will PS3's GPU be more modern than PS2's GS for its time?

[Panajev2001a]

Edited.

 

[Panajev2001a]

fair enough. to which of those 1999 cpus does the r5900 compare as a 4-5 times slower int performer?

Both Athlon 750 and P3-800.

Cheers
Gubbi

Well, that a CPU more than 2x faster in raw clock-speed, with an out-of-order issue and execute engine, more L1 cache and quite a bit of L2 cache (256 KB) can beat the R5900i is not too bad .

I still agree that a better core would have been more ideal: sacrifice VU0: make it a Macro-mode only VFPU (no VIFs would be needed for example) and build-it on top of the FPU (allowing the CPU to dual-issue with the VFPU isntructions) and use the extra space for an L2 cache for the 5900i core.

What do you think of this idea for the EE ?

 

[darkblu]

fair enough. to which of those 1999 cpus does the r5900 compare as a 4-5 times slower int performer?

Both Athlon 750 and P3-800.

Cheers
Gubbi

Pricing
The 750MHz AMD Athlon processor is priced at $799 in 1,000 unit quantities.

Pricing and Availability
The Pentium III processor 800 MHz in SECC2 packaging is priced at $851 in 1,000-unit quantities; the 750 MHz version is priced at $803. Both speeds are available now with volume ramping in Q1.

again, to which of the potential console-circa-1999 cpus does the r5900 compare as 4-5 times slower in int ops?

 

[akira888]

Panajev:

I wonder if using a more modern R8k/10k (95/97-era, the last chips MTI developed before SGI began its courtship with Itanic) with only VU1 would have provided a better balance between floating point and integer, and of course include a reasonable amount of cache.

Isn't it true (from Sony documents) that VU0 has an average utilization of 10-15 percent?

 

[Vince]

Isn't it true (from Sony documents) that VU0 has an average utilization of 10-15 percent?

SCEE's data from 2003 claimed that there was an average of 2% VU0 usage, with some games using upto 10%. Most games just didn't use VU0 when the data was compiled, although I don't know how this treats macro-mode.

 

[Gubbi]

Pricing
The 750MHz AMD Athlon processor is priced at $799 in 1,000 unit quantities.

Pricing and Availability
The Pentium III processor 800 MHz in SECC2 packaging is priced at $851 in 1,000-unit quantities; the 750 MHz version is priced at $803. Both speeds are available now with volume ramping in Q1.

again, to which of the potential console-circa-1999 cpus does the r5900 compare as 4-5 times slower in int ops?

The pricing of top dog CPUs has nothing to do with cost. I'd wager they cost no more than $75 to produce.

Anyway my point was to show that Sony clearly traded off integer (or general purpose) performance for floating point by selecting a CPU core in the bottom feeding segment and maintaning two VUs instead of sacrificing one VU and have a fast CPU core.

Cheers
Gubbi

 

[London Geezer]

Pricing
The 750MHz AMD Athlon processor is priced at $799 in 1,000 unit quantities.

Pricing and Availability
The Pentium III processor 800 MHz in SECC2 packaging is priced at $851 in 1,000-unit quantities; the 750 MHz version is priced at $803. Both speeds are available now with volume ramping in Q1.

again, to which of the potential console-circa-1999 cpus does the r5900 compare as 4-5 times slower in int ops?

The pricing of top dog CPUs has nothing to do with cost. I'd wager they cost no more than $75 to produce.

Anyway my point was to show that Sony clearly traded off integer (or general purpose) performance for floating point by selecting a CPU core in the bottom feeding segment and maintaning two VUs instead of sacrificing one VU and have a fast CPU core.

Cheers
Gubbi

You're telling me Intel and AMD can make a 1000% profit margin on CPUs?
If that's true, I'll be depressed all day.

 

[Gubbi]

You're telling me Intel and AMD can make a 1000% profit margin on CPUs?
If that's true, I'll be depressed all day.

Well, Average Selling Price (ASP) is much lower than the price of the fastest CPUs, typically $150-$200. Intel's gross profit margin typically hovers around 60%, which is quite awesome.

Cheers
Gubbi

 

[London Geezer]

You're telling me Intel and AMD can make a 1000% profit margin on CPUs?
If that's true, I'll be depressed all day.

Well, Average Selling Price (ASP) is much lower than the price of the fastest CPUs, typically $150-$200. Intel's gross profit margin typically hovers around 60%, which is quite awesome.

Cheers
Gubbi

No no no, i mean, if a CPU costs 75 quid to make, and they sell it for 750 quid, that's a 1000% profit margin... Disgusting.

 

[hey69]

that would be 9times actually not 10
ehem whatever 8)

 

[London Geezer]

that would be 9times actually not 10
ehem whatever 8)

yeah yeah yeah i know.

 

[darkblu]

The pricing of top dog CPUs has nothing to do with cost. I'd wager they cost no more than $75 to produce.

hence nobody said those cpus cost that much to produce. nevertheless i'm pretty sure they costed quite a bit given they hadn't reached any considerable volume ramping. so saying that sony could have produced a p3@800-performer cpu (btw, out of the blue) solely for their ps2...

Anyway my point was to show that Sony clearly traded off integer (or general purpose) performance for floating point by selecting a CPU core in the bottom feeding segment and maintaning two VUs instead of sacrificing one VU and have a fast CPU core.

and my point was to show that the trade off sony made was not '4-5 times inferior int performance for a second vu'.

 

[one]

A Pentium III has 9.5 million transistors while an EE has 10.5 million transistors.

 

[London Geezer]

A Pentium III has 9.5 million transistors while an EE has 10.5 million transistors.

The whole EE has 10.5M? If so, the core, which is what's being discussed here, should be MUCH smaller than that, in the region of less than 3M transistors if my memory serves me right...

 

[Fafalada]

Anyway my point was to show that Sony clearly traded off integer (or general purpose) performance for floating point by selecting a CPU core in the bottom feeding segment and maintaning two VUs instead of sacrificing one VU and have a fast CPU core

Not to dispute the logic of the argument, but can you really call this a tradeoff when it's simply following the same design patterns pretty much every console in recent history used (inlcuding every console of this generation).
Lowend general purpose and integer performance was nothing short of a design "target" for consoles to date, under the (imo mistaken) assumption that that's simply all that is practically required of console apps.

 

[Megadrive1988]

btw, I am amazed at how big my thread has become. I feel kinda humbled

 

[Megadrive1988]

A Pentium III has 9.5 million transistors while an EE has 10.5 million transistors.

The whole EE has 10.5M? If so, the core, which is what's being discussed here, should be MUCH smaller than that, in the region of less than 3M transistors if my memory serves me right...

I think EE was originally 10.5M transistors but was upped to 13 million. could be wrong though.

 

[Panajev2001a]

Isn't it true (from Sony documents) that VU0 has an average utilization of 10-15 percent?

SCEE's data from 2003 claimed that there was an average of 2% VU0 usage, with some games using upto 10%. Most games just didn't use VU0 when the data was compiled, although I don't know how this treats macro-mode.

Macro Mode is not counted in those stats AFAIK.

It has also a problem: if you issue a VU0 instuction in macro-mode you cannot dual issue anything else (I do nto think VU0 COP2 instruction can dual-issue with the R5900i instructions).

 

[Fafalada]

Macro Mode is not counted in those stats AFAIK.

It's not counted, mainly because macro mode is just part of regular CPU pipeline so to speak. Measuring VU0 macro utilization separately from everything else is akin to trying to measure separate MIPS or scalar FPU utilization.

And it's true about Dual issue but since R59k executes in order and GCC optimizer generally sucks, you get abshymal dual issue utilization with or without VU0 coop.

 

[darkblu]

Macro Mode is not counted in those stats AFAIK.

It's not counted, mainly because macro mode is just part of regular CPU pipeline so to speak. Measuring VU0 macro utilization separately from everything else is akin to trying to measure separate MIPS or scalar FPU utilization.

And it's true about Dual issue but since R59k executes in order and GCC optimizer generally sucks, you get abshymal dual issue utilization with or without VU0 coop.

apropos, are the r5900 official int performance figures of ~1.5 ops/clock effectivly sustainable?