AMD Ryzen 9 9950X

[sblantipodi]

big flop imho

are there some AVX512 tests? I want to see the improvements in that part if any...

 

[Lycanwolfen]

Yep does not surprise me one bit. Intel with there stupid P core E cores.. Like Intel has such great potential to rule the desktop market yet has no vision to do so now. If they has a pure 16 core 32 thread cpu for desktop today I sure it would smoke this one. But alais I have pretty much giving up on intel. I switched to AMD because I tired of Intel no vision for desktop.

 

[A Computer Guy]

And so the plot thickens a little bit more:

Description: Windows Bug Found, Hurts Ryzen Gaming Performance

After watching this I couldn't help but think wow and a news timeline emerges in my imagination...
> gamers enable admin user for faster performance, months go by, then followed by gamers hit hard by malware that cripples systems because they were running admin user.

 

[Knight47]

Hey wait till you find out what you just got jumping from x370 to B550 or x570 all with the same socket

Honestly upgrading from one AM4 board to another feels like a waste of money. I wanted to upgrade from 1700+X370 to 3800+X570, but I could have jumped to 11th on 12th gen intel for the same price and get better i/o. So instead I got a 3800 for the same price as the 5800 because my board didn't support 5800 for almost 2years after it's release.

 

[gffermari]

Quantifying The AVX-512 Performance Impact With AMD Zen 5 - Ryzen 9 9950X Benchmarks Review - Phoronix

www.phoronix.com

 

[Dr. Dro]

big flop imho

are there some AVX512 tests? I want to see the improvements in that part if any...

The AVX-512 tests are pretty much the only thing that Zen 5 aces

Quantifying The AVX-512 Performance Impact With AMD Zen 5 - Ryzen 9 9950X Benchmarks Review - Phoronix

www.phoronix.com

And you just beat me by a few seconds LOL

Yep does not surprise me one bit. Intel with there stupid P core E cores.. Like Intel has such great potential to rule the desktop market yet has no vision to do so now. If they has a pure 16 core 32 thread cpu for desktop today I sure it would smoke this one. But alais I have pretty much giving up on intel. I switched to AMD because I tired of Intel no vision for desktop.

Not really, no... besides, Intel's last architectural revision for their desktop CPUs actually predates this chip's predecessor a little bit. That it's still holding its own in the charts is just testament that their design works, mishaps aside.

 

[sblantipodi]

Quantifying The AVX-512 Performance Impact With AMD Zen 5 - Ryzen 9 9950X Benchmarks Review - Phoronix

www.phoronix.com

I expected more even there... damn, zen 5 is really a flop.

 

[AnotherReader]

The AVX-512 tests are pretty much the only thing that Zen 5 aces



And you just beat me by a few seconds LOL



Not really, no... besides, Intel's last architectural revision for their desktop CPUs actually predates this chip's predecessor a little bit. That it's still holding its own in the charts is just testament that their design works, mishaps aside.

Besides AVX-512, Zen 5 does well in some other workloads, e.g. node.js, databases, and memcached, but it's a dud for gaming.

 

[Dr. Dro]

Besides AVX-512, Zen 5 does well in some other workloads, e.g. node .js, databases, and memcached, but it's a dud for gaming.

LOL, @W1zzard you might have some sleepless nights ahead of you mate

I can't even. Using the hidden sys admin account changes Zen 4 and 5 performance results, basically invalidating all reviews

I wonder if this affects Zen 3 and Intel too

 

[AnotherReader]

I expected more even there... damn, zen 5 is really a flop.

I think you had unrealistic expectations. Many workloads run into memory bandwidth limitations, e.g. y-cruncher. Compare single threaded performance for y-cruncher to multithreaded performance and you can see that bandwidth bottleneck in action.

LOL, @W1zzard you might have some sleepless nights ahead of you mate

I can't even. Using the hidden sys admin account changes Zen 4 and 5 performance results, basically invalidating all reviews

I wonder if this affects Zen 3 and Intel too

Using that hidden account permanently is not the most secure practice.

 

[Dr. Dro]

I think you had unrealistic expectations. Many workloads run into memory bandwidth limitations, e.g. y-cruncher. Compare single threaded performance for y-cruncher to multithreaded performance and you can see that bandwidth bottleneck in action.

View attachment 359146
View attachment 359147


Using that hidden account permanently is not the most secure practice.

It truly isn't, but seems to work around whatever problem is going on. I truly wonder if this is some xbox game bar related bug or if it's happening to just about everyone due to security policy, more research will be needed

 

[Chrispy_]

Why run 6000 MT/S for Zen 5 when TPU made this claim back in July 1, where 6400 MT/S is the "sweet spot" for Zen 5 with 1:1?
There's no way Zen 4 shares the exact same IMC as Zen 5 does now if this fact holds credit. The Zen architecture with DDR5 is very sensitive to performance hits if RAM is not tuned properly. Combine that with 1:1 & high but stable FLCK - it would be performing better than what this review & the previous reviews for 9600X + 9700X have reported.

The source for that repost is WCCFTech, and if you read the full article you get this, which I've partially bolded for emphasis:

First of all, the integrated memory controller for the AMD Ryzen 9000 "Zen 5" CPUs is similar to the Ryzen 7000 "Zen 4" CPUs but comes with slight refinements. We have been told that the CPUs will be able to support DDR5-5600 by default and up to DDR5-6400 memory at a 1:1 fabric clock. The sweet spot is still going to be DDR5-6000 1:1 but on both X670 and X870, the upper limit will be set at 6400 MT/s.

Since WCCFTech cited no sources, it was basically just unconfirmed rumour. WCCFTech are hit and miss, but given that even AMD didn't have the answers a week prior to launch, it's probably fair to say that anything published two months ago was probably little more than guesswork.

 

[gffermari]

I would like to see benchmarks using extreme memory modules like DDR5-8000 etc.
I know that most likely the X670s won't even boot. I just want to see how bad the ram bandwidth issue is.

 

[DemonicRyzen666]

The source for that repost is WCCFTech, and if you read the full article you get this, which I've partially bolded for emphasis

First of all, the integrated memory controller for the AMD Ryzen 9000 "Zen 5" CPUs is similar to the Ryzen 7000 "Zen 4" CPUs but comes with slight refinements. We have been told that the CPUs will be able to support DDR5-5600 by default and up to DDR5-6400 memory at a 1:1 fabric clock. The sweet spot is still going to be DDR5-6000 1:1 but on both X670 and X870, the upper limit will be set at 6400 MT/s.

But is that from Marketing team or Engineering team?

If it's bandwidth starved then latency won't really matter, & you would be better off with the 2:1 divided instead. I seen what happens when Zen 3 was using 4,000mhz ram it's usually slight slower 1% lows & 0.1% lows in games by like a 1% to 4% it's so small you'd never notice it really. Not unless your average lows were brough down in gaming.

 

[AnotherReader]

I would like to see benchmarks using extreme memory modules like DDR5-8000 etc.
I know that most likely the X670s won't even boot. I just want to see how bad the ram bandwidth issue is.

Testing something obviously bandwidth bound like y-cruncher should show benefits, but outside AVX-512, workloads are more likely to be bound by latency than bandwidth.

 

[Zach_01]

When the system runs DDR5 6000MT/s the speeds go like this, right?
FCLK:UCLK:MCLK
2000:3000:3000

This is 1:1 for UCLK:MCLK but the whole thing is 2:3:3
Best case scenario since AM5 and DDR5 introduced for Ryzen

Very few could manage to run 6400MT/s
2133:3200:3200 (2:3:3)

For 8000MT/s so far the best case is
2133:2000:4000
where UCLK:MCLK drops from 1:1 to 1:2 (cant be anything else than these 2)
and probably most systems would be on
2000:2000:4000
or somewhere between 2000~2133 for FCLK

The decoupling of UCLK:MCLK from 1:1 to 1:2 creates so much latency that anything the 8000MT/s has to offer gets diminished and the overall performance stays about the same.
At least for regular desktop loads including gaming.

 

[Daven]

LOL, @W1zzard you might have some sleepless nights ahead of you mate

I can't even. Using the hidden sys admin account changes Zen 4 and 5 performance results, basically invalidating all reviews

I wonder if this affects Zen 3 and Intel too

AMD issued a statement saying you should be getting 210 fps in Cyperpunk with the 4090 at 1080p. The review shows 186 fps. That’s 13% lower.

 

[Dr. Dro]

AMD issued a statement saying you should be getting 210 fps in Cyperpunk with the 4090 at 1080p. The review shows 186 fps. That’s 13% lower.

That's wayyy too broad a statement, though. 210 fps at which settings exactly, I haven't seen that specified.

 

[HTC]

AMD issued a statement saying you should be getting 210 fps in Cyperpunk with the 4090 at 1080p. The review shows 186 fps. That’s 13% lower.

They should be getting UP TO 210 FPS, but that's if the EXACT SAME SCENE is tested.

Different scenes = different FPS, no?

 

[rv8000]

When the system runs DDR5 6000MT/s the speeds go like this, right?
FCLK:UCLK:MCLK
2000:3000:3000

This is 1:1 for UCLK:MCLK but the whole thing is 2:3:3
Best case scenario since AM5 and DDR5 introduced for Ryzen

Very few could manage to run 6400MT/s
2133:3200:3200 (2:3:3)

For 8000MT/s so far the best case is
2133:2000:4000
where UCLK:MCLK drops from 1:1 to 1:2 (cant be anything else than these 2)
and probably most systems would be on
2000:2000:4000
or somewhere between 2000~2133 for FCLK

The decoupling of UCLK:MCLK from 1:1 to 1:2 creates so much latency that anything the 8000MT/s has to offer gets diminished and the overall performance stays about the same.
At least for regular desktop loads including gaming.

Some early results from some of the prominent memory OC’ers on OCN are showing the same to slightly better IMCs on their samples (some engineering and some retail).

8300 seems to be the highest successfully tested and ~4hr stability tested (vdimm @ 1.55 if i recall).

FCLK largely seems the same, maybe one notch higher on average; few samples doing 2233 there.

Slightly lower bandwidth than Zen 4 at similar memory speeds in AIDA64

 

[R-T-B]

It's not subjective. It's called cherry picking. You selected to show all the benchmarks with the least performance relative to the 7950X. How cunning of you but I'm not sure why you took so much time to copy and paste the worst data into your comment. Here is the chart for the rest of us not looking for the worst performance but performance across all apps. Take what you like from it but this is at least ALL the data versus the 7950X.

View attachment 358926

I mean I agree he's cherry picking, but it is to highlight a pretty legit issue. Frankly one or two regressions is understandable, but there are a lot more than that.

I'm hoping microcode fixes this because this is looking bad.

LOL, @W1zzard you might have some sleepless nights ahead of you mate

I can't even. Using the hidden sys admin account changes Zen 4 and 5 performance results, basically invalidating all reviews

I wonder if this affects Zen 3 and Intel too

This honestly stinks of an on-silicon bug of some kind. What kind of cpu can't handle permission levels well?

 

[mkppo]

I expected more even there... damn, zen 5 is really a flop.

For AVX512? It's a 27.3% performance increase at 12% less power vs Zen 4. With the same ish node and same die area, I'd say it's pretty great.

Even without specifically AVX512 workloads, they have a solid 18% gain across all tests with a ~10% power reduction to boot. Keep in mind this is without PBO, where Zen 5 gains another few % over Zen 4 (since it scales past the point where Zen 4 stops scaling).

Zen 5 for servers and workstations is anything but a flop IMO.

But is that from Marketing team or Engineering team?

If it's bandwidth starved then latency won't really matter, & you would be better off with the 2:1 divided instead. I seen what happens when Zen 3 was using 4,000mhz ram it's usually slight slower 1% lows & 0.1% lows in games by like a 1% to 4% it's so small you'd never notice it really. Not unless your average lows were brough down in gaming.

It's both IF and Memory bandwidth bottlenecked, y-cruncher shows the bandwidth bottleneck and games show the latency issue (both memory and inter CCD). Problem is, increasing memory bandwidth without increasing the IF doesn't bring about as much benefit as it should have but it still allows some nice gains to be had. What DDR5-8000 truly allows is lower effective latencies. When tuning for 6000Mhz, you quickly run into the latency floors for that speed but DDR5-8000 with relatively tight-ish timings allow both a reduction in latency and increase in bandwidth. Not saying it's easy, but I think that's what AMD are targeting with X870 boards.

I'm going to do the tests in a couple of months after X870 launch just to see the impact it has and how the scaling differs from the 7950X3D it will replace. My prediction is that when memory is tuned properly at 8000mhz, and a slight bump to IF, it should be pretty even or close to the 7950X3D in games.

 

[R0H1T]

That it's still holding its own in the charts is just testament that their design works, mishaps aside.

IPC gains outside FP/vector heavy computations have been few & far between even for Intel. Take AVX, AVX2 or AVX512 out & Intel's IPC gains would probably be singly digits in 10+ years. Although for some reason they've consistently(?) performed better than AMD in games. Maybe it's because of ringbus & much simpler uarch maybe it's a legacy thing? We'll get more answers as they introduce chiplets themselves.

 

[rv8000]

For AVX512? It's a 27.3% performance increase at 12% less power vs Zen 4. With the same ish node and same die area, I'd say it's pretty great.

Even without specifically AVX512 workloads, they have a solid 18% gain across all tests with a ~10% power reduction to boot. Keep in mind this is without PBO, where Zen 5 gains another few % over Zen 4 (since it scales past the point where Zen 4 stops scaling).

Zen 5 for servers and workstations is anything but a flop IMO.



It's both IF and Memory bandwidth bottlenecked, y-cruncher shows the bandwidth bottleneck and games show the latency issue (both memory and inter CCD). Problem is, increasing memory bandwidth without increasing the IF doesn't bring about as much benefit as it should have but it still allows some nice gains to be had. What DDR5-8000 truly allows is lower effective latencies. When tuning for 6000Mhz, you quickly run into the latency floors for that speed but DDR5-8000 with relatively tight-ish timings allow both a reduction in latency and increase in bandwidth. Not saying it's easy, but I think that's what AMD are targeting with X870 boards.

I'm going to do the tests in a couple of months after X870 launch just to see the impact it has and how the scaling differs from the 7950X3D it will replace. My prediction is that when memory is tuned properly at 8000mhz, and a slight bump to IF, it should be pretty even or close to the 7950X3D in games.

From past experience with the 7900X3D I use to have, the difference was pretty moot. 7800/8000 c36/38 was generally worse latency, 5-8% bandwidth increase and incredibly difficult to stabilize with the exception of using the Tachyon/Gene. Meanwhile 6400 c32 was much easier to run/stabilize (as well as getting others OCN’ers stabilized here), with better latency, and slightly less bandwidth at much more sane voltages.

Remains to be seen if DDR5 8000 is worth it, when something along the lines of 6400 with tight timings provides similar results. This is especially true when only very specific workloads or synthetics can take advantage of this; potentially useless for 90%+ of users.

 

[Dr. Dro]

I mean I agree he's cherry picking, but it is to highlight a pretty legit issue. Frankly one or two regressions is understandable, but there are a lot more than that.

I'm hoping microcode fixes this because this is looking bad.


This honestly stinks of an on-silicon bug of some kind. What kind of cpu can't handle permission levels well?

It'd be especially egregious considered Windows only utilizes two of the four security rings in x86 CPUs, the innermost "supervisor" and outermost "user" mode rings (rings 0 and 3). As I understand, though, even with operating under a root-level account, common applications will still be in user mode, there just wouldn't be anything to inhibit system-wide (including kernel, protected and reserved regions) access.

In the name of science, one can always try to run something with NT Authority permissions to see if that alone bypasses the problem...