AMD's Upcoming $750 Ryzen 9 3950X (16C, 32T) Shown Beating Intel's $2,000 i9-9980XE (18C, 36T)

[windwhirl]

I find hilarious that the uploader of these benchmark results is called "blueleader". Oh, the irony...

 

[Hifihedgehog]

I find hilarious that the uploader of these benchmark results is called "blueleader". Oh, the irony...

And we all know how well things went for Blue Leader. That's why you never see Blue Squadron at the Battle of Yavin.

 

[RichF]

Unless the latest Intel hardware mitigates against the security flaws and their performance regressions then people should take those into full account for all of these comparisons as well.

For example, having to disable hyperthreading, which several companies have said is necessary for full security, is a big blow to some of Intel's parts. Some people will respond by saying they're not that worried about security. Well, maybe they aren't but some are and should be.

 

[nemesis.ie]

Regarding the comments on the IF divider, 3733 is the last step at 1:1, the next one up is 2:1. So 3733 is considered the "sweet spot" unless you have really fast RAM and can get the latency settings way down.

Additionally, there should be enough IF b/w at 2:1 as it's supposed to be 2.5 or 3x the speed of Zen+ IIRC. Even if it's "only" 2x, it would be no worse at the 2:1 divider but the RAM (and the IF) will still be going faster vs Z+.

I can't wait to play with it.

 

[Captain_Tom]

Unless the latest Intel hardware mitigates against the security flaws and their performance regressions then people should take those into full account for all of these comparisons as well.

For example, having to disable hyperthreading, which several companies have said is necessary for full security, is a big blow to some of Intel's parts. Some people will respond by saying they're not that worried about security. Well, maybe they aren't but some are and should be.

It's a blow to ALL of Intel's parts. I have noticed in my system (It is Intel) a performance regression over the past 12 months. That ~10% drop in my applications is actually VERY noticeable.

What people don't seem to realize is that AMD is clearly still holding back some of their cards. That benchmark where they showed Intel without hardware patches was very intentional - they are excited for the final reviews to make Intel look even worse than they did.

 

[svan71]

I hesitant to believe this just yet. The 18 core Intel should be a least negligible faster than the AMD 16 core counterpart.

Perhaps when they get a non ES chip running 200 MHz faster than the one tested you will believe it.

 

[Tomgang]

I found a result with an Ryzen 5 3600 in geekbench.

Gigabyte Technology Co., Ltd. X570 I AORUS PRO WIFI - Geekbench

Benchmark results for a Gigabyte Technology Co., Ltd. X570 I AORUS PRO WIFI with an AMD Ryzen 5 3600 processor.

browser.geekbench.com

 

[ArchStupid]

Sorry but no, I just find it hilarious the lengths certain folks go to defend Intel. I buy and build from both camps.

No one even mentioned Intel here, you're delusional.

 

[windwhirl]

What people don't seem to realize is that AMD is clearly still holding back some of their cards. That benchmark where they showed Intel without hardware patches was very intentional - they are excited for the final reviews to make Intel look even worse than they did.

I just imagined Dr. Su rubbing her hands together with a smirk on her face lol

 

[Alpha_Lyrae]

Regarding the comments on the IF divider, 3733 is the last step at 1:1, the next one up is 2:1. So 3733 is considered the "sweet spot" unless you have really fast RAM and can get the latency settings way down.

Additionally, there should be enough IF b/w at 2:1 as it's supposed to be 2.5 or 3x the speed of Zen+ IIRC. Even if it's "only" 2x, it would be no worse at the 2:1 divider but the RAM (and the IF) will still be going faster vs Z+.

I can't wait to play with it.

Yeah, IF bus width has been increased from 256-bit in Zen/Zen+ (AM4 only) to 512-bit (same as Zen/Zen+ Threadripper). So, IF is pretty well overprovisioned vs DDR4 in 1:1 mode. The numbers seem a bit ridiculous until you consider that IF SDF also handles all other communication too, which can saturate the bus internally. Once RAM access is needed, that becomes the primary performance limiter. Ryzen seems ready for DDR5, honestly.

1866.5 (1:1 mode) * (512/8) * 2 lanes = 238.912GB/s
- Maximum DDR4 bandwidth at 1866 (3733MT): 1866.5 * (128/8) * 2 = 59.728GB/s
- IF bandwidth is well in excess of maximum DDR4 speed

2000/2 (1:2 mode) = 1000 * (512/8) * 2 lanes = 128GB/s
- Maximum DDR4 bandwidth at 2000MHz (4000MT): 2000 * (128/8) * 2 = 64GB/s
- bandwidth isn't an issue, but since IF speed is halved, latency increases slightly

These numbers exclude overhead costs.

I'm still not clear on if the SDF is a 512-bit bidirectional link (2 lanes). I'm assuming it is for this exercise.

 

[xpredator_13]

No one even mentioned Intel here, you're delusional.

Take this man to the infirmary.

 

[Zotz]

"Almost as big a win as that huge stone on Lisa's hand."

The most photographed gem since the Hope diamond.

 

[fireedo]

it is overclocked but, at least with the same geekbench application (different OS I know) 7980XE win a quite much

7980XE Geekbench result

 

[phill]

it is overclocked but, at least with the same geekbench application (different OS I know) 7980XE win a quite much

7980XE Geekbench result

I think the clock speeds where different to what they are meant to be released at... But it's a good result

 

[HwGeek]

Do we know if 3950X reviews are coming on 7/7? since Der8auer already got his
I know it will be out only on September

 

[windwhirl]

Do we know if 3950X reviews are coming on 7/7? since Der8auer already got his
I know it will be out only on September

Did he now? Huh.

Well, he may have had to sign a NDA, so he probably won't tell us much until we are closer to launch. Besides, I'd wait until at least a few major tech sites have their reviews ready.

 

[medi01]

I hesitant to believe this just yet. The 18 core Intel should be a least negligible faster than the AMD 16 core counterpart.

Because we have never seen AMD beat Intel while on the same process, or because TSMC process node is not ahead of Intel's.
Oh wait.

 

[Frick]

"Almost as big a win as that huge stone on Lisa's hand."

What?

 

[HUSKIE]

AMD Myrtle?

Sounds like nearby same as our street "Myrtle" Road Sheffield UK. LOL

 

[RichF]

It's a blow to ALL of Intel's parts. I have noticed in my system (It is Intel) a performance regression over the past 12 months. That ~10% drop in my applications is actually VERY noticeable.

Which is why I continue to make the point that people need to keep these performance regressions front and center in their minds and in their writing, when they do evaluations. Too often I see presentations that seem to be based in an alternate reality where there are no performance regressions.

Some have seemed to suggest that the latest Intel CPUs have less performance degradation because there are some baked-in mitigations, in hardware. I want clarification about that in any performance comparison between Intel and AMD until it's no longer relevant. I want hyperthreading to be disabled on any Intel parts that require that for full security, in any performance comparison, so people can see what the impact is.

What people don't seem to realize is that AMD is clearly still holding back some of their cards. That benchmark where they showed Intel without hardware patches was very intentional - they are excited for the final reviews to make Intel look even worse than they did.

That's nice for that artificial financial instrument's empty pursuit of profit, I suppose. I'd rather have reality taken into full account, not marketing magic hold the day. Right now, reality appears to be that we have serious performance regressions due to serious security vulnerabilities. Those need to be taken into account in all performance appraisals. Now, not later.

 

[Berfs1]

2063 MHz is 4133 MT/s (DDR4-4133). What is truly interesting is that AMD has shown that performance may actually be reduced by running the memory at anything over about 3600 to 3733 MT/s (1800 to 1866 MHz).

The problem with that is, other geekbench submissions show the full ram speed and not the halved