Why do caches keep getting bigger? why can't they get smaller?

[Anarchist420]

i am so impressed by the Katana design (i.e., the Hitachi SH-4 had 24KB total cache over the loser blackbelt one which would've used a Power PC likely with not less than 160 KB total cache).

it seems to me like intel just wants higher numbers of everything rather than originality when they never had to go with higher numbers.

i mean, one physical core would've been fine especially with more and maybe unique threads (logical cores). The Hitachi SH-4 in the Dreamcast was incredible as was the Power VR CLX2

 

[KWiklund]

A larger cache helps boost performance. Basically, it takes longer to retrieve data from main (external) memory than it does to retrieve data stored in the cache. The more data (or program) you can fit in cache, the less time you spend waiting for that data to be retrieved from further away. Of course, cache is only one part of the performance puzzle, since you also need good programmers who can optimize their use of cache, as well as a CPU to match.

Outside of possible constraints on power and footprint, there's no reason not to have a large cache.

 

[VirtualLarry]

Total cache size went down, between Core2Quad and Nehalem. However, memory speeds went up and latencies went down, due to the IMC.

 

[witeken]

Because RAM is slower and has a higher latency.

It's also useful for graphics.

 

[DrMrLordX]

Intel's cache/prefetch/etc. designs are nothing short of extraordinary. Why else do you think that consumer Haswell processors gain so little in performance on a wide variety of apps from lower system memory latency? Intel chips do a great job of making sure that the data needed is in cache, even on processors that have lower L3 cache-per-die (such as the G3258). Sure, there's probably room for improvement, but many advancements have been made since the days of the Dreamcast. Their cache density is also really good. Intel doesn't have to allocate much die space to L2 or L3 these days. Compare that to, say, a Piledriver FX.

 

[Cerb]

i am so impressed by the Katana design (i.e., the Hitachi SH-4 had 24KB total cache over the loser blackbelt one which would've used a Power PC likely with not less than 160 KB total cache).

While its uniqueness and microcontroller-like ISA held it back, as it matured, the SH-3 and SH-4 were novel implementations of novel and well-tested ideas, that paid off. It was among the few RISCs that could pull off code density better than x86, and decent typical performance per clock outside of maths, for a simple/embedded RISC CPU. As the future came, it needed lower real and clock-level latencies, which SH simply could not provide, but SH-3 and 4 are a lot like the early ARMs, being rather special for their time, and technology available. IoW, they are historically impressive, and not really comparable to more modern general-purpose CPUs, nor even more costly mainstream CPUs of their eras.

i mean, one physical core would've been fine especially with more and maybe unique threads (logical cores).

By the time they got around to implementing HT again, it wasn't going to be any cheaper to do it that way, so only the Atoms got that. One core could be fine, but 2 of the same core will be better. If getting 2 functional cores presented yield problems, I'm sure we'd have seen 1C2T lower-end CPUs, instead of 2C2T, but it didn't go that way.

was incredible as was the Power VR CLX2

Then everyone else, over the course of a few years, got similar non-infringing tech, then they left for embedded pastures; and I'm pretty sure that today, Imagination doesn't envy PC GPU companies, despite their smaller size.

Now, as far as the title question:
Core 2 (Conroe): 4MB/8MB
Core 2 (Penryn): 6MB/12MB
i3 (Nehalem): 3MB
i5 (Nehalem dual): 4MB
i5 (Nehalem quad): 8MB
i5 (SB, IB): 6MB/8MB
i3 (Haswell): 3MB/4MB
i5 (Haswell): 6MB

They really haven't been getting bigger, these days. Intel has been improving their speed, and their speculative efficacy, so they've been able to not have to increase the size. Even with AMD, you can see the 2-module Richland and Kaveri CPUs trounce their predecessors, despite some of those CPUs having roughly 3x the cache, plus complete dedicated cores, all while being clearly slower than those same predecessors per single thread.

 

[soccerballtux]

working sets get larger necessitating larger cache to keep the processor fed

 

[Exophase]

Like almost every other CPU used in a console, Dreamcast's SH4 was garbage compared to contemporary PC processors. At least when it came to doing anything outside of big blocks of vector transforms. Which the other consoles offloaded to coprocessors or accelerators. SH4 was fine for Dreamcast games, but for desktop apps of the time it would have been a disaster.

I have no idea what you find impressive about it. If you think lacking cache is a great thing for that time go ahead take a 1999-era Pentium 2 and remove its L2 cache. The result will be the original "Covington" Celeron which had such abysmal performance that Intel had to make a new Celeron with a new die containing integrated L2 to make up for it.

 

[tweakboy]

So next time you want it,, baaam it will come in a instant.

 

[Anarchist420]

Alright thanks guys; but I really think that four cores at least like intel has done them is enough. I hope they make an original breakthrough soon, because AMD long ago said the future is fusion.

But how could the power pc black belt design have been a better idea?

Was the sh4 CPU arch to blame for the ATT web browser being slow?

Wasn't the smash pack decently emulated other than the pixellization and audio issues?

Would an SH-X be a good idea IN YOUR OPINION for a dream cast successor? Explain please? It could stimulate japan's economy. I was think a dream cast 2020 with a modern off the shelf power vr gpu doing normal rasterization and a powerful dual physical core (no extra logical cores) HITACHI SH processor with 64 bit ISA was a good idea. The powervr could have an Audio DSP in it.

Steve snake helped with the smash pack if i am mistaken and kega fusion is pretty versatile. It is not too inaccurate from a processing stand point; at least from the VDP and the audio processor. The z80 and 68k not all that accurate, but good enough for me anyway. Audio sounds DAMN good on my setup.

Personally, I think that with amd's possible management change they could finally get back to really good products like they did from 03 to 05 except only be profitable. They have suffered from mediocre to poor management for too long but the most original engineers win out in the end anyway.

I just don't want to see them eaten up by nvidia but fortunately I don't think they will be. I spent way too much money on nvidia stuff but I had to because it was fool proof and the abilify made me a fool. AMD hardware was always more versatile just with worse coolers and their driver team adapts to changing game engines better; rather than just a few of the newest games. They also allow their drivers to be more open. I don't like JHH's extreme imposition of uniformity. He is a good man, but their prices are way too high for me given the lack of his originality. in fact,I would say gk110 was better; the new AA modes in GMZ are retarded to me. Those who focus the Most on efficiency and consolidation usually wind up being less efficient and holding back good original production.

 

[Exophase]

But how could the power pc black belt design have been a better idea?

You mean with the rumored PowerPC 603e 166MHz instead of a 200MHz SH4? They would have needed to do something else for the vector FPU capabilities. Other than that, the processors seem pretty similar. 603e doesn't have L2 cache either, and the execution resources are comparable. It has an advantage due to having more associativity in its L1 cache (it also has double the L1 icache but that's offset by having 32-bit instructions). It's hard to say which would be better just from what I know of either.

Was the sh4 CPU arch to blame for the ATT web browser being slow?

I'm sure it didn't help.

Wasn't the smash pack decently emulated other than the pixellization and audio issues?

So? They probably took shortcuts (or didn't know what they were doing) with the FM emulation, which saved CPU cycles. Maybe it was being offloaded to the DC's AICA hardware but not really fit properly.

As a comparison point, my old 75MHz Pentium could also handle Genesis emulation decently (with the Genecyst emulator) if audio quality was similarly gimped (by being offloaded to OPL2 hardware). That was a 1994 processor.

Would an SH-X be a good idea IN YOUR OPINION for a dream cast successor? Explain please?

No. SH was a deadend, SH-X is too limited to embedded applications. Hitachi didn't have a competitive processor around 2005-2006. And they hadn't done a custom processor for Sega yet so why start now as Sega would surely be in an even poorer position to buy this? This is assuming they had the resources to do a competitive design anyway.

Japanese console companies would be wise not to rely too much on Japanese parts and manufacturing when it doesn't make sense to. Nintendo for example put themselves at risk relying on Renesas to manufacture the Wii U GPU.

It could stimulate japan's economy. I was think a dream cast 2020 with a modern off the shelf power vr gpu doing normal rasterization and a powerful dual physical core (no extra logical cores) HITACHI SH processor with 64 bit ISA was a good idea. The powervr could have an Audio DSP in it.

64-bit ISA for that generation is pointless, even if various consoles got it anyway. It only makes sense with the latest gen. I don't know why a PowerVR GPU would have an audio DSP in it.

I believe that IMG would have also struggled to have a competitive GPU product in the 2005-2006 timespan. They wouldn't have been in a position to provide programmable shaders, for instance.

Steve snake helped with the smash pack if i am mistaken and kega fusion is pretty versatile. It is not too inaccurate from a processing stand point; at least from the VDP and the audio processor. The z80 and 68k not all that accurate, but good enough for me anyway. Audio sounds DAMN good on my setup.

Steve Snake worked on Genesis emulators for over a decade. Just because KEGA Fusion would eventually become extremely accurate doesn't mean that the Smash Pack was.

Personally, I think that with amd's possible management change they could finally get back to really good products like they did from 03 to 05 except only be profitable. They have suffered from mediocre to poor management for too long but the most original engineers win out in the end anyway.

It's going to be a hell of a lot harder to do what they did with K7 and K8, where they benefited from various misexecutions on Intel's part. This was back when were more or less keeping up with manufacturing capability and even had some advantages like copper interconnects. AMD has never been as behind as they are now, it'll take a lot more than better management to completely catch up.

Kind of weird how you segued into AMD and nVidia here, I'm having a really hard time seeing how you got there from the original thread or even the post you were making

 

[Cerb]

Woulda coulda shoulda...who knows? That was 15 years ago. SH isn't quite dead, still exists (legacy), and has gone 64-bit. IIRC, some of those that worked it went on to help with Renesas' RX, which itself is a pretty cool line of chips (again, for that they are and what they are meant to do). But, SH could not have, like ARM, gotten hugely boosted performance over time, with faster wider speculative cores. While it was relevant, they smartly stuck to ease of code optimization, and low power usage for embedded devices they already had a foothold in. ARM did the same, but was able to start growing up.

A few years after the DC came out, 200MHz was not going to cut it even for a cheap game console, and the relative latency of main memory was going to be too much for such processors, running at higher speeds. Only in very simple MIPS and GFLOPS benchmarks could they look good, going forward. Speculative application CPUs, like we're used to, would have been needed to extract more real performance, but that would also require more real R&D money, which was already getting thin for these kinds of companies, relative to what it would cost to go head to head against AMD, Intel, or even IBM/Motorola. Plus, SH would have had to break its forward compatibility scheme.

Was the sh4 CPU arch to blame for the ATT web browser being slow?

A similar MIPS or PPC CPU would have also been pretty slow for web browsing, even considering the time frame. As of the Pentium, no cost-effective RISC was going to be equal or better in performance for what we used PCs for; with the P6 putting the last nail in the coffin. Anything affordable, even Apple's expensive computers, except for a few brief periods, could be measured in how much less worse they were than older PCs, how useful they were where you didn't want a normal PC (small quiet PCs weren't a big thing until the Mac Mini got clones, for instance), and then other tertiary feature concerns.

The SH-4 and how it was used in the DC seems pretty cool due to it being a $200 box, in 1999, that Western devs afraid of the Saturn could work with. It was not going to compete with any PC from years before it, much less at the time, or even low-end much later, in any way but mul then add throughput. For a cheap game console box, Sega made something pretty impressive for the day. But, it was a cheap console box, and it's actually all rather pitiful if that context is removed, and you were to compare it to a low-end AMD-based gaming PC of the day.

 

[sefsefsefsef]

working sets get larger necessitating larger cache to keep the processor fed

I actually don't think this is true. There have always been some applications that have large working sets, and many that don't. Do you have any particular class of applications in mind where they could benefit from more cache every year? For many workloads with large working sets it's nearly as easy (and less expensive) to have a moderately sized cache and then just prefetch the working set into the cache each time it is used.

Also, caches aren't getting bigger. Last level cache sizes have held constant at ~2MB/core for years now (from both Intel and AMD).