A CPU cache mod idea :)

Beeblequix · 01-24-2003, 03:31 PM

I get great ideas all the time. But the fact is I'm F**<>>G lazy, so I just usually do the hand-off approach and continue being lazy...

What we need is a L4 cache upgrade for our CPU's. Like, you'd drop your socket A Athlon with 384L2 cache on top of a similar form factored memory bank with the same pinouts, but becomes a plug-in-able (a word?) cache expansion that benefits the system by:

1) not having to chug through the relatively slow memory bus to system memory
2) very low tracelengths for REAL quick access
3) your applications which crave a large cache will boost in speed immensely without investing in a Xeon or RISC class enterprize server
4) you'll be cooler

Issues to face:
1) by altering the structure as it is with regards to the voltage requirements & signal corruption due to longer tracelengths, the cache add-on would need to stabilize the current with capacitors. Would it need a slight voltage increase?
2) would it be necessary to create a mini-memory controller chip to handle the memory IO in this extended cache for hire?

I don't know -- I'm just bored at work.

But could this have potential?

ßeeßle Popeil

AllGamer · 01-24-2003, 03:39 PM

How about.....

Scrap the whole idea

and instead

De-Solder the old smaller L2 and/or L3 caches chips

then Install and Solder new and Larger Capacity L2 / L3 caches replacing the old ones

much easier, a lot less complicated, and wont need any additional electronic modification

BTW, that's exactely what Intel did with the Xeons

causticVapor · 01-25-2003, 02:03 AM

B.Quix, I've been thinking about a similar thing, like the return of the COAST slot cache. Detachable, upgradeable cache. No desoldering required.

Could mean more business for Rambus, and would be a lot faster than main memory while being dedicated to the CPU. The only problem is, price. Running mid-price cache at even 1/4 the speed of a 2GHz CPU is A) hard B) consequently expensive. And synch-- 2GHz+ cache in sizes over 1MB -- for now, forget it. This one reason is why the P2/P3 xeon cost so much -- the L2 cache ran at CPU speed. Itanic and others' clock speed ramping are hindered by high transistor counts, partially related to the cache.

But the idea is quite sound. It just needs to become mainstream... psst... a very good alternative for increasing IGP performance, instead of soldering mem chips on the motherboard, put them in a COASTlike BEACH (But Ever Always Close Home) slot. Saves on board real estate and is optional as well, so it can save money and hassle. Coast/BEACH L3/IGP cache... hmmm...

causticVapor · 01-25-2003, 02:06 AM

Quote:

Originally posted by AllGamer
How about.....

Scrap the whole idea

and instead

De-Solder the old smaller L2 and/or L3 caches chips

But then you have to face the fact that the CPU has a limited L3 cacheable range. For instance, if I put 2MB chips on my MVP3 board, which already has 2x1MB chips, it probably wouldn't work. At best, it would recognize only 2MB of the 4MB, and be error prone due to signals straying in the nonaddressed region. Besides, who would want to desolder all the pins, make the board look crappy, and then crappily solder them back on?

And who wants to go through the trouble of getting a high-grade, high-precision iron to do it? No mobo manf. would want to....

Now I would certainly do it if I knew the CPU/northbridge were able to address more and if it were equipped with a small amount, like 512KB. Then we take Mr. Happy Iron!

And as for everything else....

I smell a cache on the COASTline.....

sm8000 · 01-25-2003, 12:55 PM

How about CPU makers have the L2 cache on-die so that the mobo's L2 cache becomes L3? Like what happened with the K6-2+ and K6-3+? Or is it the mobo makers who need to embed L2's on boards again? If so, I do like the idea of upgradeability there.

OpK Chowdy · 01-25-2003, 01:24 PM

This sorta reminds me of that "hard drive" which goes in a PCI slot and you plug in up to 4 PC-133 ram sticks in it for a maximum of 4gb of really, really fast storage. There's a separate PSU for it too. Duno..that just reminded me of this thread

causticVapor · 01-25-2003, 05:44 PM

Quote:

Originally posted by sm8000
How about CPU makers have the L2 cache on-die so that the mobo's L2 cache becomes L3? Like what happened with the K6-2+ and K6-3+? Or is it the mobo makers who need to embed L2's on boards again? If so, I do like the idea of upgradeability there.

sm8000, I mentioned that in my previous two posts. Do you want $4000 athlons running at 2.25GHz? If you say yes, then there you go, 3MB synch on-die L3 cache. It's very hard to make cache run at such high speeds and yields are very low. Think about way back in the days of the Pentium Pro. That's exactly the problem that would occur now if we tried to integrate so much cache running at such high clock speeds.

A better solution would be to use some dedicated cache running at maybe 1/2 or 1/3 the CPU clock speed, and in a COAST-like slot so sizes can be upgraded and swapped. With soldered on cache, the motherboard automatically costs more and you have to throw it away if the cache goes bad. Implementing a slot would be a lot less expensive initially and you could opt to have it empty. If a cache module is bad, you can simply swap it.

Performance motherboard makers could allow you to adjust the cache divider in the BIOS, just like with the classic athlons. And yes, they'd allow voltage increases and whatnot for it, as well as cooler manufacturers coming up with interesting "cachesink" designs.

We need a lot of people to submit this proposal to Intel and AMD. Perhaps then we will arrive with the return of external L3 cache.

causticVapor · 01-25-2003, 05:46 PM

Quote:

Originally posted by OpK Chowdy
This sorta reminds me of that "hard drive" which goes in a PCI slot and you plug in up to 4 PC-133 ram sticks in it for a maximum of 4gb of really, really fast storage. There's a separate PSU for it too. Duno..that just reminded me of this thread

Even though solid-state is more expensive, it's better. What happens to your PC133 "hard drive" when the power goes out? All current is lost and due to the volatile nature, so are the signals (data). Battery, yes. Batteries die.

OpK Chowdy · 01-25-2003, 06:15 PM

Yeah, that's why i wouldn't trust one of those suckers with imporant data, and the only ones who'd gain, really, are servers...so oh well. I think it's a pretty nifty idea, though. It costs $5,000 for the 4GB version BTW.

causticVapor · 01-25-2003, 06:36 PM

Hell, for that price you can get FIVE gigs of solid-state storage.

OpK Chowdy · 01-25-2003, 09:15 PM

what is solid state?

causticVapor · 01-26-2003, 12:50 AM

Solid state media is like flash memory -- non-volatile, and keeps its information even when its power is not refreshed. Somewhat like EEPROM, somewhat not. The USB keys use it.

The only hindering factor for now is $. Since it costs $1 per megabyte, a 30GB unit would cost $30,000.

OpK Chowdy · 01-26-2003, 08:33 PM

is solid state fast? comparatively speaking.

fizur2002 · 01-28-2003, 02:03 AM

I have a stupid idea, but here it is anyways. You have your rimm, dimm, and simm slots right. Well separate from your normal ramm slots, you have a separate slot next to or really close to your processor and have a separate high speed and high MB count stick in it to make it your next L# cache. Told you it was stupid, but im bored and wanted to type something.

causticVapor · 01-28-2003, 01:34 PM

Quote:

Originally posted by OpK Chowdy
is solid state fast? comparatively speaking.

To quote allgamer -- it's not fast, it's INSTANT and just simply amazing