Speculation: Xenon cpu 360 gflops?

[bbot]

The consensus on the gflops rating of the Xenon cpu seems to be about 90gflops. This is based on the assumption that a single Xenon cpu core will issue two instructions per cycle, which means the Xenon cpu issues six instructions per cycle, and 18 billion ips at 3 ghz. But in reality Xenon dev kits are using two PPC970's. PPC970 issues eight instructions per cycle. Together, they are issuing 16 instructions per cycle and 32 billion ips at 2 ghz. I find it strange that for the final version of the cpu, MS would choose a less powerful chip. So let's consider the possibility that a single core for the Xenon cpu will issue eight instructions per cycle. Estimates of gflops rating would have to be increased fourfold. So that would mean instead of an aggregate gflops rating of 90 gflops, that number would have to be revised to 360 gflops.

360. Haven't we heard that before? Yes, in the name Xbox 360.

 

[Edge]

You making the assumption that all the execution units are SIMD floating point units, and they are not. You have integer units, load/store units, floating point units, and VMX units.

Note, that the 80 GFLOP estimated rating for Xenon's CPU is what the VMX units can achieve, and that's seperate from the standard integer or floating point units. The VMX processor is a specialized co-processor, similar to Intel's SSE units.

You can't get floating point operations done on integer or load/store units, and the regular floating point units, are not able to do SIMD calculations.

 

[Brimstone]

Speculation

If the X-Box 360 CPU is based off of the Broadband Processor Architechture like CELL, the lower performance level is a result of design goals. Microsoft is intrested in portability between PC and XB-360 games so that is certainly an major influencer.

 

[Deadly Towers]

I would guess the differences between the CPUs has more to do with the graphics chips (the nvidia chip could be rasterizer only)

 

[a688]

Speculation

If the X-Box 360 CPU is based off of the Broadband Processor Architechture like CELL, the lower performance level is a result of design goals. Microsoft is intrested in portability between PC and XB-360 games so that is certainly an major influencer.

What is "Broadband Processor Architechture"? I doubt Microsoft cores are that similar to cell at all besides being multicore.

 

[bbot]

My guess is the Xenon cpu core is based on POWER4 or POWER5 architecture.

 

[Brimstone]

Speculation

If the X-Box 360 CPU is based off of the Broadband Processor Architechture like CELL, the lower performance level is a result of design goals. Microsoft is intrested in portability between PC and XB-360 games so that is certainly an major influencer.

What is "Broadband Processor Architechture"? I doubt Microsoft cores are that similar to cell at all besides being multicore.

Titianio pointed this report out in another thread.


Microprocessor Report - Cell Moves into the Limelight

 

[a688]

^^ I was trying to point out that "Broadband Processor Architechture" is a crappy buzz word.

You making the assumption that all the execution units are SIMD floating point units, and they are not. You have integer units, load/store units, floating point units, and VMX units.

Note, that the 80 GFLOP estimated rating for Xenon's CPU is what the VMX units can achieve, and that's seperate from the standard integer or floating point units. The VMX processor is a specialized co-processor, similar to Intel's SSE units.

You can't get floating point operations done on integer or load/store units, and the regular floating point units, are not able to do SIMD calculations.

Are you trying to say that "regular" floating point units can't be added to the flop rating? Also check out the 360Gigaflops the finaly Blue Gene computer is supposed to be capable of. OMG, xbox is the blue gene supercomputer.

 

[aaronspink]

What is "Broadband Processor Architechture"? I doubt Microsoft cores are that similar to cell at all besides being multicore.

its a meaningless buzzword. kinda like "media processor", etc.

Aaron Spink
speaking for myself inc.

 

[Titanio]

Are you trying to say that "regular" floating point units can't be added to the flop rating?

The 90Gflop number includes a FPU.

 

[Shifty Geezer]

I'm sure we've been here a dozen times before! As I understand it, based on articles, rumours, devs blogs etc, it sounds like Xenon's cores are dual issue, in-order processors. That is, as described for the 90 GFlop number, 2 instructions per core.

Now bbot says...

the Xenon cpu issues six instructions per cycle, and 18 billion ips at 3 ghz. But in reality Xenon dev kits are using two PPC970's. PPC970 issues eight instructions per cycle. Together, they are issuing 16 instructions per cycle and 32 billion ips at 2 ghz. I find it strange that for the final version of the cpu, MS would choose a less powerful chip.

Firstly it's said Xenon CPU will be 90 GFlops - not less than the dual PPC of the dev kits. Secondly, pehaps the reason they've provided different tech in the kits is because the Xenon CPU isn't at manufacturing yet, and they haven't got any? The Dev kits have been out a long time, using available components. And thirdly, wouldn't a dev kit have more power than the target console anyway, to aid faster development, run debugging etc. on executables, and so forth?

I don't see more power in dev kits meaning underestimation of final performance myself.

 

[Titanio]

Are we confusing instructions issued with flops performed? I always thought the 2flops/cycle for the core itself (minus the VMX unit) simply came from the FPU unit being able to perform 2flops/cycle, not that the core could issue two instructions per cycle. Or am I the one getting confused?

 

[bbot]

The dev kits have two PPC970 running at 2ghz. So the total gflops should be 160 gflops, which is more than the final cpu configuration. And then there's the matter of PPC970 being out-of-order.

 

[Crazyace]

970 has 2 scalar FPU's and 1 VMX .. so theoretical peak would be 2x2 + 1x8 = 12 flops/cycle, so 2 PPC970's at 2GHz would be 48GFlops

(I dont know where you get 160GFlops from)

 

[aaaaa00]

Also 970FX is 5 dispatch into the out-of-order queue, 8 issue out of the out-of-order-queue into the pipelines -- so really the max throughput per clock is 5 instructions on the average, and that's assuming good ILP. (If I read it right.)

(http://www-306.ibm.com/chips/techlib/techlib.nsf/techdocs/AE818B5D1DBB02EC87256DDE00007821/$file/970FX_user_manual_v1.41.pdf)

 

[Legend]

But J Allard Mention that Xenon Will “ deliver a teraflopâ€￾ in his GDC speech.

 

[bbot]

On second thought, most likely MS won't be using three G5 cores. Three G5 cores running at 3ghz would run too hot.

 

[Phil]

also, dev kits cost a magnitue more than consoles and aren't sold at a loss... powerful dev kits can also help the development process.

 

[Fox5]

On second thought, most likely MS won't be using three G5 cores. Three G5 cores running at 3ghz would run too hot.

Hmm, I thought the rumor was that the cores would be all new designs(at least not seen in a mac), but similar in performance to G4 cores.

 

[Megadrive1988]

this thread has me CONFUZZLED

most widely expected estimate for Xenon CPU system:
84 to 90 to 100 to slightly over 100 gflops


highend hyped estimate:
a teraflop (and yes, for the CPU alone according to certain statements which i do not think are slip of the tongue.)


bbot's guess which is now the middle gound estimate:
360 gflops



I'm going to stay with ~100 gflops. thats a reasonable leap from ~3 gflops of the Intel CPU in Xbox