AMD Ryzen 9 3950X breaks three benchmarking world records

[HEXUS]

LN2 cooled CPU becomes 16C champ in Cinebench R15 / R20, and Geekbench 4 multi.

Read more.

[scaryjim]

Loving the retro Athlon 64 sticker on the LN2 flask

[Tabbykatze]

Those scores are very impressive, even if the Cinebenches are slightly AMD leaning over Intel.

[Zak33]

Loving the retro Athlon 64 sticker on the LN2 flask

i thought that too... I havfe one of those stickers somewhere

[Zak33]

Cinebench R20: AMD Ryzen 9 3950X @ 12167 points
Previous holder: Intel Core i9-7960X @ 10895 points

that has to hurt.....11.67% faster.

[eHM]

Are these results post or pre security patches?

Big numbers talk louder than words

[Tabbykatze]

Are these results post or pre security patches?

Big numbers talk louder than words

Because it's Intel i would expect they had them turned off. Considering their Ts&Cs basically say testing has to be off.

[Irien]

Whilst cinebench might be seen as AMD-favouring, it is more a case of it being a thread-friendly program that is also cache-friendly. The fact that the previous record-holders were listed as Intel chips not threadrippers shows that Intel CPUs also excel in the benchmark - it is just well written/designed for modern CPUs rather than favouring one vendor over another.

Where Intel has traditionally been stronger are less well designed titles that hammer single threads at the expense of load spreading - it is common to find one demanding thread bottlenecking things, whilst the other cores are less heavily loaded. Such titles (eg. older games) are where Intel seem to be throwing down the gauntlet at the moment (presumably because they are feeling threatened in other areas). Newer games tend to be more thread-aware due to consoles having large numbers of slower cores, demanding programmers learn to adapt their coding styles.

What is significant to me is that the AMD parts seem to be excelling without obvious drawbacks - Threadripper was/is great, but seemed to need a lot of "fixes/modes" to cope with its more unique architecture. So far, it would appear that these high-core-count Ryzen CPUs are less fussy and generally just perform well. However, the proof will be in the eagerly-awaited Hexus review(s)!

[KultiVator]

Yep - the new Ryzen's look to have matured nicely - I'll be watching for Hexus to put the 12 core part through the mill. I'm looking forward to building a new system that can chew through video rendering tasks in a fraction of the time I've been used to on my 8+ year old Intel 2700K rig - and still play the odd game when I'm in need of some downtime.

I'm less certain about the AMD RDNA video cards though - my guess is that these will not be powerful enough to really exploit the PCI4.0 bus or offer much of a performance or price differential over their closest Nvidia counterparts. Would love to be wrong - and to be honest, I don't like Nvidia's current (rip-off pricing) direction, so I may be tempted go with a tactical vote once some authoritative reviews are in.

An interesting few months ahead - and AMD has definitely pulled a blinder not just in the PC space, but in being at the centre of all the hardware driving the upcoming Xbox, PlayStation and Google Stadia offerings.

[Xlucine]

Where Intel has traditionally been stronger are less well designed titles that hammer single threads at the expense of load spreading - it is common to find one demanding thread bottlenecking things, whilst the other cores are less heavily loaded. Such titles (eg. older games) are where Intel seem to be throwing down the gauntlet at the moment (presumably because they are feeling threatened in other areas). Newer games tend to be more thread-aware due to consoles having large numbers of slower cores, demanding programmers learn to adapt their coding styles.

I disagree on the cause - console developers have had to contend with the old AMD CPU cores for ages (with the original PS4/xbone combo released in 2013, both of which offer 8 naff cores), yet it's only recently that games have started showing advantages for >4 cores. It looks more likely that desktop ryzen making >4 core CPUs widely available is what made developers start using more than 4 cores

What is significant to me is that the AMD parts seem to be excelling without obvious drawbacks - Threadripper was/is great, but seemed to need a lot of "fixes/modes" to cope with its more unique architecture. So far, it would appear that these high-core-count Ryzen CPUs are less fussy and generally just perform well. However, the proof will be in the eagerly-awaited Hexus review(s)!

One of those issues was just due to software not expecting so many cores, with some applications giving up if they saw more than 20 cores in the system (which is why they needed to turn off parts of threadripper for some games). That's not a fault with the design, and the fix is due to software updates not improvements in zen2 silicon.

We also don't know how much of an issue going across CCXs is, until the reviews hit

[watercooled]

Whilst cinebench might be seen as AMD-favouring, it is more a case of it being a thread-friendly program that is also cache-friendly. The fact that the previous record-holders were listed as Intel chips not threadrippers shows that Intel CPUs also excel in the benchmark - it is just well written/designed for modern CPUs rather than favouring one vendor over another.

Where Intel has traditionally been stronger are less well designed titles that hammer single threads at the expense of load spreading - it is common to find one demanding thread bottlenecking things, whilst the other cores are less heavily loaded. Such titles (eg. older games) are where Intel seem to be throwing down the gauntlet at the moment (presumably because they are feeling threatened in other areas). Newer games tend to be more thread-aware due to consoles having large numbers of slower cores, demanding programmers learn to adapt their coding styles.

What is significant to me is that the AMD parts seem to be excelling without obvious drawbacks - Threadripper was/is great, but seemed to need a lot of "fixes/modes" to cope with its more unique architecture. So far, it would appear that these high-core-count Ryzen CPUs are less fussy and generally just perform well. However, the proof will be in the eagerly-awaited Hexus review(s)!

It's not just down to threading either. You also have to factor in that for things like games, aside from being lightly threaded, they have typically been targeted at Intel CPUs given that's where the market was for quite some time. Architecture-specific optimisation can make a substantial difference in performance. As an extreme example, check pretty much any of the old Pi-calculating apps vs y-cruncher - it even has a dispatcher to choose the optimum binary depending on which CPU family it detects (the blog makes for some very interesting reading). The compiler itself (one of which is produced by Intel) can make a difference WRT optimisations too, Ryzen didn't exist when many older games still widely used for benchmarking were compiled so no family-specific optimisations could have been implemented. Ryzen is a proper ground-up design, not an iteration of previous designs - Intel haven't done such a thing in an awful long time, so it's competing on a playing field where in some cases you're effectively comparing code optimised for one family, yet totally unaware of another.

[KultiVator]

From my (limited) understanding, the biggest issue with ThreadRipper CPUs is the 'fabric' used to connect the multiple CPU dies to each other and to the memory channels... or in simpler terms, not all parts of the CPU have direct access to each area of memory. This can result in a thread on one part of the CPU having to pull the contents of *distant* memory over a relatively 'slow' link, whilst other threads might be talking to *close* memory over a faster connection... resulting in some oddball performance inconsistencies.

I gather that the optimisations AMD and Microsoft have introduced, try to align threads and memory more carefully, thereby avoiding some threads getting a raw deal.

Looks like the Zen2 architecture has made some positive steps forward... but it'll be good to hear what Hexus have to say once they've had some review units in the lab for a while.




[QUOTE=Xlucine;4105070]
One of those issues was just due to software not expecting so many cores, with some applications giving up if they saw more than 20 cores in the system (which is why they needed to turn off parts of threadripper for some games). That's not a fault with the design, and the fix is due to software updates not improvements in zen2 silicon.

[afiretruck]

From my (limited) understanding, the biggest issue with ThreadRipper CPUs is the 'fabric' used to connect the multiple CPU dies to each other and to the memory channels... or in simpler terms, not all parts of the CPU have direct access to each area of memory. This can result in a thread on one part of the CPU having to pull the contents of *distant* memory over a relatively 'slow' link, whilst other threads might be talking to *close* memory over a faster connection... resulting in some oddball performance inconsistencies.

I gather that the optimisations AMD and Microsoft have introduced, try to align threads and memory more carefully, thereby avoiding some threads getting a raw deal.

Looks like the Zen2 architecture has made some positive steps forward... but it'll be good to hear what Hexus have to say once they've had some review units in the lab for a while.

That kind of architecture is known as Non-Uniform Memory Access, or NUMA. It's Typically seen in multi-socket systems (servers, high-end workstations, someone's mad science experiment, etc), where the latencies between NUMA domains are much more pronounced than between the Zen core complexes. NUMA was the biggest issue for Threadripper (not so much for Ryzen, but still a problem) as most games and software assumed it was just a single, fat memory domain. It's less of an issue for EPYC as server workloads tend to (not always, mind) be lots of separate tasks, instead of one task with lots of threads, so they can be scheduled to remain within their own NUMA domain, instead of trying to spread out among all of them. A dual-socket EPYC system can have as many as 8 domains. The high core-count Threadrippers (32,24) have 4 - two without direct access to memory (ouch).

Judging by the architecture of Zen 2, they'll remove the NUMA issue entirely, as all memory channels are now connected to a single, central controller, so access characteristics are the same, regardless of which core is talking to which bit of memory.

[KultiVator]

That kind of architecture is known as Non-Uniform Memory Access, or NUMA. It's Typically seen in multi-socket systems (servers, high-end workstations, someone's mad science experiment, etc), where the latencies between NUMA domains are much more pronounced than between the Zen core complexes. NUMA was the biggest issue for Threadripper (not so much for Ryzen, but still a problem) as most games and software assumed it was just a single, fat memory domain. It's less of an issue for EPYC as server workloads tend to (not always, mind) be lots of separate tasks, instead of one task with lots of threads, so they can be scheduled to remain within their own NUMA domain, instead of trying to spread out among all of them. A dual-socket EPYC system can have as many as 8 domains. The high core-count Threadrippers (32,24) have 4 - two without direct access to memory (ouch).

Judging by the architecture of Zen 2, they'll remove the NUMA issue entirely, as all memory channels are now connected to a single, central controller, so access characteristics are the same, regardless of which core is talking to which bit of memory.

Many thanks for the fuller explanation afiretruck... I am optimistic AMD will convince me to part with some cash this summer ;-)

[philehidiot]

Many thanks for the fuller explanation afiretruck... I am optimistic AMD will convince me to part with some cash this summer ;-)

Same here. I've only recently added more DDR3 RAM in the expectation that the CPU market was going to remain stagnant. Unfortunately, not only has my relative performance declined with these new chips, absolute performance has been hammered due to multiple Intel vulnerabilities being patched. It looks like AMD have now got on board with Microsoft to deal with many of the performance issues which is encouraging.

The simple fact is there is nothing in the current Intel range which appeals to me. Then you look at AMD and there's a chance that their AM4 socket will still be in use next time I want to upgrade as well as things like PCI-e 4 which is useless now, but I'm buying for a few years down the line.

AMD are buying a lot of consumer good will and confidence and on top of that, decent products.

I strongly suspect that once the platform has matured a little and when the benchmarks come out proving its worth, I'll be upgrading mobo, CPU and RAM. Ryzen chip, lower RAM prices and mobos with forward looking features... the only thing I need is a mobo that also will take my old PCI sound card....

[afiretruck]

Many thanks for the fuller explanation afiretruck... I am optimistic AMD will convince me to part with some cash this summer ;-)

No worries! I hope you enjoy your new build

... the only thing I need is a mobo that also will take my old PCI sound card....

They do exist. At least there are B350 boards with a PCI slot or 2 in the mix. My friend had the exact same need when he bought his Ryzen 2 years ago. However, you may struggle to find an X570 board with one.

Though I strongly suspect, unless you specifically require the unique features of your sound card, you'll find the motherboard's onboard audio to be more than acceptable.