AMD Ryzen 9 5900X CPU-Z Bench Score Leaks, 27% Higher 1T Performance Over 3700X

[Dredi]

Because the "IPC" gains cannot be normalized for multithreaded workloads. Not only does every core clocks differently but also without fixed clocks, which I doubt this test was run on, you're just relying on the "TDP" & the boost algorithm to measure the performance, not to mention different motherboards make the test even more unreliable. For almost any workload you can think of, the IPC gains will not be 1:1 coming from ST to MT tasks. Though the length of the tests also matter, without fixed clocks & TDP restrictions (removed) the bigger core siblings will be at a disadvantage in short duration benches.

IPC cannot be ’normalized’ for single threaded applications either, because it is always by definition an application specific metric.

 

[R0H1T]

Right & in this case we're speculating how it affects CPU-Z Bench.

 

[Dredi]

Right & in this case we're speculating how it affects CPU-Z Bench.

WCCFTech compares it to 3900X too:

"As for performance, the chip scored 652.8 points in the single-core test which is 27% faster than the AMD Ryzen 7 3700X and up to 25% faster than the Ryzen 9 3900X.

Coming to the multi-threaded performance test, the alleged AMD Ryzen 9 5900X CPU scored a total of 9481 points which is a massive 75% improvement over the Ryzen 7 3700X & a 15% improvement over the Ryzen 9 3900X."

Why only 15% higher score than 3900X in multicore? That's lower than the alleged IPC uplift - for almost 50% higher TDP, presumably. Something's not right.

Zen2 had uplift in single and in multicore compared to Zen1 or Zen+. 2700X to 3700X had 12% increase in single core CPU-Z score (457 to 509), but 13% increase in multicore (4839 to 5465).

It's strange that 3900X to 5900X wouldn't follow the same pattern, especially with the TDP and frequency uplift. I could imagine this result if processor was thermally or power limited.

This time the IPC gains of about 10% come from the CCX structure changes that affect single and multi threaded workloads differently. For single threaded workloads you effectively have double cache resources at your disposal, where as the MT workload has about the same as before per thread. Depending on if the worker threads share data there could be some benefit from the restructuring, but in this case I think that might not be the case and the workers in the MT workload do not share data and thus the IPC gain would be the about 10% lower than in the ST workload.

 

[Bwaze]

That's an enormous difference just for the cache size.

 

[Dredi]

That's an enormous difference just for the cache size.

Yup, but 32 megs of cache for a single thread is a lot compared to 16 megs on zen2. On MT workloads the cache size is the same per thread as it was on zen2.
It might also be possible that they have figured out how to pump more juice to a single core, but have kept the TDP the same, meaning that the single core performance would be higher but MT would be proportionally more power limited.

 

[TheoneandonlyMrK]

Shit, n I wasn't going to upgrade, might have to

 

[jesdals]

Would like to know the Infinity fabric speed of these 5xxx cpu’s

 

[Xzibit]

So this person can get his hands on unreleased hardware but can't update to the latest CPU-Z in a year

 

[Vayra86]

If that single thread is noticeably higher than what Intel does on CFL, I'm getting rid of this setup and upgrading too, most likely.

This 8700K still isn't my favorite CPU. Still does odd things now and then, spiky voltage/temp behaviour at times, its hard to put the finger on it, but meh - and the OC potential just isn't there unless I go exotic on cooling measures, which I'm not doing.

 

[Calmmo]

3900x, I'm sorry but i don't think we were meant for each other, it's not because of you, it's me..

If that single thread is noticeably higher than what Intel does on CFL, I'm getting rid of this setup and upgrading too, most likely.

This 8700K still isn't my favorite CPU. Still does odd things now and then, spiky voltage/temp behaviour at times, its hard to put the finger on it, but meh - and the OC potential just isn't there unless I go exotic on cooling measures, which I'm not doing.

oh boy, i hope you are prepared for the zen3 drama then, it was nuts with zen2 voltages last year for the first 2-3 months

 

[Vayra86]

oh boy, i hope you are prepared for the zen3 drama then, it was nuts with zen2 voltages last year for the first 2-3 months

Yeah... def waiting for dust to settle. I'm also looking at Ampere... damn what a shit show.

As far as OC goes... I've let go of that idea, its not worth it anymore. They can do that for me now, apparently

 

[Calmmo]

Yeah, OC is very different to what I was used to from my quad core i7 days, it's all about per ccx oc as quality per ccx varies wildly. You go from cores that can do 4.4 at 1.275 to cores that cant even do 4.3 at 1.4v.

 

[Xaled]

If the 652 single-core score is true, the whole current cpu lineup is KOed..
So I'll take that with a pinch of salt

 

[R0H1T]

Which lineup?

 

[TheLostSwede]

I've been predicting for months that between the IPC increase, frequency increase/node improvement, and the doubling of cores per CCX, that the core for core performance lift could be upwards of 30%, although this isn't total confirmation, it's in the right direction.

TPUs own analysis showed that the performance difference between the 3100 and the 3300x is 12% due to the respective two vs one CCX design, and since then I've postulated that this (along with the cache redesign, 17%-20% IPC gains and 200-300mhz frequency bump), the total performance uplift could be upwards of 30%...Can't wait to upgrade my 2700x for a 5900x and my X470 for an x670

There won't be any new motherboard chipsets for the 5000-series CPUs.

 

[FinneousPJ]

About 85% over stock R7 1700. Yeah, I think I can justify the upgrade

 

[Berfs1]

Yeah when CPU-Z 1.93 is out, definitely use version 1.90 for a "leak"!

 

[Chrispy_]

Hope this transaltes into significantly better gamingperformance. Wonder how far we get on ram support, since Renoir APUs can du 2200-2300 infinity fabric I`m hoping Zen 3 can do atleast that.

RAM support is still something that AMD needs to work on. I still regularly encounter RAM that simply doesn't work with an XMP setting. I know XMP is intel-optimised timings but when you get a common, affordable DDR4-3200 kit and it's not happy on Ryzen, your average Joe is going to be running it at 2133 JEDEC default and horrible 1066MHz Infinity Fabric, simply because the timings for XMP aren't quite there on the AMD platform.

Given that 95% of RAM kits available to consumers rely on XMP, it kind of sucks that the XMP settings still don't work for a non-trivial number of combinations. Sure, most TPU readers know to plug in the timings from the 1usmus DRAM calculator to get their AMD systems running at or close to the advertised speeds, but I'm willing to bet that most users aren't going to be interested in a 3-step process that involves manually inputting dozens of secondary and tertiary RAM timings buried in a completely different place under advanced BIOS settings for each motherboard.

Only this week I encountered a mainstream Gigabyte B550 board that failed to run XMP settings on two different RAM kits, and earlier during lockdown, Asus and Asrock B450 boards choked on a popular Kingston 3200 kit but were happy with (supposedly terrible for AMD) Corsair LPX and budget Patriot 3600 kits respectively.

AMD really needs XMP to "just work" and their AGESA firmware should really really loosen timings by 25% or so on the memory training before giving up and running 3600MHz DDR4 at almost half it's rated speed.

 

[Dredi]

RAM support is still something that AMD needs to work on. I still regularly encounter RAM that simply doesn't work with an XMP setting. I know XMP is intel-optimised timings but when you get a common, affordable DDR4-3200 kit and it's not happy on Ryzen, your average Joe is going to be running it at 2133 JEDEC default and horrible 1066MHz Infinity Fabric, simply because the timings for XMP aren't quite there on the AMD platform.

Given that 95% of RAM kits available to consumers rely on XMP, it kind of sucks that the XMP settings still don't work for a non-trivial number of combinations. Sure, most TPU readers know to plug in the timings from the 1usmus DRAM calculator to get their AMD systems running at or close to the advertised speeds, but I'm willing to bet that most users aren't going to be interested in a 3-step process that involves manually inputting dozens of secondary and tertiary RAM timings buried in a completely different place under advanced BIOS settings for each motherboard.

Only this week I encountered a mainstream Gigabyte B550 board that failed to run XMP settings on two different RAM kits, and earlier during lockdown, Asus and Asrock B450 boards choked on a popular Kingston 3200 kit but were happy with (supposedly terrible for AMD) Corsair LPX and budget Patriot 3600 kits respectively.

AMD really needs XMP to "just work" and their AGESA firmware should really really loosen timings by 25% or so on the memory training before giving up and running 3600MHz DDR4 at almost half it's rated speed.

I don’t think AMD can affect the XMP profiles memory manufacturers decide to put on their modules. Because of that the only way XMP could ”just work” on AMD is if they commit to designing their memory controller to be compatible with the same timings as the intel memory controllers, which is a silly proposition. The way forward is (sadly) QVL lists for memory, until memory manufacturers stop making their XMP settings incompatible with AMD memory controllers.

 

[cueman]

nice but,true performance we see then when intel get ready even 10nm cpu....and what leaks and so on, amd ryzens not shine so bright then.

its useless comapare 7nm cpu for 14nm cpu...and specially tell it 'fast'.


few month and we see intel 10nm side and 6 month we see equal battle

 

[Dredi]

nice but,true performance we see then when intel get ready even 10nm cpu....and what leaks and so on, amd ryzens not shine so bright then.

its useless comapare 7nm cpu for 14nm cpu...and specially tell it 'fast'.


few month and we see intel 10nm side and 6 month we see equal battle

Few months, according to what? There are no 10nm desktop processors in any timetable.

 

[Steevo]

WCCFTech compares it to 3900X too:

"As for performance, the chip scored 652.8 points in the single-core test which is 27% faster than the AMD Ryzen 7 3700X and up to 25% faster than the Ryzen 9 3900X.

Coming to the multi-threaded performance test, the alleged AMD Ryzen 9 5900X CPU scored a total of 9481 points which is a massive 75% improvement over the Ryzen 7 3700X & a 15% improvement over the Ryzen 9 3900X."

Why only 15% higher score than 3900X in multicore? That's lower than the alleged IPC uplift - for almost 50% higher TDP, presumably. Something's not right.

It might be a lot of clock speed for the single core increase.

 

[Sandbo]

I thought I didn't have to upgrade from my 3600X.

 

[Fleurious]

We must be getting close to actual benchmarks, wasn’t there supposed to be something in October for Zen3 and big navi?

Hopefully, these leaks are true as it will be nice to see them compete in high refresh rate gaming.

 

[Space Lynx]

We must be getting close to actual benchmarks, wasn’t there supposed to be something in October for Zen3 and big navi?

Hopefully, these leaks are true as it will be nice to see them compete in high refresh rate gaming.

october 7th/8th for cpu and 27th and 28th for gpu. you will see official benches then