardware guys to pack their bags then...

Lol Eug is right on all counts (even if some had been brought up already )
Especially on the last point: what's that cell processor bit all about anyway?
IBM now has its fingers in more hardware pots than it has in 20 years with the Xbox CPU (read: commodity) PPC being opened up, Apple, cell processor in the playstation 3 and other tiny, tiny devices (read: HUGE commodity) working with Nvidia and AMD....
pioneering and having the most advanced SOI in the world (which they used to bail out AMD)
and their brand new multi billion $ fab.
Yeah, looks like IBM is not interested in hardware...

 

It is IBM who is sitting pretty, especially in the hardware department....allying themselves with AMD for R&D and the like, selling systems with Intel chips (yes the have AMD-based ones too), and making the PPC chips for the G5....they're in a great position. I don't think they'd care to buy AMD, at least not yet.

 

[AIWGuru]

Originally posted by: sm8000
It is IBM who is sitting pretty, especially in the hardware department....allying themselves with AMD for R&D and the like, selling systems with Intel chips (yes the have AMD-based ones too), and making the PPC chips for the G5....they're in a great position. I don't think they'd care to buy AMD, at least not yet.

Don't forget about the cell processor...

 

[Eug]

Originally posted by: AIWGuru
Don't forget about the cell processor...

And whatever processors will be in Xbox 2 and GameCube 2.

 

[Schadenfroh]

Originally posted by: Eug

Originally posted by: AIWGuru
Don't forget about the cell processor...

And whatever processors will be in Xbox 2 and GameCube 2.

IBM is going to have it made console wise, they are providing chips for Xbox2, Gamecube2, and Playstation3. they cant lose

 

Damn! Didn't even know about all those....been too busy lately with school winding down (hopefully for good!).

 

[ingenuiti]

Although hardware is still a large part of IBM, it is positioning its future with services with its "On Demand Strategy." As we all know margins on hardware related products have been decreasing every decade from the 70s, 80s, and 90s. On Demand will offer the full spectrum of services IBM foresees for all competitive businesses. IBM will always be involved with hardware, but it will become less of its core in the future.

If you don't believe me take a look at their annual report or take it from someone who works there.

 

[Eug]

Originally posted by: ingenuiti
Although hardware is still a large part of IBM, it is positioning its future with services with its "On Demand Strategy." As we all know margins on hardware related products have been decreasing every decade from the 70s, 80s, and 90s. On Demand will offer the full spectrum of services IBM foresees for all competitive businesses. IBM will always be involved with hardware, but it will become less of its core in the future.

If you don't believe me take a look at their annual report or take it from someone who works there.

It's a common belief, but one that I don't think that is quite accurate. I think people continue to underestimate the importance of hardware and hardware technology at IBM. There is huge money to be made from services and consulting work, but one of the reasons IBM succeeds is because it's a very diversified company and one that is always ahead of the game. Even as (some) people say that IBM is focusing on services, IBM remains the leading company in the world in granted patents on a yearly basis, on a wide range of things.

And indeed, one of the purposes of IBM's technology is just not to profit off the hardware. As you suggested, they sell the services you talk about based around that hardware. ie. They make super-duper POWER5 big iron servers, and then they sell the OS, software, and most importantly the rollout and support. If they only had the services part that'd be a big problem. eg. If they were hawking Itanium 2 stuff, then they're only one of many potential providers of that hardware and Intel makes the cash on the CPUs. OTOH, if they're marketing POWER5 stuff, they're really the only game in town for support (because they own the technology), and they get some money in the hardware as well.

No, IBM is not a services company. IBM is an everything company.

Anyways, back to CPUs. I hope IBM can correct the problems with their 90 nm process soon. It's interesting to note that Apple put the blame on the delayed Xserve G5 rollout squarely on IBM. They essentially said that everything from all suppliers was already in place, but they just couldn't get the CPUs in quantity. They didn't talk about the Power Macs, but obviously if the low volume Xserves are delayed by IBM, then so are the high volume Power Macs, which are now overdue for an update. IBM has already announced the G5 2.5 GHz 970FX, and that is needed for Apple to keep up with the 3.2 GHz Northwoods out there. Power Mac sales are already dropping, as astute Mac buyers are waiting for the refresh, thought to come 2 months ago.

Interestingly, the 970FX at 2.5 GHz is actually quite a low wattage part. Once the problems are ironed out in the fab, one might expect them to ramp up to 3.0 GHz fairly quickly. I suspect for at least floating point, a 3 GHz G5 might best a 3.6 GHz Prescott, and given that Intel is having 90 nm problems too (and AMD is probably gonna take a while too), it's possible that IBM could take the speed crown with its 970FX if it can get the 3.0 GHz out by fall.

That's a lot of ifs though, and Intel is one mighty rival. Competition is good...

 

[reever]

Originally posted by: AIWGuru

Originally posted by: Cerb

Originally posted by: AIWGuru

You mean 76% more (unless I missed their 500m transistor GPU somewhere...).

No...
P4northwood-55million transistors
NV40-220million transistors
That's 4 times.
Me thinks you need to go back to math class

P4 Prescott: 125m transistors.

most of which is cache. GPUs have barely any cache because of their SIMD nature.

Registers count as cache, and as PS/VS revisions go up, so do the register numbers

 

[AIWGuru]

Originally posted by: reever

Originally posted by: AIWGuru

Originally posted by: Cerb

Originally posted by: AIWGuru

You mean 76% more (unless I missed their 500m transistor GPU somewhere...).

No...
P4northwood-55million transistors
NV40-220million transistors
That's 4 times.
Me thinks you need to go back to math class

P4 Prescott: 125m transistors.

most of which is cache. GPUs have barely any cache because of their SIMD nature.

Registers count as cache, and as PS/VS revisions go up, so do the register numbers

True, but they surely do not add up to 512k worth as in the northwood core still giving NV40 more than 4x the active transistors.

 

[Eug]

Silicon Strategies:

One of the problems is ongoing yield issues within the company's 300-mm fab in East Fishkill, N.Y. For months, IBM has been struggling with yields in the fab, a 130- and 90-nm plant.

"We do see demand, but we need to make the products," Joyce said. "We need to improve our yields in our 300-mm plant. Our yields did see some improvement, but not as fast (as the company had hoped)."

IBM is seeing better-than-expected yields within its 200-mm plant in Burlington, Vt., he said. But profits within the 200-mm fab did not offset ongoing losses within the 300-mm fab and a sudden drop in IP revenues in the first quarter, he said."

 

[rbV5]

I think we are seeing some assumptions concerning the Programmable Video Processor. My understanding is that NV has identified codec support, but hasn't coded it all yet. Its not simply going to be available in X set of drivers, but that driver support will be added over time and as they are programmed into the VP and then they will take over from the installed codecs. They won't be using the VPU and shader support like ATI is currently doing(although they can use the shader pipeline if they wish), it will be handled by the VP and the CPU working in conjunction.

The MPEG Encode/decode diagram I saw at B3D will still require CPU cycles, the VP does alot of the work, it is still passing off work to the CPU as well. I've seen no other diagrams, but I'll assume thats the way it will work. I'd say there is a lot to be discovered on the inner workings yet.

 

[jrphoenix]

Originally posted by: Eug

There has been news that all of IBM capacity for 90 nm chips is being used to assist AMD?

IBM doesn't make AMD chips. IBM helps AMD develop the process. IBM's 90 nm CPUs (G5 PowerPC 970FX) are almost all going to Apple, which is a problem since it's not enough for Apple, and certainly not enough for Apple + IBM's JS20 eServers (for 64 bit clusters) + 3rd parties.

So yeah, Intel is having 90 nm problems, but so is IBM.

If IBM ever wanted to get back into the chip wars they could buy AMD and buy a big stake in Nvidia...

IBM's high end is POWER, and they've been pushing development of PowerPC for the low end. In fact they've opened up the development of the POWER/PowerPC architecture to 3rd parties. For PowerPC they are pushing 970FX for the low end, but right not heavily partially because they've been having problems ramping the chip, and also because they've got a lot invested in Opteron machines for the low end.

BTW, IBM is collaborating with ATI for future consoles.

IBM has positioned its future in services not hardware.

Hmmm... I guess we'd better tell the IBM hardware guys to pack their bags then...

I may have been duped (or not?) by a story by the Inquirer that IBM is using all of their capcacity for 90 nm chips now, by fabing some AMD chips: IBM fabing AMD?. If this is true... it is a very intersting turn of events?

 

[jrphoenix]

Originally posted by: rbV5
I think we are seeing some assumptions concerning the Programmable Video Processor. My understanding is that NV has identified codec support, but hasn't coded it all yet. Its not simply going to be available in X set of drivers, but that driver support will be added over time and as they are programmed into the VP and then they will take over from the installed codecs. They won't be using the VPU and shader support like ATI is currently doing(although they can use the shader pipeline if they wish), it will be handled by the VP and the CPU working in conjunction.

The MPEG Encode/decode diagram I saw at B3D will still require CPU cycles, the VP does alot of the work, it is still passing off work to the CPU as well. I've seen no other diagrams, but I'll assume thats the way it will work. I'd say there is a lot to be discovered on the inner workings yet.

What would be very intersting is that if the GPU can offload enough of the encoding work from the CPU... that AMD / Intel processor are essentially equal in peformance in this field, AMD could have the all around better processor. Encoding / decoding was the primary reason to choose an Intel chip over AMD because AMD was realtively quicker in gaming. Of course things could change... but this could be a big jump in performance for AMD (without them having to do a thing)....

 

[addragyn]

If IBM start fabbing AMD anything there will be press releases from both companies. Nowhere in that "story" does it say anything about any one chip using all of IBM's 90nm capacity.

http://www-306.ibm.com/chips/index.html
Go to the IBM Microelectronics homepage and get the info straight from the horses mouth. What architecture is features prominently? What architecture will the all of the big 3 all be using in their next gaming consoles?

Tom Yager has written on IBM's Fishkill fab. He says, IBM is building the Death Star of the semiconductor industry.

 

[Davegod]

Originally posted by: miniMUNCH

Originally posted by: Regs
The current CEO of AMD is too passionate about his technology to sell off his company. But an hostile take over could be a possibility if the stock owners wanted to jeopardize their own pipe line.

A CEO really can't do that much to stop their company from being sold...if IBM or someone else wants to pay enough for AMD, the board directors has the responsibility to do whats best for stock-holders and take the deal.

Also, the CEO has plenty to profit from the deal...he too gets the benefit of the poison pill which is often really tough to pass up.

true. fwiw though, hostile takeovers are generally significantly more expensive.

Personally I think IBM's current strategy is getting it a lot of fingers in a lot of pies, with good strategic oppertunity and flexibility.

 

[ingenuiti]

I hope you guys realize that IBM's semiconductor business has reported a loss of $150 million. I'm not sure what type of strategic opportunity there is for IBM to take on the risk of buying AMD or Nvidia as the original poster just mentioned.

 

[Sunner]

Originally posted by: ingenuiti
I hope you guys realize that IBM's semiconductor business has reported a loss of $150 million. I'm not sure what type of strategic opportunity there is for IBM to take on the risk of buying AMD or Nvidia as the original poster just mentioned.

Well, for one thing they provide POWER chips for IBM's Regatta line, which is doing better and better, which in turn sells DB2 and Gloabl Services.