Nekochan Net

Except for the "coolness" factor, it wouldn't make much sense. A 195MHz R10K is already pushing the limits of the Indigo2 architecture.
Go to Ian mapleson's site and compare the performance of the R10K CPU in Indigo2, Octane, PowerChallenge and Origin systems.

What's the point of putting in a faster CPU if it's waiting for memory all the time, and the disk system is narrow SCSI ?

How fast are the RM7900 CPU? How much performances can we expect over the 600Mhz, I just want to know if it's worth asking SGI over and over again to gain only a small 10% speed burst.

Is it legal to reverse-engineer a PROM? If the PROM doesn't boot somehow, is it possible to 'reflash' it using a PROM programmer, where is it physically located?

Thanks

I'm sure I'll be corrected if I'm wrong because I'm too lazy to hit the PMC Sierra website.
But I think the RM7900 is basically a 900 MHz version of the RM7K. So it should still have the 256KB L2 cache on-die?

I believe Chicago Joe already posted benchmarks somewhere comparing the RM7K versus the R12K and it was actually better. My guess is another 50% clock speed increase would give somewhere near 50% speedup in most cases where the CPU is the limiting factor, as I doubt bandwidth is much of a limitation IRL for the CPU.

Personally I was surprised how good the RM5200/300 is compared to the R12K/300, now that I have both. I'm told the R10K/R12K caches are not effective in the O2. The R12K brute forces its way about 10-30% faster in CPU intensive stuff, in my experience, so I'm sure the RM7K would be a greater force to be reckoned with. Fewer ALUs, but double the MHz, plus full cache efficiency *and* on-die L2. That and much less power consumption. My R12K/300 gets hot within minutes.

The impact of the on-die L2 cache can be large in some cases, I'm sure. My closest experience was comparing my K6-2 (512KB off-die L2 only) to my K6-3+ (256 KB on-die L2 only). The CPUs didn't change too much, aside from extra write combining and a few tweaks. And the K6-3+ is even clocked 10% lower. But often, for moderately sized tasks, it could be nearly twice as fast.

The advantage of the on-die cache disappears completely on large tasks, unsurprisingly. I'm willing to believe the RM7K has similar behavior. Smallish operations that fit inside the L2 probably run much faster on RM7K than RM5200, clock for clock. Largish operations that don't fit in on-die cache will still run faster than RM5200 mainly because of the higher clock rate.

At some point, we're getting hit by the memory bus bandwidth anyway... which is why we won't scale up linearly.

also:

Is it legal to reverse-engineer a PROM? If the PROM doesn't boot somehow, is it possible to 'reflash' it using a PROM programmer, where is it physically located?

Maybe. Most likely on O2 with the UMA memory doing 3D graphics at the same time.
But for all-out CPU activity we may luck out, especially for smallish things that fit mostly in cache.
I was also a bit surprised to learn my K6-x systems rarely saturate their 800 MB/sec FSB, (don't ask how I found out) so odds are they could scale up in MHz, too.

Beyond certain thresholds bandwidth isn't always a limit. For some architectures--like K8--it doesn't buy much of anything. For example: http://www.anandtech.com/IT/showdoc.aspx?i=2291&p=19 Generally the trend is to be core-limited which is why everyone can get away with dual cores and not be bandwidth limited. The question is whether/how often the O2 will still be core limited at 900 MHz.

What is the FSB on the O2 anyway?

One of my concerns is:

If we hack the boot prom, then SGI would have no problem breaking our hack with the next IRIX release*. Just in fear of seeing a company rising from the nekochan community selling commercial 900Mhz upgrades for O2, or rebranded O2 with 900Mhz cpus for peanuts. SGI already has a hard time selling its fuel, if one comes up with a $1000 O2 clocked at 900Mhz that might sell really good, or not... due to the lack of applications for the SGI/MIPS.

(*) Offer a new prom with the next irix release and perform a check that the prom is current from the OS during boot time. We already saw SGI doing these tricks with the origin 200 framebuffer on gigachannel with 6.5.23.

SGI doesn't care about Fuel sales... and they probably no longer have the MIPS/IRIX engineering know-how to even add such a block.

CPU SysAD bus is similar to Octane... 90 - 100 MHz to the CPU.

I believe the RAM thruput is even greater than that on Octane, but of course it's used for graphics too, so the real world performance numbers are lower.

[on a mostly unrelated note]

Well here's a question... if you could do that, rebuild O2s at 900MHz, how much would you charge for one, taking into consideration the cost of upping the mem to at least 512mb, and at least one 18gig harddrive, probably an AV module, and the upgrade to the processor itself... along with the time invested in... putting software on it.

(Not much if you can make a "standard" O2 restore tape and just generate installs of IRIX for your O2s from that.)

As far as software goes... I'd suggest compiling/preinstalling a bunch of opensource stuff, GIMP, OpenOffice, Scribus or whatever.. enough to do some stuff with it.

I know that I would be interested in purchasing an upped O2 or just the processor module.... (they just pop into an R5k/180 motherboard, (the 5200 modules) would they?

(like, if I just happened to stumble upon an R5k/180 for myself )

I surely know that I'd forgo IRIX upgrades if it meant running at 600, 835 or 900MHz speeds, especially if it would be possible to buy a cleaned and well-softwared O2 with a good (768mb+) amount of memory.

Well... I'm going to try something out in inDesign for "NekO2" tonight... I'd love to see what people would think about for spec-sets and prices.

[and just because it's required.]

I can't say how impressing this really is, to see that 600MHz SGI O2 systems are in action, and that there are so-many people who are continuing to work on increasing this number, and it seems like people want to put it into a 4 digit number. I can say that it's rather-exciting to see this continue to happen.

_________________
I [heart] the Performer Town Demo

I wouldn't pay a huge amount of money for a 900 MHz O2, but I would pay a a significant premium over what a similar Octane would cost. The small size is important to me, I just wish I could have had a 600 MHz module made up before LACE Technology quit doing the mods.

6.5.21m is all I need.

Maybe, but SGI isn't doing much improving on Irix anyway, so being "stuck" at 6.5.26 or so would be no great hardship. About being competitive with a Fuel, well ... I don't think so

As good of an idea as making money modding the O2 CPU modules sounds, there's not much chance of it happening. My cost (in parts and 3rd party rework charge) to mod an RM5200-300MHz board to an RM7000C-600MHz board is about $250 without S&H. Add to that price the cost of the "doner" 300MHz board, about $250 from eBay, and you can see the cost (without any value for my time) for the mod is very high. The cost of the RM7900-900MHz chip is about $150 more than the RM7000C-600MHz chip, so to make a commercial 900MHz product that is worth someones time to do, you are going to have to charge at least $1000 just for the CPU module. The real question is: how many people would buy the faster module at $1000 if you could offer one? I don't think enough people to make it worth anyone's time to produce. The eBay seller nelscon has been selling 400MHz R12KA O2 modules for less than $200 each for several weeks and the 400MHz R12KA module is going to give you about 85 - 90% of the performance of the 600MHz RM7000C module for about 20% of the cost.

At this point I don't think SGI cares what we are doing as a hobbiest community but I know they know we are modding these because I sent them 2 emails asking to license the CPU PCB design. Make no mistake though, if someone started to offer a commercial 900MHz O2 module (with a hacked ip32PROM) SGI has people who could (and I'll bet would) cripple IRIX to make it unuseable with the faster CPU modules.

I do have a new rework shop I have been working with but they are doing a fair job, not great. About 1 in 3 boards are having to be redone and this is very time consuming. I am sending another batch of O2 boards early next week (1 modded to 2M of cache), if it seems they are doing a better job, I'll post their contact information.

Joe

Sounds like the killer $$ is all the rework .... if one could obtain or reverse-engineer the cpu card, life would be a lot easier. And much cheaper ..... maybe even feasible. Reverse-engineering multi-layer pc boards is no picnic tho

I wonder what the legalities are. Perhaps one of the many Mac CPU upgrade card companies would be interested.

(My all time favorite Mac CPU upgrade was the CPU on a SRAM module... replaced the L2 cache SRAM stick on some "non-upgradable" Macs)

IIRC, SGI's old O2+ webpage specifically said "near-real-time" playback of DV files on a 400 MHz R12K. That "near" part always bugged me. I think it would be nice if an O2 could at least play those full-speed. I'd have to refer back to the Mplayer thread about MPEG2 performance on O2, but from past experience a VCD will eat up about half of a K6-3+ 450, and I would guess DVD playback is at least 4-6X the ops/sec (certainly at least 4X the resolution to fill!) Between VICE and an RM7900 it would maybe put the O2 on the edge of viability for DVD?

So while I can't deny the cost explosion for a homebrew RM7900 rework relative to R12K (a cheap PC can do these tasks fine, too), there is something charming about the idea of a little O2 reaching these previously unheard of levels with a nice IRIX desktop. And a lower wattage CPU.

Heck, any chance of overclocking VICE while we're at it?