AMD Ryzen 9 3900X, SMT on vs SMT off, vs Intel 9900K

[W1zzard]

By community request, we present our findings on how the AMD Ryzen 9 3900X performs with SMT disabled. This approach has potential, especially for gaming, because it ensures more physical hardware units are available for each thread, and could also benefit the processor's power management.

Show full review

 

[TheRagnarok]

Thanks! Glad to see this although I went with the 3700X.
Still good to know.

 

[ShurikN]

Just saw Hardware Unboxed video on the same topic with 36 games and essentially in some of them you get performance, in some you lose. The majority is within 2-3%.
And in the end, they all kinda average out at 1% difference. Which is negligible.
Basically it depends on the game, but in the grand scheme of things it doesn't matter...

 

[KarymidoN]

Zen is trully impressive on some aplications, SMT is really making a difference in some work aplications, its really clear now that if you do anything other than pure game on your PC, AMD is the clear choice, Intel must adjust prices.

 

[Deleted member 158293]

Guess this idea useful information, and Zen's tiny overhead is impressive. Seems people who just want to play games should simply opt for a smaller core count CPU to begin with.

 

[bug]

With SMT disabled, the IPC of the "Zen 2" core seems to go up by roughly 2 percent. This directly impacts performance of software that don't need too many cores, namely games. We theorize this could be due to two distinct things happening with the processor. For one, with the SMT off its shoulders, the power-management of the processor is left to spread its power-budget and Precision Boost headroom across fewer logical processors, so each of the cores runs at higher boost frequencies, positively impacting per-core performance. Secondly, to a smaller extent, performance is also benefited by the on-core schedulers not having to juggle resources between two logical processors.

Just nitpicking here, but it's the performance that goes up by ~2%. Because if it's the former, it's the same IPC combined with slightly higher clock speeds

 

[Manu_PT]

Anyone else noticed how the real "winner" here gaming wise appears to be the i5 9400F? 5% slower than zen2, 9% slower than 9700k/9900k, costs 140€....... Really impressed with these charts. Media seems to ignore this little chip, but damn it has a great perf vs price ratio.

Glad you included it in your tests! Thanks

 

[moproblems99]

Did I read the graphs wrong or is it showing the 9900k being less power hungry?

 

[Metroid]

Thanks for the test, smt off is only good if you are sure you will play a game that benefits from it for a long time, beginning to end, around 20 hours or more. So you dont need to go all the time into bios and change it. Few games benefit, nowadays many game engines have added smt in its code to benefit from it. Few games have the inverse, their 1% lows are much deeper if smt is off. Few people on reddit are planning a list of games that benefit from having smt off.

in 2008 when hyper-threading was released, smt off was a must, it was a huge performance impact leaving it enabled on games, you had to turn it off, 5 to 25% or so. Windows and game devs have improved the use of smt a lot.

Anyone else noticed how the real "winner" here gaming wise appears to be the i5 9400F? 5% slower than zen2, 9% slower than 9700k/9900k, costs 140€....... Really impressed with these charts. Media seems to ignore this little chip, but damn it has a great perf vs price ratio.

Glad you included it in your tests! Thanks

Since its release, the 9400f is still the best cost benefit modern cpu.

 

[catulitechup]

very good tests

maybe add same tests between 9700K / 3700X and 9600K / 3600X will be good

sadly amd dont offer cpus without ht and lower price example: 3600X without HT around 150us maybe awesome

 

[Khilos]

If your trying to squeeze the most performance out of the 3900x, you can overclock each CCX in Ryzen Master individually rather than an all-core OC. You'll get 5 to 10% more performance versus the all-core overlock or the PBO/AOC combo. Typically CCX 0 and CCX 1 overclock to 4.5 GHz and CCX 2 and CCX 3 overclock to 4.2 / 4.3 GHz.

 

[Deleted member 158293]

very good tests

maybe add same tests between 9700K / 3700X and 9600K / 3600X will be good

sadly amd dont offer cpus without ht and lower price example: 3600X without HT around 150us maybe awesome

To be fair AMD doesn't charge extra for SMT on their CPUs, so turning it off technically doesn't cost you anything. It's only relative to Intel's higher cost for an HT option that it looks like it should be cheaper.

 

[dirtyferret]

Anyone else noticed how the real "winner" here gaming wise appears to be the i5 9400F? 5% slower than zen2, 9% slower than 9700k/9900k, costs 140€....... Really impressed with these charts. Media seems to ignore this little chip, but damn it has a great perf vs price ratio.

Glad you included it in your tests! Thanks

Actually I would say the real winner is the PC gamer as the Intel i5-8400/9400F and AMD Ryzen 2600 both give you over 100FPS in virtually every game while costing $140 or less. Everything else is basically fan boys (with some enthusiasts) arguing over how much extra juice they can squeeze out of their lemon to either fit their specific niche and/or justify their ego & purchase.

 

[bug]

Actually I would say the real winner is the PC gamer as the Intel i5-8400/9400F and AMD Ryzen 2600 both give you over 100FPS in virtually every game while costing $140 or less. Everything else is basically fan boys (with some enthusiasts) arguing over how much extra juice they can squeeze out of their lemon to either fit their specific niche and/or justify their ego & purchase.

Well, there are those that do stuff besides gaming on their rigs

 

[Wavetrex]

Not worth turning it off unless low-thread software (or games) is being used most if not all the time.
AMD's SMT implementation is simply amazing (and is rumored to get even better with Zen 3 ... possibly 3-4 thread/core)

 

[IceShroom]

@W1zzard can you drop 720p benchmark, as i haven't seen a monitor with that resulation and replace that with 768p.

 

[dirtyferret]

Well, there are those that do stuff besides gaming on their rigs

very true but the the topic was gaming

 

[bug]

Not worth turning it off unless low-thread software (or games) is being used most if not all the time.
AMD's SMT implementation is simply amazing (and is rumored to get even better with Zen 3 ... possibly 3-4 thread/core)

I doubt they can pull that off. That many cores would need a hell of an instruction decoder. Plus, the more cores (physical or logical) the sooner you become bandwidth constrained.

 

[Wavetrex]

I doubt they can pull that off. That many cores would need a hell of an instruction decoder. Plus, the more cores (physical or logical) the sooner you become bandwidth constrained.

Threadripper has sufficient bandwidth.

TR 4000 series... 32 cores - 128 threads ( 4 chiplets + quad channel IO die) - Aimed at professionals.
Possible ? Of course !
The question is... will they actually do it ?

Mainstream (AM4) very likely won't need that, and will stay as the "gamer" CPU, with less but faster cores, and max 2T/c

 

[A.Stables]

Smells funny

 

[Metroid]

@W1zzard can you drop 720p benchmark, as i haven't seen a monitor with that resulation and replace that with 768p.

720p like 1080p, 480p, 1440p is standard.

 

[Dante Uchiha]

@W1zzard Could you add the 1% low to the charts? The analysis would be more interesting with this information.

 

[HD64G]

Just saw Hardware Unboxed video on the same topic with 36 games and essentially in some of them you get performance, in some you lose. The majority is within 2-3%.
And in the end, they all kinda average out at 1% difference. Which is negligible.
Basically it depends on the game, but in the grand scheme of things it doesn't matter...

That review has exactly the same results as the @W1zzard 's one as in the 1080P tests the difference is less than 1% on average. Good to see something so repeatable and clear.

 

[Xaled]

When multithreading really counts, an SMT acts as half a core, in Rendering/baking tests 12cores/24 threads is AVG. %50 fasters than 12cores. That's really impressive imo.

 

[guser]

well , just cant belive that 7nm amd cpu loose intel 14nm cpu??!!!
intel win.

hope techpower put same test then, when intel release comet lake and of coz 1when intels 10nm intel ice lake is out.

again, excellent test and,tx!