Speculation: Ryzen 3000 series

[inf64]

Allegedly, Francois Piednol really wanted Intel to do just that. Years ago. He was rebuffed.

I checked FP's twitter account, dude is suffering from serious case of cognitive dissonance. His reality is vastly different of reality everyone else experience. It's funny how now he does not mention any of the "low clocks" and "Chinese are not impressed with Zen2" comments he made 4 months ago. He finally realized that was nonsense. It's a start at least.

 

[exquisitechar]

People are still expecting up to 5GHz on all-core OCs? Sorry, but I'm calling it now. That isn't happening. Not until Zen 3.

The only Zen 2 chip I saw at 5GHz was on LN2 at 1.6V.

(I would be happy to be wrong)

I'm not even expecting 5GHz on a single core, let alone all. Down the line, super binned Threadripper 3 might boost to 4.8GHz, I guess. I'm expecting a bigger variation in OC potential between different samples/SKUs than the previous gens.

I checked FP's twitter account, dude is suffering from serious case of cognitive dissonance. His reality is vastly different of reality everyone else experience. It's funny how now he does not mention any of the "low clocks" and "Chinese are not impressed with Zen2" comments he made 4 months ago. He finally realized that was nonsense. It's a start at least.

Don't forget about his 1.4GHz base clock prediction for Rome and his gloating when low power engineering samples popped up that matched them...only for 225W Rome to be revealed to have a ~2.25GHz base clock. It was impossible, he said, it would break the laws of physics, he said! After that, he was silent about the matter and turned to spreading FUD about the Infinity Fabric instead...

I believe him about him suggesting that Sunny Cove should be backported to 14nm, though.

 

[Kenmitch]

People are still expecting up to 5GHz on all-core OCs? Sorry, but I'm calling it now. That isn't happening. Not until Zen 3.

The only Zen 2 chip I saw at 5GHz was on LN2 at 1.6V.

(I would be happy to be wrong)

It's just a numbers game in the end. The only thing that really matters is the overall performance.

I'm looking forward to seeing the reviews. All eyes on AMD when they drop.

 

[Tup3x]

So does that mean it has much lower latency than Zen 2000 series ?

I think it means next Threadripper is going to be awesome.

 

[DAPUNISHER]

Anyone post this yet?

https://browser.geekbench.com/v4/cpu/13698306

 

[realibrad]

Anyone post this yet?

https://browser.geekbench.com/v4/cpu/13698306

I'm trying to figure out what a stock 9900k would score, but, I'm seeing a mountain of different postings of scores. Is there a ballpark score for the 9900k at stock?

 

[exquisitechar]

I'm trying to figure out what a stock 9900k would score, but, I'm seeing a mountain of different postings of scores. Is there a ballpark score for the 9900k at stock?

I think it's something like 6200-6300, not really sure. It depends on a lot of things, but the overall score in GB is useless, IMO. Better to look at the sub-scores and individual tests. From what I can see, if this result is indicative of the average, I don't think that the 3900x will match the 9900k in GB single core, ignoring the memory score.

At this point, though, I'm kind of done with the hundred different GB results out there. We'll get all the answers soon enough.

 

[lobz]

I think it's something like 6200-6300, not really sure. It depends on a lot of things, but the overall score in GB is useless, IMO. Better to look at the sub-scores and individual tests. From what I can see, if this result is indicative of the average, I don't think that the 3900x will match the 9900k in GB single core, ignoring the memory score.

At this point, though, I'm kind of done with the hundred different GB results out there. We'll get all the answers soon enough.

@ stock it's in the lower 6100s

 

[realibrad]

I think it's something like 6200-6300, not really sure. It depends on a lot of things, but the overall score in GB is useless, IMO. Better to look at the sub-scores and individual tests. From what I can see, if this result is indicative of the average, I don't think that the 3900x will match the 9900k in GB single core, ignoring the memory score.

At this point, though, I'm kind of done with the hundred different GB results out there. We'll get all the answers soon enough.

That is roughly what I was expecting. Thanks.

 

[exquisitechar]

@ stock it's in the lower 6100s

Could be, I don't keep track of the average overall single core scores that much. My overall comparison still stands, though. I looked at multiple 9900k results to draw my conclusion, but like I said, we'll need to wait and see.

 

[IEC]

The really interesting bit is that the combination of central IO die, top binning for Threadripper chips, and CCX-aware scheduler in Windows 10 build 1903 means there is a potential that Threadripper could become the king for gaming chips this generation. Probably not in value, but in chasing that last 5% of performance there certainly seems to be that potential. Whether or not it can be realized remains to be seen. Just like max turbo is increased in the 3950X I imagine there will be Threadripper bins capable of 4.8GHz low core count turbo.

 

[exquisitechar]

Interestingly, if you look at that recent huge GB single core score of the 3600, Zen 2 seems to have higher IPC in GB than Skylake when running on Linux, but not under Windows 10. When running on Linux, Ryzen gets a huge boost in both the SC and MC score in GB. While in Cinebench, Zen 2 firmly seems to have 10% higher IPC than SKL.

Sunny Cove seems to have ~10% higher IPC than Zen 2 in Geekbench, by the way.

 

[maddie]

So does that mean it has much lower latency than Zen 2000 series ?

The purpose of cache is to mitigate or compensate for latency issues. The doubled size cache in Zen2 will allow better performance even if latency remains the same.

I have some issue with measuring raw latency without also compensating for larger and/or better cache. It gives a false impression to stress latency values alone without seeing the complete picture.

 

[PotatoWithEarsOnSide]

Indeed.
The latency number on its own means very little. If you're having to go to memory 36-40% less often then Zen 2 will effectively the same overall latency as the best Intel has to other.
The question for me is whether it is possible to measure, under a specific load, how often the CPU is forced to go to memory.

 

[maddie]

Indeed.
The latency number on its own means very little. If you're having to go to memory 36-40% less often then Zen 2 will effectively the same overall latency as the best Intel has to other.
The question for me is whether it is possible to measure, under a specific load, how often the CPU is forced to go to memory.

In the interview with The Full Nerd, Robbert Hallock claimed ~ 15-20% increase in game performance just by the increased cache size and before any additional IPC improvements. Although to be honest they are also intertwined.

 

[PotatoWithEarsOnSide]

I'd love to know how they isolate it in order to be able to say that. Are they able to limit them to 2MB per core to have a direct comparison?
It is nice though, and explains a whole lot about how they managed to turn the CS:GO tables.

 

[DiogoDX]

Don't know if this is relevant here:
https://pc.watch.impress.co.jp/docs/news/1193327.html

TSMC recent talk about 7nm, 7nm+ and 5nm process.

 

[DrMrLordX]

I checked FP's twitter account, dude is suffering from serious case of cognitive dissonance. His reality is vastly different of reality everyone else experience. It's funny how now he does not mention any of the "low clocks" and "Chinese are not impressed with Zen2" comments he made 4 months ago. He finally realized that was nonsense. It's a start at least.

Yeah he's a character all right. I do think he might have gone into a board meeting at Intel (when he worked there) and popped off about porting Sunny Cove to 14nm though. It's the kind of thing he might do.

@IEC

I haven't seen much evidence for 5 GHz overclocks either. Only saw some commentary that the 6-core chiplets can actually get there and that 8-core chiplets can't (without exotic cooling). I definitely would not expect it on air, nor would I expect it from the 8-core chiplets under water (except maybe 3950X, due to binning).

 

[Mopetar]

People are still expecting up to 5GHz on all-core OCs? Sorry, but I'm calling it now. That isn't happening. Not until Zen 3.

The only Zen 2 chip I saw at 5GHz was on LN2 at 1.6V.

There was a recent result that got 5.2 GHz on LN2, but it's hard to draw many conclusions from one or two results.

From the data that Silicon Lottery has collected, only 10% of 9700k CPUs can hit 5.2 GHz, 36% (26% if you discount those that can go higher) can hit 5.1 GHz. Most (89%) will be capable of 5.0 GHz though.

It's likely that there's a somewhat similar curve for maximum OC potential for Matisse as well. If you just have one (or two results) it's not likely that you'll have received a golden sample and seen the potential.

At least we're close to release and will find out soon.

 

[jpiniero]

The really interesting bit is that the combination of central IO die, top binning for Threadripper chips, and CCX-aware scheduler in Windows 10 build 1903 means there is a potential that Threadripper could become the king for gaming chips this generation. Probably not in value, but in chasing that last 5% of performance there certainly seems to be that potential. Whether or not it can be realized remains to be seen. Just like max turbo is increased in the 3950X I imagine there will be Threadripper bins capable of 4.8GHz low core count turbo.

I think you are going to be sorely dissapointed. Strictly AMD vs AMD, the gaming results should be pretty much identical from all 5 chips at release, beyond clocks. For reviews, I guess it's a question of what the sustained clocks are at stock settings, but since they are all unlocked it's kind of moot. Threadripper shouldn't be any worse however, as it is now.

 

[darkswordsman17]

Don't know if this is relevant here:
https://pc.watch.impress.co.jp/docs/news/1193327.html

TSMC recent talk about 7nm, 7nm+ and 5nm process.

N6 is EUV? I'm confused now as it seems worse than 7+ (which claims 1.2x density, while they're claiming 1.18 for 6). I thought 6 was just going to be an improvement to the base 7 (so no EUV). Are they doing something where EUV is being used to improve the normal 7 node (but doesn't need any special work to reap the benefits)?

I think you are going to be sorely dissapointed. Strictly AMD vs AMD, the gaming results should be pretty much identical from all 5 chips at release, beyond clocks. For reviews, I guess it's a question of what the sustained clocks are at stock settings, but since they are all unlocked it's kind of moot. Threadripper shouldn't be any worse however, as it is now.

I think that's somewhat the point. The chips with more cores also seem to be the highest clocked (at least stock, we'll see how things work out in the real world and overclocking). I wouldn't be surprised to find that they bin the highest clocking chips for Threadripper. Combine it with the ability to potentially distribute the heat across a larger area, and the extra memory channels (I won't be surprised if they go to 6 or even the full 8 on some specs), and I could definitely see this next gen of Threadripper being the absolute best for even gaming. Obviously you'll get diminishing returns hard, so plenty of people will be happy with even a 6 core for gaming.

 

[Kenmitch]

I think that's somewhat the point. The chips with more cores also seem to be the highest clocked (at least stock, we'll see how things work out in the real world and overclocking). I wouldn't be surprised to find that they bin the highest clocking chips for Threadripper. Combine it with the ability to potentially distribute the heat across a larger area, and the extra memory channels (I won't be surprised if they go to 6 or even the full 8 on some specs), and I could definitely see this next gen of Threadripper being the absolute best for even gaming. Obviously you'll get diminishing returns hard, so plenty of people will be happy with even a 6 core for gaming.

I'd never buy one, but the next Threadripper will most likely be a beastly offering for those whom need the cores. Maybe they'll do something special like a 5GHz boost under 8c, 12c, 16c loads?

 

[RTX2080]

https://translate.google.com/transl...mod=viewthread&tid=2011003&page=1#pid42540506

1. almost every result about Zen2 that already leaked is not accurate, due to Resonant Clock being locked, boost frequency is not accurate.
2. 8 core Zen2@4.5Ghz against 9900k@5Ghz has similar performance
3. 4.8Ghz below 1.4v is easy except 3700 (and 3800 maybe)
4. clock mesh (similar to intel vid?) is not open yet, so ES OC limit is not accurate

 

[Markfw]

Guys, since pre-orders start Monday, do you think reviews will be out then also ? I need to unlock the review thread the night before the reviews come out, since somebody will beat be to it that morning.

 

[Karnak]

Pretty sure it's monday for pre-orders and 7/7 for reviews.