What if Intel / AMD designed CPUs with lower clocks / higher cores (like GPUs)?

[PianoPlayer88Key]

So I was thinking ... what might happen if Intel (or AMD) designed their CPU core count & clock speeds to be more similar to GPUs at similar prices?  (Basically significantly lower clocks, but many more cores.)

 

For example...

Spoiler

(btw, I'm using both texture units and cuda / amd-equivalent for the gpu core comparison.  I'm not sure which should be used.  I'm thinking Tex is more comparable, but "cuda" is what's called "cores" so...)

~$1000 i7-X / ? = 1480 MHz, 240 / 3840 cores (Titan Xp)

~$600 i7-K / ? = 1480 MHz, 224 / 3584 cores (1080 Ti)

~$500 i7-K / Ryzen 7 > 1607 MHz, 160 / 2560 cores (> 1080)

~$350 i7-K / Ryzen 7 = 1506 MHz, 120 / 1920 cores (1070)

~$250 i5-K / Ryzen 5 = 1506 MHz, 80 / 1280 cores (1060 6g) (or) 1257 MHz, 144 / 2304 cores (580 8g)

~$200 i5 / Ryzen 5 = 1506 MHz, 72 / 1152 cores (1060 3g) (or) 1257 MHz, 144 / 2304 cores (580 4g)

~$175 i5 / Ryzen 5 = 1168 MHz, 128 / 2048 cores (570)

~$140 i3 / Ryzen 3 = 1290 MHz, 48 / 768 cores (1050 Ti) (or) 1090 MHz, 56 / 896 cores (460 4g)

~$110 i3 / Ryzen 3 = 1354 MHz, 40 / 640 cores (1050) (or) 1090 MHz, 56 / 896 cores (460 2g)

~$80 Pentium / ? = 1100 MHz, 32 / 512 cores (550)

 

While per-core performance would be down significantly (compared to, say, a delidded i7-7700K under LN2, overclocked so hard it hits 100°C and thermal throttles), would the multi-threaded performance be significantly improved to more than make up for it?  Or, might it be an incentive for software developers to expedite optimizing for parallel workloads?  Or what other repercussions might that have?

 

Spoiler

Also I'd like to see multi-CPU support on mainstream / consumer platforms, like AMD CrossfireX allows multiple GPUs.  For example, the ~$80 RX 550 supports 2-way crossfire, and the ~$170 RX 570 supports 4-way crossfire.  Yeah I know games have the risk of not scaling well with multiple GPUs, but what about workloads typically done on CPUs, or multitasking / VMs?

 

[themctipers]

single threaded performance is going to hit Pentium 4 levels again

multi threaded would be the same, assuming there's no IPC increase.. however, what program do you know that can use ~2000 cores at once, CPU cores at once..

 

also the cpu price would skyrocket, as having more cores and threads on a single piece of silicon is hard to manufacture, and has lots of faults..

 

[ARikozuM]

This would be a very bad idea. GPU's have to work on many small pixels and textures where the sheer numbers makes cents ... 

Spoiler

I'm terrible.

... but CPU's have to handle many random operations that require stronger cores.

[SpaceGhostC2C]

It's more about the nature of the task than software optimisation in most cases. 

 

In any case, you may want to check GPGPU and Xeon Phi. 

[Dabombinable]

AMD already did that with Bulldozer and its derivatives. It didn't work out well for AMD.

[Guest]

I'd assume it'd work well.... if every program in existence (including operating systems) recoded to work well with it. AND you are fine with higher prices. So basically, not a good idea...

[mariushm]

They did.

Intel made a processor with lots of cores, each core with a smaller set of instructions, it's called Xeon Phi : http://www.intel.com/content/www/us/en/products/processors/xeon-phi/xeon-phi-processors.html#dynamicProductCardBlade

Apparently they go up to 72 cores.

 

There's also "many cores" processors like for example Ambric AM2045 with up to 344 tiny cores (up to 333 Mhz core frequency) : https://en.wikichip.org/wiki/ambric/am2000/am2045

There was also ClearSpeed CSX700 with 192 tiny cores : https://en.wikipedia.org/wiki/ClearSpeed

 

[SCHISCHKA]

6 minutes ago, Dabombinable said:

AMD already did that with Bulldozer and its derivatives. It didn't work out well for AMD.

uh not quite. AMD thought they could increase clock speeds

[SpaceGhostC2C]

1 minute ago, SCHISCHKA said:

uh not quite. AMD thought they could increase clock speeds

I think it was just a bad joke. Bulldozer has nothing to do with what OP describes. 

Phis, on the other hand, are a step in that direction. But they are meant to shine in the scenarios in which GPGPU was already profitable. 

[mariushm]

AMD squeezed 8 cores into a processor by making pairs of cores share the floating point unit, betting that people won't use a lot of applications that need floating point or don't use it a lot... and they were right, but there were also some tiny tradeoffs that affected performance.. and applications and games were too slow of converting to multithreading (they had no incentive, consoles only had a few cores)

They were royally screwed because the GlobalFoundries factory that made their processors couldn't upgrade as fast as they estimated so they still had to make the processors on 32nm and around the time Bulldozer showed up TSMC also cancelled some upgrade to a smaller process node so they had to make some of their processors on 28nm (the FM ones if I remember correctly)

It was way too expensive to convert the Bulldozer designs to 28nm and start making them on TSMCs factory (probably around half a million dollars or more and up to around 6 months to first batch of sellable cpus ) and AMD also had/has a contract that forces them to manufacture a minimum number of wafers at GlobalFoundries every semester so they were kind of forced to leave those FX processors on 32nm and just make small revisions ... they sucked it up and started work on Zen with the 14-16nm figures in mind from the start .. GlobalFoundries screwed up even that upgrade process, they just gave up in the end and licensed technology from Samsung to get things working.

Nowadays AMD won the processors and gpus for major consoles, they have zen competitive, they have HSA and other technlogies that could allow them to run code in cpu cores or gpu cores transparently etc etc

 

[i_build_nanosuits]

1 hour ago, PianoPlayer88Key said:

So I was thinking ... what might happen if Intel (or AMD) designed their CPU core count & clock speeds to be more similar to GPUs at similar prices?  (Basically significantly lower clocks, but many more cores.)

it would be aweful because like 90%+ of the softwares out there are programmed to make use of 2 to 4 CPU threads, and when they can use more they still rely on great single-threaded performance...basically, single-thread performance is still a lot more beneficial for the vast majority of tasks.

[Dabombinable]

1 hour ago, SCHISCHKA said:

uh not quite. AMD thought they could increase clock speeds

The aim with Bulldozer was more threads without that associated costs (eg. "16 core" Opteron which excelled at server orientated tasks and file compression but sucked at everything else).

AMD might have panicked on the consumer side of things and started releasing higher clocked versions but the IPC was utter shit at 5GHz as opposed to lower clock speeds around 2-3GHz.

Because at around 1.6GHz Bulldozer's Trinity derivative is as fast as Arrandale (around the same age) at 1.6GHz, while at 5.4GHz its slower than Wolfdale-which is 5 years older-at 4.4GHz.

1 hour ago, SpaceGhostC2C said:

I think it was just a bad joke. Bulldozer has nothing to do with what OP describes. 

Phis, on the other hand, are a step in that direction. But they are meant to shine in the scenarios in which GPGPU was already profitable. 

Read above-the ludicrous clock speed goals were there because AMD realised that they'd fucked up and that even i5 2500K were able to beat the FX 8350 in a large amount of productivity orientated tasks and in all games.

[SpaceGhostC2C]

5 hours ago, Dabombinable said:

(...) 

You are going sideways big time. Bulldozer has nothing to do with the gpu-like processors OP is talking about. You may want to read above indeed. 

[Dabombinable]

10 minutes ago, SpaceGhostC2C said:

You are going sideways big time. Bulldozer has nothing to do with the gpu-like processors OP is talking about. You may want to read above indeed. 

"CPU's with lower clocks higher cores". That's what Opteron's were.

[SpaceGhostC2C]

6 minutes ago, Dabombinable said:

"CPU's with lower clocks higher cores". That's what Opteron's were.

That's also what Xeons were. Actually, that's what Xeons and Opterons are. 

Except it would have served you to read the full post instead of stopping at one sentence. Then you would have realized that "let's increase Phenom II clocks and add CMT" how nothing to do with HUNDREDS of cuts running at very low clocks. 

If you also read beyond the first post, you will also notice that many of us already pointed out that this is the idea behind Xeon Phi. A very different animal from Bulldozer. 

[rockon5622]

So, it doesn't work quite like that. This would also defeat the point of a CPU. If you wanted a million very simple cores, you get GPU. Clock speeds would not really affect the number of cores that much other than the trade of area for speed. If you want more cores, you need a bigger chip, a more complex chip, a harder to make chip, a more expensive chip. The core architecture of a CPU core is insanely complex and has a many instructions that do very complex things. While GPU cores are much simpler. This actually kinda gets into the debate of RISC vs CISC. It also begs the question on how much can you parallelize the code you want to run on a CPU. 

 

You also have another issue that you will run into, communication between cores. That and the amount of L1 Cache per core would not be able to stay the same for the area. I can go on, but basically, the CPU wants to have a very very fast complex single core and when thats not possible have more of those fast cores. If you could have a single core CPU with 4x performance as a quad core it would be  better than the multicore counterpart. I suppose the thread switching time now comes into play. It's a trade off in the end. 

[PianoPlayer88Key]

Okay, so I know that a lot of current software doesn't do as well with many lower-clocked cores.  But I was thinking ... it's gotten much more difficult to improve single-core performance over time, than it used to be.  

 

Spoiler

For example, compare the single-core performance of the i7-2600K vs the i7-7700K.  Looking at a couple benchmark sites, they appear to be about 35% different.

 

Back in 1989, my dad bought a 286-10.  In 1995, he bought a 486DX4-120.  Looking at the Wikipedia MIPS page, between two similar CPUs, the (486)DX4-100 is about 54.7x faster, single-threaded, than the 286-12.  There'd be a bit wider difference between my dad's CPUs.  Also he paid 1/3 the price, counting the motherboard & RAM (bundled with the 286, separate with the 486), for the 486 as he did with the 286.

 

Also, I think back in that era, the per-clock IPC improved dramatically.  Per my research, the 286 was about 2x faster than the 8086, clock-for-clock, and the 486 was similar over the 386, and the original Pentium (P5) over the 486.

 

I wish we could catch back up to where we would be on price to performance, if that trend hadn't slowed.  I realize we won't get that drastic of an improvement per year on single-core performance anymore, which is at least partly why I'm thinking about multi-core, like GPU style.

 

Spoiler

 

I choose GPU style, because from what I can tell, their raw processing power is a lot higher than CPUs, at the same price.  (For example, a GTX 1060 or RX 480 might be around 4000-5000 or so GFLOPS or so, whereas an i5-7600K seems to be closer to 27 or so GFLOPS according to a SiSoft reference I found online.

Even with some margin for error, maybe that's not the best benchmark to compare, and differences in the real world (for example RX 480 & GTX 1060 TFlops appear ~25% different but in games they're similar performing it seems), I still think that's a pretty wide gap, for similarly-priced products.

 

 

And, isn't GPGPU at thing, too?

 

Also what would it take to give programmers a kick in the gluteus maximus, so that they'll expedite optimizing for more parallel workloads?  I don't think, unless some miracle happens, that we're gonna suddenly jump up in single-core performance or IPC like what was happening between the 8086 and the original Pentium, or even the Pentium III or 4, right?

 

I'm just getting tired of lackluster CPU upgrades per generation.  Nvidia and AMD are taking a step in the right direction with their CPUs, and Ryzen is a big improvement over Bulldozer for AMD CPUs.

 

Intel, where are you?   In encoding H.265, 4K, 30fps, 200mbps, all-intraframe video at about 1 fps per pass in Handbrake, my i7-4790K gets about 1 FPS, and i7-6700K maybe around 3 fps or so.  

Spoiler

(I might want to run those tests again, though, or is QuickSync or something a factor, even though the iGPU isn't used on the laptop that has the 6700K?)

I want my next ~$300 CPU, sooner rather than later (like 2018 or so), to encode similar or higher quality video at faster than realtime, be able to encode & live stream on the fly, etc.  And if I end up waiting until like 2020-2022, or, for example, when my PSU's warranty expires, or when DDR5 & PCI-E 5 are out, or when I can get an 8K 30/60fps camera for around $800-1K or so, then I'll want be able to live-encode/stream that 8K video.  (For now, I'm using a Panasonic FZ1000, which tops out at 4K, 100mbps, 30fps, and I believe doesn't have HDR during video recording.)

 

 

 

Anyway ...

 

what would it take to DRAMATICALLY improve the performance of CPUs and the software run on them, since we can't really improve single-core performance nearly as much as in past decades?

Spoiler

I guess that last pseudo-statement or whatever could be kind-of like a tl;dr?

 

[ARikozuM]

1 minute ago, PianoPlayer88Key said:

snip

If game engines (or programs in general) would start using features that are present on the newest cores leaving the generations n-1 or n-2 behind, we may see some more performance coming out of the CPUs. Remember that programmers aren't programming just for today's CPUs, but the ones from yesteryear and beyond. 

[PianoPlayer88Key]

Yeah true, I suppose, @ARikozuM.

 

There's another thing I thought of, too.  What about making legacy software run faster on newer CPUs, for when newer software doesn't do the desired task, or exceeds budget, or whatever?

[ARikozuM]

3 minutes ago, PianoPlayer88Key said:

Yeah true, I suppose, @ARikozuM.

 

There's another thing I thought of, too.  What about making legacy software run faster on newer CPUs, for when newer software doesn't do the desired task, or exceeds budget, or whatever?

There's very little incentive to do so. Legacy becoming faster is good, but the engines will limit the extent to which it can go. Magic Carpet, a game from the 90's, runs very well on older hardware at 30-120Hz, but on newer hardware where the game gets 600Hz, it becomes totally unplayable as the engine has no handling of time at those high frame-rates.

[PianoPlayer88Key]

Just now, ARikozuM said:

There's very little incentive to do so. Legacy becoming faster is good, but the engines will limit the extent to which it can go. Magic Carpet, a game from the 90's, runs very well on older hardware at 30-120Hz, but on newer hardware where the game gets 600Hz, it becomes totally unplayable as the engine has no handling of time at those high frame-rates.

Ahh.  And that reminds me of some really old games that I've played, whose performance was tied to the CPU clock.  Gapper (from 1986), played well enough on our 286, but try it even on our 486 and it was way too fast, making it unplayable.

I wasn't actually thinking of games though.  I was thinking of other tasks, like effects in audio editing programs like Audacity, among other things.  (I *think* Audacity doesn't support multithreading, but I'm not 100% sure.)  Another example would be rendering in something like Sketchup (which I've read somewhere also may not support multithreading, but I'm not sure on that either).

[ARikozuM]

Fortunately and unfortunately, most programs will make multiple versions and cut off the upgrades based on the hardware of the system being checked. It'd be great if the P4 could have CS6 (as an example, I don't know and haven't tried), but can only run the 2000 version at best. Should it be possible? Yes. Should Adobe allow it? No. There comes a point where legacy has to be dropped in favor of advancing. This is the reason why Adobe utilizes as many cores as you give it (up to a point) and prefers faster cores. Adobe's customers (prosumers and artists) have the money for quad-core and higher count CPU's. Game developers have to target their demographic consisting of up to hexa-core processors. There's little to gain from programming for 20-core Xeons that run at 1.8GHz. Single-threaded performance will tank due to speed, while multi-threaded becomes a headache to explicitly split between logical cores as there are too many. 

[PianoPlayer88Key]

Interesting, @ARikozuM.

 

Yeah, I suppose dropping legacy is something to do eventually.  For example, my current motherboard (ASRock Z97 Extreme6) only has PCI-E slots and SATA ports (with a couple M.2 slots).  There's no built-in PCI, AGP, ISA, IDE, ST-506, floppy controller, etc.  My previous board (Gigabyte GA-MA69G-S3H) had a mix of PCI and PCI-E, as well as IDE and SATA.  I think one of my brother's or dad's computers maybe a few "generations" older had ISA slots.  ("generation" meaning our upgrade cycle.)

 

I've heard a little of rumors that Intel might be dropping some legacy instructions with the revision after Tiger Lake.  I was thinking .... for someone who needs those instructions, but also needs the capabilities of the new platform, could a PCI-E card or something be made that would support those instructions?  (It would be sold separately, kind-of like the math co-processors back in the days of 286/287, 386/387, 486/487.)

 

Something I've wondered about, and maybe this should be a different topic, but .... do I (and my dad) have a longer upgrade cycle than a lot of typical computer users, even gamers?  

Spoiler

(1989, dad: 286-10.  1995, dad: 486-120.  1999, dad: Pentium 166, bought from my bro, who upgraded to Pentium II-400. 2002, dad: Athlon 1.4.  02-2008, me: Athlon 64 X2 4000+. 08-2008, dad: Core 2 Duo T7250 (laptop).  01-2015, me: i7-4790K (desktop).  12-2015, me: i7-6100 (laptop, upgraded 11-2016 to i7-6700K.))

I'd love to have a similar price/performance jump from my 4790K to my next CPU, that we had from the 286 to the 486.  I really don't want to have to wait almost forever for it, though.  5 or 6 years, the time frame from the 286 to 486, would be acceptable.

[ARikozuM]

I don't think we'll see another grand improvement for awhile. We'd need something akin to XP to Vista to make it happen. The underlying layers all have to be better to make the most difference. Even back then, upgrading the CPU was a 5-7 year thing. There wasn't much reason to build a PC if you already had one that worked unless you just needed a new dGPU or expansion card.