Giant 3rd party L3 Cache add-on- why not?
People upgrade with gobs of memory all of the time, right? Why isn't there at least one specialty store out there that sells L3 upgrades? Something like a 64 or 128 MB L3 plug-in? It would be an expensive 64 MB, but surely you could buy one for the cost of an expensive videocard ($300-400)? It is an expensive upgrade, but for the power-hungry professional user, it is pocket change to justify if it does improve the system. There may not be enough money in the world to buy a 2 Ghz G4, but you can buy something that will allow your 1.2 Ghz G4 to run at full tilt, all of the time. ...Or is this L3 cache still on the "other end" of the FSB bottleneck from the CPU?
Could this be what we need to get snappy OSX performance like we never thought possible or truly sustained Altivec operations 16 GFlops, as is claimed in Apple literature? Yeah, I know- I'm always ranting on about large, fast caches with unsubstantiated performance benefits. I just wonder what this could do.
Could this be what we need to get snappy OSX performance like we never thought possible or truly sustained Altivec operations 16 GFlops, as is claimed in Apple literature? Yeah, I know- I'm always ranting on about large, fast caches with unsubstantiated performance benefits. I just wonder what this could do.
Comments
I've seen the VRAM chips on the 2nd gen iMac done as a flip-chip slide-in, so that is why I ask.
[ 08-26-2002: Message edited by: Randycat99 ]</p>
My thought was that the typical outboard cache would be similar to the high-bandwidth VRAM used in videocards.
If this is not the case, would there be any advantage (size over latency/speed) if VRAM type of memory was used as an L3 cache?
The backside cache memory is static RAM. As opposed to dynamic RAM (DRAM) technology, this memory does not need to refresh so it is 2 to 5 times as fast as any DRAM tech (SDRAM, DDR SDRAM, RDRAM) due to its short latency. Unfortunately, a 1 MB chip of SRAM can be as expensive as the CPU itself.
VRAM is a dual ported memory technology, ie, two devices can access it at the same which is necessary for video cards since both the video chip and the CPU may access it at the same time. It is not faster than SRAM.
Now however, a gigantic backside cache is a very interesting idea. A quad-channel single-chip-per-channel Rambus backside cache would be very intriguing. That can be 128 MB backside cache at 8.4 GB/s bandwidth using PC1066 DRDRAM...
[ 08-27-2002: Message edited by: THT ]</p>
Why do you suggest Rambus, BTW? Isn't it always said the latency effects would make it horrible for cache use?
Well, damn it (about the VRAM idea). How about that 1T SRAM stuff? Would that make for good cache memory?</strong>
Don't really know much about 1T-SRAM. I would say that there is no such thing as a free lunch, and 1T-SRAM is a DRAM altered to behave like SRAM, so there will be some consequences due to that. IBM and Nintendo use it for the Gamecube main memory, but I'm not sure if it performs better than DDR SDRAM or RDRAM. I doubt it performs better than SRAM.
<strong>Why do you suggest Rambus, BTW? Isn't it always said the latency effects would make it horrible for cache use?</strong>
A single RDRAM chip should have the latency of DDR SDRAM, and on top of that, it would have all of the sense-amps to itself increasing both bandwidth utilization and latency further.
The Rambus solution you are thinking about (used in P3 and P4 system) consist of 16 or 32 chip RDRAM channels which carries with it some consequences that really hurt its latency because of the serial nature of the tech. However, reducing it to one chip per channel gets rid of a lot of headaches. One potential is clocking RDRAM higher, say 600 MHz.
The Playstation 2 uses dual-channel single chip per channel (and I believe on-chip memory controller) RDRAM solution for its main memory, so it can't be hurting it too much...
[quote]Originally posted by THT:<strong>
The Playstation 2 uses dual-channel single chip per channel (and I believe on-chip memory controller) RDRAM solution for its main memory, so it can't be hurting it too much...</strong><hr></blockquote>
Thank you MUCH for that! I have wondered about that very thing for some time. Specifically, if latency effects are comparable between Intel's i850 and Sony's PS2 since they both use Rambus or are they different due to the specific implementations of Rambus in their respective systems.
As always, THT, you da man on hardware!
[ 08-27-2002: Message edited by: Randycat99 ]</p>
<strong>Well, damn it (about the VRAM idea). How about that 1T SRAM stuff? Would that make for good cache memory?</strong><hr></blockquote>
1T SRAM would behave in a very similar way to DRAM. In fact, 1T SRAM is DRAM. The "1T" part means that it's a DRAM, and that means one transistor (1 T) and only the SRAM meaning is that it uses the same process steps as SRAM does. This means you don't get the great deep trench/finned capacitor to give the capacitance. 1T SRAM is basically two transistors per bit, one of which is behaving as a capacitor, the other is the access method.
1T SRAM is good for some things -- where silicon area is expensive (either for $$$ or timing) typically. Real SRAM is good when fast accesses are necessary with small paths and small amounts (up to 6-8 MB for example). Real DRAM is good for problems that can be solved with some latency, bursty accesses and the decreased utilization for refreshing the capacitors are acceptable.