News AMD previews Ryzen 3rd generation at CES

[Hitman928]

More info to come. . .

 

[tamz_msc]

You can bet your shirt that they selected the best combination of hardware that made the 9900k look as weak as possible,just as they did at the original presentation of ZEN.
Being on par in cinebench doesn't mean much for userbench is what I'm saying,the 2700 especially the x was already pretty much on par,just minus 5% against the 9900k at stock.
https://www.techpowerup.com/reviews/Intel/Core_i9_9900K/9.html

It's funny that you don't read your own sources thoroughly because if you did you'd know that the 9900K in the TPU review was not maintaining boost clocks, hence making the gap to the 2700X look closer than it really is under proper turbo configurations.

 

[JoeRambo]

So you are saying the benchmarks that were shown today at CES are fake ? That the new cores are now NOT on par with the 9900k ?

We don't really need AMD siege mentality anymore in 2019, there is not a single mention of word "fake" in his thread. I think TheELF was simply pointing out that we were shown Cinebench only. And it is reasonable to assume that AMD has chosen the best case scenario for their chip.
Cinebench scales real great on systems with horribad memory hierarchies ( compare TR2 2950X vs TR2 2990WX ) and it is obviously not that sensitive to memory latency and/or bw. So at least in Cinebench everything is fine for new chip, IPC and clocks are there at nice power consumption.

But a proper (CES spirited) test for new chip would have been some CPU limited gaming scenario, AMD has just moved I/O and MC further away from cores, that's where THE real questions are.

 

[Arzachel]

But a proper (CES spirited) test for new chip would have been some CPU limited gaming scenario, AMD has just moved I/O and MC further away from cores, that's where THE real questions are.

Cinebench mimics a real world workload and scales well enough that the results can be extrapolated to other CPU-heavy tasks pretty reliably. It's also well-known and has a massive database of results. You propose they instead use a CPU limited gaming scenario where the results are by nature going to be extremely application-specific, a huge pain to verify and revelant only to a very small subset of consumers.

Might as well ask them for javascript benches

 

[JoeRambo]

Cinebench mimics a real world workload and scales well enough that the results can be extrapolated to other CPU-heavy tasks pretty reliably. It's also well-known and has a massive database of results. You propose they instead use a CPU limited gaming scenario where the results are by nature going to be extremely application-specific, a huge pain to verify and revelant only to a very small subset of consumers.

Might as well ask them for javascript benches

Why not show them all? I have nothing against CB, but i feel it plays into AMD core strength too much, while hiding memory latency/bw effects too much. Why not add some gaming, GeekBench 4 into the mix? And JavaScript is fine to me as well, i am certain ZEN2 will shine in it due to great cache subsystem.

 

[Arzachel]

Why not show them all? I have nothing against CB, but i feel it plays into AMD core strength too much, while hiding memory latency/bw effects too much. Why not add some gaming, GeekBench 4 into the mix? And JavaScript is fine to me as well, i am certain ZEN2 will shine in it due to great cache subsystem.

I also want a pony for my birthday but expecting a full suite of benchmarks 5-6 months ahead of launch is a bit unrealistic when they don't even have the final silicon yet.

 

[Abwx]

..................................

Why not show them all? I have nothing against CB, but i feel it plays into AMD core strength too much, while hiding memory latency/bw effects too much. Why not add some gaming, GeekBench 4 into the mix? And JavaScript is fine to me as well, i am certain ZEN2 will shine in it due to great cache subsystem.

Although being mainly a FP test CB use instructions up to SSE4.1/2, no AVX and other FMA to skew the IPC comparison, as such it s not too far from usual integer code based benchs when it comes to execution improvements.

Also the tests you re suggesting could be actually too favourable if ever they improved parameters like latency, in this case some games could see an higher uplift/Hz than CB, much like Zen+ whose perf/Hz gained 3% in CB and 5% in games...

 

[scannall]

I mean according to AMD the first ZEN was already supposed to blow intel out of the water,they made it look sooooooo gooooood against intel hedt in their presentation.

According to AMD the first Zen was supposed to be %40 better than the construction cores. Which would have put it about Ivy Bridge. Instead, it was %52 faster. They sandbagged Ryzen 1, not over-hyped it. Forum enthusiasts hyped it to the moon, but AMD didn't.

I have to wonder if they are sandbagging here too? They did say it wasn't running at final clocks. And whatever clock they chose was just enough to pass the 9900K. We shall see when the parts hit the market.

 

[dacostafilipe]

Why not show them all?

They used CB for all demos since the release of Zen, how are people not expecting them to continue with this trend?

 

[JoeRambo]

They used CB for all demos since the release of Zen, how are people not expecting them to continue with this trend?

And how it did turn out for Ryzen? It turned out quite deficient in gaming, while having almost a parity in CB per watt. Sounds like best case scenario to me. Widely known, but irrelevant for general performance.

 

[PotatoWithEarsOnSide]

General performance does not equal gaming.

 

[JoeRambo]

General performance does not equal gaming.

Nor does CB.

 

[Arzachel]

Nor does CB.

This is a silly both-sides-ism.

I mean, I would love to see the latency numbers cause de-integrating the mem controller is a bold move but a) it's way too early and b) the result would be meaningless for the vast majority of users who will be GPU limited anyways.

 

[coercitiv]

I have to wonder if they are sandbagging here too? They did say it wasn't running at final clocks. And whatever clock they chose was just enough to pass the 9900K. We shall see when the parts hit the market.

That's exactly what's happening, the Matisse sample was running at roughly 35W lower power than nominal values for AM4 TDP. We have every reason to believe this can translate into another 10% performance at launch, that is unless they unleash the power beast (ignore TDP on X570 boards) like Intel did with 9900K, in which case we're likely to be bound by 7nm max frequencies and can expect higher than +10% performance.

General performance does not equal gaming.

Even if it did, we're bound to see performance gains there as well. Even if it's just the equivalent of 2700X running at 4.7Ghz, it still means a healthy 10% gain.

The thing with this CES demo is AMD showed us they either have a very strong IPC gain or a very healthy frequency gain. It can be one of these, a mix of these, but cannot be none of these.

The CB15 demo may be less relevant for overall performance than we'd like (considering we have yet to know arch specifics in relation with latency bound loads), but what it does tell us is AMD has a strong advantage in manufacturing and looks poised to bring significant all-round performance increase through brute force if they have to.

 

[Atari2600]

This is AMD showing that they can beat Intel with one hand behind their back.

Quite literally if those power consumption and performance are accurate.

Wonder how much of the CPU power consumption is in the chiplet and how much is in the I/O. Could they double the core count and stay within a 125W envelope with frequencies very similar to this?

 

[IRobot23]

I believe those were system power draw, not CPU power draw.

Well I think that was CPU power.
Anandtech said 170W for CPU i9 9900K at 4,7GHz. They are basically contradicting themselves with AMD article.
https://www.anandtech.com/show/13400/intel-9th-gen-core-i9-9900k-i7-9700k-i5-9600k-review/21

SO can someone explain?

Spoiler: She does say system power indeed

We have to understand that I/O die probably takes up 20-30W so ryzen CPu die is eating only 45-55W? Seriously?

If that I/O die is design for low latency then probably it could burn up to 40W easily or even more (remember if they went for low latency). While cores to around 90W which is actually not a lot... the chip has 32MB of L3 running above 4GHz. And for whole package only 75W ( as they say) yeah, that is amazing.

i9 9900K is running at 4,7GHz on Core+L2 and 4,3GHz on L3. Remember than. Putting 16Mb of L3 to 4,7GHz will burn more than just few W.

 

[coercitiv]

Could they double the core count and stay within a 125W envelope with frequencies very similar to this?

Based on the numbers Anandtech gave us (75W) and assuming 15W for the I/O they could probably fit the second chiplet with only a 5% drop in clocks for this workload.

 

[Arzachel]

Well I think that was CPU power.
Anandtech said 170W for CPU i9 9900K at 4,7GHz. They are basically contradicting themselves with AMD article.

Direct quote from the Anandtech article: "The Intel system, during Cinebench, ran at 180W. "

 

[PotatoWithEarsOnSide]

That's exactly what's happening, the Matisse sample was running at roughly 35W lower power than nominal values for AM4 TDP. We have every reason to believe this can translate into another 10% performance at launch, that is unless they unleash the power beast (ignore TDP on X570 boards) like Intel did with 9900K, in which case we're likely to be bound by 7nm max frequencies and can expect higher than +10% performance.


Even if it did, we're bound to see performance gains there as well. Even if it's just the equivalent of 2700X running at 4.7Ghz, it still means a healthy 10% gain.

The thing with this CES demo is AMD showed us they either have a very strong IPC gain or a very healthy frequency gain. It can be one of these, a mix of these, but cannot be none of these.

The CB15 demo may be less relevant for overall performance than we'd like (considering we have yet to know arch specifics in relation with latency bound loads), but what it does tell us is AMD has a strong advantage in manufacturing and looks poised to bring significant all-round performance increase through brute force if they have to.

Very much so.
What interests me is that with Zen+ a 3% CB improvement resulted in a 5% gaming improvement, primarily because the gains were down to improved memory latencies.
If we're going to see the same with Zen 2, then I suspect that the uplift in gaming will be pretty significant. As you said, the clocks alone will provide a chunky uplift, moreso if the ST boost is up near that 5.0GHz mark.
My only concern really us that IF changes haven't really been talked about much anywhere, as there were rumours of decoupling IF clocks from Memclock. If that is the case, then we're definitely likely to see memory latencies decrease, even if somewhat offset by having gone the MCM route.
I don't envision the Ryzen 3xxx CPUs to be behind in gaming at all, except in some very niche cases.

 

[PotatoWithEarsOnSide]

Based on the numbers Anandtech gave us (75W) and assuming 15W for the I/O they could probably fit the second chiplet with only a 5% drop in clocks for this workload.

The other aspect is that they were likely running increased voltages for stability as it was a live demo. It's possible that production SKUs consume even less power. That being said, the ES was not at final clocks, so a bit of give and take in terms of what production SKUs might be capable of.

 

[IRobot23]

Direct quote from the Anandtech article: "The Intel system, during Cinebench, ran at 180W. "

"by contradicting" I mean ther own test with i9 9900K eating on HW info64 just below 170W.

 

[TheGiant]

Quite literally if those power consumption and performance are accurate.

Wonder how much of the CPU power consumption is in the chiplet and how much is in the I/O. Could they double the core count and stay within a 125W envelope with frequencies very similar to this?

It is not clear IMO atm with the "beat". The power looks very nice, but we don't know the clocks. And with 7nm the die area is going to be a lot smaller than with 12nm so cooling the same amount of power will be much more difficult. I am not that convinced that intel or AMD can run their 10nm/7nm parts with 125W easily as now.
They should give us more versatility of the benchmarks as others pointed out.
So be on parity with performance with 9900K is not enough as a signal atm. AMD needs to give a very strong signal we are better with lot less power.
The higher power doesn't prevent sales, as was in the days of P4 or is now in GPU comparison of RX580 with gtx1060, where the RX card is simply awful.
So can be with new zen and 9900K, performance on parity while 9900K needs a personal reactor and people will buy both except pure enthusiasts.
AMD did a very strong signal with EPYC, where Intel is already covering in fear but this consumer part is too weak IMO.

 

[The Alias]

"by contradicting" I mean ther own test with i9 9900K eating on HW info64 just below 170W.

That's like a 6-7% uptick in power usage. Hardly contradictory when you haven't verified they're using the same tool to measure power usage.

 

[Abwx]

The other aspect is that they were likely running increased voltages for stability as it was a live demo.

That s about 100% sure, in the range of 15% excess power out of the 65W or so it consumed in the demo.

The 75W stated by AT is the estimated delta measured at the main assuming that system idle power was 55W, actually the delta should be a little higher at 84W, wich is 65W at the SoC level.

15% higher score than a 2700X at 65W is perfectly in line with 2x the perf/watt at isofrequency but not isoperf, the 15% delta in perf imply 0.5 x 1.15 = 0.575x the power, wich is exactly the ratio between this SKU and a 2700X, so much that this point to the test being done at 4GHz wich is the 2700X all core turbo in CB.

FTR the numbers stated by AMD imply that power increase by at least 16% from 4 to 4.2GHz, if it were such a frequency clocking the Zen 2 there would be no way that it could be at only 65W when we add an excess voltage, 15% higher throughput and an I/O still in 14nm.

 

[Arzachel]

"by contradicting" I mean ther own test with i9 9900K eating on HW info64 just below 170W.

The article you linked measures the maximum power draw with POV-Ray which uses AVX2 instructions while CB doesn't. Intel has always let AVX2/512 workloads run out of spec and the power draw numbers aren't indictive of general use.

 

[frozentundra123456]

Cinebench mimics a real world workload and scales well enough that the results can be extrapolated to other CPU-heavy tasks pretty reliably. It's also well-known and has a massive database of results. You propose they instead use a CPU limited gaming scenario where the results are by nature going to be extremely application-specific, a huge pain to verify and revelant only to a very small subset of consumers.

Might as well ask them for javascript benches

Wow, I was just playing cinebench last night. Really had a lot of fun. Cant wait to get home from work tonight and play it again. Guess I can ignore my whole steam library now. I have no problem with cinebench as a preliminary benchmark, but it *is* pretty much a best case scenario for AMD. We will just have to wait for more benchmarks, but gaming is still a question, and a primary use case for many (most) users. Final clocks may be better, but actually the only thing I am impressed with so far is the power consumption. By the time it comes out it will be almost a year after the 9900k, and to only match it in a best case artificial benchmark is underwhelming to me.