Hacker News Reader

Comment by traden209 | Hacker News Reader

You still can't beat intel at the high end. Both in terms of heat and performance. If you have a task that doesn't parallelize well, you can't beat intel.

Games are traditionally single threaded and hard to parallelize.

However: you don't need to be the best, you just need to be at the "agreed upon" high end for game requirements, which is the xbox and ps4 (soon to be next gen).

And so, if you're trying to build a "generation proof" gaming PC in a standard sized case, the 3800X is currently your best bet for a new build (or waiting for zen3). If you're trying to do SFF, Intel may be a better choice.

Regardless: AMD's competition has dropped the prices in the "enthusiast" segment. My next build will be AMD, simply because I don't need the best of the best, but rather, good enough.

View on HN

> Games are traditionally single threaded and hard to parallelize.

On my 3900X I've seen even single player action titles utilize 16+ threads with overall CPU utilization reaching peaks of 80-90 %.

Yes, it maybe be hard to parallelize game engines, but all major ones heavily use multiple cores. A bunch of years ago (4-5) people would recommend an i3-K dual core for gaming, because you could overclock it to 5+ GHz on air. That was sufficient for many titles back then. (And these i3s would sometimes give you higher FPS compared to the more expensive quad-core options). But a CPU like that is far too weak for newer titles.

> You still can't beat intel at the high end. Both in terms of heat and performance. If you have a task that doesn't parallelize well, you can't beat intel.

In terms of heat, that is, performance per Watt, Intel's high end is quite a bit worse than AMD's. It is true however that you can get slightly higher single-threaded performance out of Intel's highest-end CPUs, but not because they are more efficient (- AMD is better both perf/W and IPC-wise), but just by raw, unadulterated clock speed, pushing them to 5+ GHz. And that causes them to burn a lot of power.

Zen 2 has consistently ~10-15 % higher performance clock-for-clock compared to Intel's best while achieving approximately 60-80 % higher perf/W.

This LTT review captures the massive advantage Zen 2 has in efficiency: https://www.youtube.com/watch?v=ZYqG31V4qtA

The efficiency at load is great, the power usage at idle could use some improvement (I feel there's a missing update). My previous system (6700k boasting to 4.6ghz) used about 50 watts at idle (from the wall). After I switched motherboard and CPU (a 3950x, the rest is the same, plus new Windows install) the power usage at idle is ~90 watts. The power management is too eager to use all the cores, I'm pretty sure their would be zero decernable difference if they shutdown one of the CCDs (chiplets) when utilisation is low. I've played around with the power settings and ~90w is about the best it gets.

Under load the system is brilliant (it's 4x faster with the same power usage, makes using my laptop painfully slow in comparison). The power management on these CPUs is incredibly advanced (each core runs at different speeds & voltages), I'm just surprised that it doesn't do better at idle.

The 3900X and 3950X sadly still have issues with idle power, the lower end chips do much better. A 3400G for example self-reports less than 10 W at idle, entire system is around 20 W out of the power socket. The uncore of my 3900X consumes more than that in idle according to itself...

The usual desktop (3500-3800 series) chips perform similarly well, it's just the two-chiplet AM4 processors that have been somewhat plagued by problems like this.

I was running into issues where just having Steam open, not doing anything at all, caused a R5 3600 to go from 2 watts, to 13 watts power usage. Start closing background processes. There might be one doing weird stuff keeping cores awake

>On my 3900X I've seen even single player action titles utilize 16+ threads with overall CPU utilization reaching peaks of 80-90 %.

which ones? genuinely curious. I have two Intel E5 eight-cores in an old workstation and rarely ever see a game using more than 2-4 threads.

I guess i'm not spending my time on the latest and great AAA games as soon as they come out, but i've spent quite a few (shameful) hours gaming on fairly recent titles.

Doom Eternal doesn't even have a main thread and uses a dependency graph of jobs, so should paralellize reeeeally well across multiple cores

The Division (a game I worked on) is quite well optimised for multi core.

In fact it’s one of the reasons we made our own game engine.

Rise and Shadow of the Tomb Raider is very CPU intensive and seems to use as many cores as you can feed it.

That's a bit outdated. Yes, the fastest cpu for games is still an Intel cpu, the i9-9900K. But the distance is not big. The 3950X is very close, the more reasonable AMD cpus - Ryzen 7 3700X and Ryzen 5 3600 - are also close. I collect benchmarks, comparison here [0]. When the next Ryzen generation arrives on the desktop even that small top spot for Intel will likely end.

It's not only that the turbo boost of those Ryzen cpus is high and at the same clock Intel does not have a lead anymore. It's also that recent games are already pretty parallelized. There is no need to wait for the next console generation to see that effect.

Btw, the 3800X is not a good pick. It is just a minimally higher clocked 3700X. If you want 8 cores, get the 3700X, the higher price of the 3800X is just a waste of your money. But for gaming you want to stick to a Ryzen 5 3600.

[0]: https://www.pc-kombo.com/us/benchmark/games/cpu/compare?ids[...

> Btw, the 3800X is not a good pick.

That really depends on the pricing situation, very much like with the 3600x.

Over the months I've seen the price difference between a 3800x and 3700x see as low as 20€, the same with the 3600 and the 3600x.

That's because the 3600 and 3700x are recommended, and sold so often that retailers sometimes end up sitting on large stocks of 3600x/3800x too long and clear them out at discounts.

Imho even for gaming the 3800x might be a pretty good long term choice. The higher clocks are particularly relevant yet you still get the core advantage of the 3700x, without taking the same clock hit.

But if it's all about the budget and price/performance then it's really difficult to beat the 3600.

There is no big clock hit. The turbo clock is 4.4GHz vs 4.5GHz, that's nothing. In practice the difference is mostly 1 FPS. The base clock difference looks bigger, 300 MHz, but Ryzen does not stay at the base clock like old Intel cpus did, it always turbos.

Now, if it's the same price there is no reason not to get the 3800X over the 3700X :)

The 3700X is a 65W part, so its quieter for those who dont like listening to fans spin loudly.

Higher base clock might mean you can overclock it higher. Of course, most people don't overclock their CPUs, even if they are unlocked.

I don't see why the base block would be the factor for that. The Ryzen cpus overclock on its own, it should be the turbo clock and PBO that shows how well they overclock. I just linked in another comment to an article that explains that while the 3800X does overclock better, it's a minimal difference, right in line with their turbo clock specs.

Presumably it basically is the same CPU, potentially just better binned ones being flashed as 3800x?

Exactly. https://www.techspot.com/review/1899-ryzen-3800x-vs-3700x-di... explains more details:

> The top 20% of all 3800X processors tested passed their 4.3 GHz AVX2 stress test, whereas the top 21% of all 3700X processors were only stable at 4.15 GHz. Also, all 3800X processors passed the test at 4.2 GHz, while 3700X processors were only good at 4.05 GHz, meaning the 3800X has about 150 MHz more headroom when it comes to overclocking.

That's a minimal difference of course, but right at the clock wall of Ryzen, so AMD has to do binning here to enable the 3800X.

I think average frame (which I assume this is) is much less useful than minimum, 0.1%, and 1% frame times. As a human, I'm more sensitive to instantaneous gameplay experience, not average. It would be very interesting to see graphs that show these more telling numbers.

Yes, these are average FPS. Minimum is a useless metric that is usually caused by benchmark issues, but 0.1% and 1% would be interesting. But they are usually correlated to average frame times, it's too rare that in specific games one processor has good average fps but sporadically dips to make the enormous effort to collect that data as well. When that happens it is because the architecture does not work well for gaming (Threadripper 1000 and 2000, fixed with the 3000 series) or there are not enough cores, something the site addresses in the build guide. When that happens the average FPS also are not great when looking at enough games, which the benchmark does for almost every modern processor.

It would not make a difference when judging Intel vs AMD or Ryzen 3000 vs Intel 9000 series.

The source articles for the benchmarks are linked though, often enough they also have those 1% or 0.1% benchmarks.

IIRC correctly the 1% and 0.1% frame times tell you how the frame rate varies i.e. if most of the time it is running at 60FPS and every so often tanks to 1 FPS. So some average FPS numbers might look okay but the frame rate might be tanking like every minute.

its close any way you cut it. theres a billion benchmarks out there with the data you want

> You still can't beat intel at the high end. Both in terms of heat and performance.

This sentiment is almost a year out of date. Since Zen 2 launched, Intel has a measurable disadvantage in IPC. Intel's single-thread performance is only partially won back with higher clock speeds at the expense of far greater power usage.

>Games are traditionally single threaded and hard to parallelize.

ECS and being able to safely run "systems" within your game concurrently is becoming a thing rapidly as "lists of objects" do not scale.

I just checked Doom Eternal: it's running 81 threads. (Of course, not all of them are always active or doing things concurrently. It's just a datapoint.)

Intel basically only wins by a non-significant margin at video encoding nowadays thanks to AVX512 which AMD does not implement. AMD wins or is comparable at a lower TDP on everything else.

I read a Twitter thread the other day about id Tech 7 (the new engine underpinning Doom Eternal). Apparently there is no “main thread”, everything is scheduled to a set of jobs that can be executed concurrently.

https://twitter.com/axelgneiting/status/1241487918046347264?...

I’ve been wondering if this will become de facto standard for multi-threaded/multi-core systems going forward.

The EA/DICE Frostbite engine has worked like this for many years now.

Learning about ECS and how to use them made gamedev so much easier for me. Although now I'm hamstrung by only being able to work on games in languages/frameworks with a solid ECS library (does a good/community agreed upon one exist for rust yet?).

ECS, the observer pattern, and behavior trees are probably the three main things that answered all the "how on earth do they build something this complex" questions for me.

> (does a good/community agreed upon one exist for rust yet?).

I just ran across this: https://csherratt.github.io/blog/posts/specs-and-legion/

> does a good/community agreed upon one exist for rust yet?

Specs? At least that's what Amethyst uses.

I would add state machines to any list of amazing techniques

Definitely. Behaviour trees are kind of a special case/evolution of regular ol state machines, but I find both useful in different situations.

I was considering learning them soon for enemy ai. Are they useful in other situations too?

Basic state machines are useful for controlling flow between different menus/game screens - this "current mode" state exists across iterations of the game loop so you need some kind of data based state machine.

e.g. in say a jrpg you have "new game/continue" menu -> world map -> battles and character menu and back again.

Edit: oh I misunderstood you - you were asking about behaviour trees. I haven't used them for anything other than controlling game unit behaviour (although not necessarily just enemy ai - one other example is using them to combine a bunch of simple actions into one thing that the user can tell a unit to do - so like an attack move in an RTS is actually "loop (is enemy nearby? -> no -> move towards location -> yes -> attack (which is probably it's own behaviour tree))

From what I have seen, the single threaded tasks are slightly better on intel but even in this linked article the AMD chips are winning slightly in single threaded. And in multi threaded the AMD CPUs absolutely blow away intel by a huge margin at a lower price.

> Games are traditionally single threaded and hard to parallelize.

That's some dinosaur games from the past. Modern games are using all available CPU cores with something like Vulkan and saturate GPU fully. So Intel CPUs are left in the dust with all the cores you get from high end Ryzens.

Yeah, this was true for basically everything like, 12? years ago. But it has been steadily getting less and less true as time goes on. Doesn't stop it being repeated constantly though.

There should be a formalized law for this - anything that was once widely known as true (because it was) will continue to be treated as gospel far after it ceases to be the case.

In technology, it seems to be the case that once some fact is old enough to be known as gospel, the world will have moved on enough that it's not true anymore.

And when challenged by the new norm, the old guard repeat their gospel louder and in greater numbers! So if you hear some technological rule of thumb repeated often enough, you should doubt it even more than if you hadn't heard it at all.

>Games are traditionally single threaded and hard to parallelize.

It is not that the Gamers and Prosumers dont know.

But they are willing to trade ~10% of Single Thread Performance for Double the Core at a similar or less price point.

Sounds like a very reasonable decision to me.

Also games are hardly limited by CPU at all a 10% loss in single threaded won't be noticed at all but more cores will make a huge difference for other tasks.

Tell that to Cities Skylines or any game based on Unity

Unity is transitioning to DOTS, a vastly more performant and parallelisable architecture. Obviously this isn't relevant to legacy titles, but "games are mostly single-threaded" is increasingly a thing of the past.

https://unity.com/dots

And Unity is pretty much dead last in that spot too. Remember that both the PS4 and Xbox One are (formerly) high core AMD cpus. Anything that runs on them can handle multi-threaded processors pretty well

I haven't played in a while but I was happily playing it on an i7 4770k which is way under what both Intel and AMD sell now.

It doesn't appear to scale past a certain size of city very well. Lot's of people get best case baseline fps of around 50. Great game for the most part though.

Add Ashes of the Singularity to that list too, great worst case for benching gaming rigs cpu bottleneck.

> Games are traditionally single threaded and hard to parallelize.

Citation needed. Games are about the least "Traditional" software engineering culture on the planet. We've had multicore CPUs for the last two gens of consoles (and I've seen multi-core programming on the Saturn and N64). On PC we've had mainstream multicore since 2009-ish. That's about when I left games. But given the need to win big on consoles, I can't imagine anyone has a main loop anymore that does everything on one thread.

What N64 had multicore?

>> traditionally

> anymore

> does everything on one thread

Multithreading is useful for concurrent programming separately from its value in parallel programming.

The N64 had two MIPS processors.

It did not.

Yeah, it really did. The RSP was just another MIPS processor that couldn't access main memory (like the SPEs on the PS3). Usually developers treated it like a graphics card, feeding it display lists that were interpreted and rendered by Nintendo's own microcode. But if you got hold of the RSP Dev Kit, then you could write your own microcode. And if you were really ambitious, you could write an MPEG player that used both the RSP and the main CPU to deliver FMV off a cartridge. That was Todd [0]. That's why N64 emulators couldn't run Resident Evil 2 for the longest time: they had native implementations of the RSP-as-a-graphics-card, and just drew the display lists. But RE2 didn't use the normal microcode, and it couldn't be played until someone wrote an emulator for the RSP.

[0] https://ultra64.ca/files/other/Game-Developer-Magazine/GDM_S...

What N64 had multicore?

Most CPU reviews are showing up to twice the heat or more for around 5% better performance on single-cpu loads.

Most games are lightly threaded, with newer games being much better optimized for 6+ cores.

Even for Gaming, only the 9900K/KS are even in consideration for a good choice, and if you're doing anything besides gaming or quicksync loads, odds are you're better off with AMD. The 3900X is less expensive and does as good or better in most workloads, and even in gaming, most wouldn't be able to tell the difference.

As to best of the best for higher end, the Threadripper has no peer.

I wonder what games really stress the CPU these days and I wish to get some example beyond dwarf fortress of games that actually stutter without a top end system, I got a midrange one built around a old core i5 and a modern Radeon 570 and even physic bound games like beam.ng and games with lots of agents like arma 3 run just fine.

Battlefield V likes as much CPU power as it can get. But it also likes GPU...

There is a free to play game Armored Warfare (crysis engine?), which I genuinely used to test stability of my cpu. 8 threads fully loaded and probably with some form of AVX.

Reviews of upcoming Mount&Blade also note how it loads 16+ threads (mostly for NPC on battlefield).

Arma3 was notorious for being bound by single thread performance and low latency memory.

https://www.youtube.com/watch?v=Gw_RH0RQ2oA

on a old midrange i5 6500 b5 does a solid 30+ fps on ultra and cpu load doesn't top out so I find the claim that it needs a top of the line cpu very dubious. I'll try armored warfare since it's free.

so just tested out armored warfare and at full had maximum quality I get 60fps with 70% cpu utilization on a i5 6500 with the rx570 gpu

If you are interested in cpu being utilised, should lower graphics setings from ultra. On ultra GPU is the main bottleneck. Try same on Low settings

There is some difference. The most marked difference is between dual core, and quad core. There are quite a few youtube videos out there.

https://www.youtube.com/watch?v=1eMh-SZQuWM

yeah so it just does 80-90 fps? I'm not talking about "any difference", I'm talking games right now that require a high end cpu, since everyone upward in the thread was claiming gamers need top end cpu, and I can't think of a game that stutters with a mid range one.

also, the fps reduction seems more in line with the low end intels having less pci lanes, and not the game being cpu capped.

Unless you're a competitive gamer, and young with very good vision, keeping above 100fps at 1440p or better, you're very unlikely to notice the difference.

Very few people genuinely notice the difference. I ran a 3600 while waiting for a 3950X, and if I had it to do over, would probably have just gone 3900X to start with... For the most part, I don't notice the difference, only after I'm running a couple databases and a few services that are active to I even come close to seeing a difference in general use.

I know GPU and Monitor options are similar, and CPU/GPU combination in terms of balance are important. Unless you're going RTX 2080 or higher, then it's much less of a difference in general use anyway.

the new ryzens compile gcc things faster, even on one thread. and you get more cores per dollar all the way up

I would think as of now, the main lead that Intel has in with AVX-512 computations.

Reply on news.ycombinator.com