Hacker News Reader

Comment by formerly_proven | Hacker News Reader

5900X beats a 10900K by 100 points in Cinebench R20 Single-Thread. 100. Points. That's 20 % higher per-thread throughput.

20 % IPC increase, 20 % higher perf/W on the same process, effectively double the cache size, and reduced memory latency. Absolute insanity from AMD.

I'm almost a little disappointed they didn't introduce a 5960X -- they could probably claim 2x-3x performance over the Intel part.

View on HN

Even more significant is that it beats the newest Intel Tiger Lake by 6% at equal clock frequencies (at 4.8 GHz).

That means that AMD will enjoy a better IPC than anything Intel for at least 1 year, until the end of 2021, when Intel will launch Alder Lake.

Intel Rocket Lake, expected in March 2021, cannot have an IPC better than Tiger Lake, so it will also have an IPC slower by at least 6% than Zen 3.

Because Zen 3 tops at 4.9 GHz, Rocket Lake will have to reach at least 5.3 ... 5.4 GHz to match or maybe exceed Zen 3.

Is Rocket Lake expected to scale up to 16 cores?

No, the top model of Rocket Lake will have only 8 cores & 16 threads.

Therefore, for multi-threaded tasks it will be much slower than Zen 3.

However, it is expected to have an IPC comparable with Ice Lake and Tiger Lake, so if it would reach 5.4 GHz, it might be a little faster than Zen 3 for single-threaded tasks and gaming.

Although everyone is saying that Intel is dead in single thread (which is definitely far more true than previously), I don't see 5.4GHz as infeasible for them - tenth gen goes up to 5.3GHz at the very high end and I would be surprised if they didn't reach at least 5.4GHz.

Looking historically, the 6700k reached 4.2GHz, the 7700k 4.5GHz, the 8700k 4.7GHz, the 9900k 5.0GHz, and the 10900k 5.3GHz - which would imply 5.5 or 5.6GHz for what will (presumably) be the 11900k.

Power consumption will probably curb that frequency hike above 5GHz.

Haven't people been saying that for a long time?

We've also been stuck at ~4GHz for a long time. The 10th gen Core only manages more because of their ridiculous 300W power consumption.

Everyone here is missing the most important concept in semiconductor manufacturing: binning and yield [1]. They're interconnected.

It's not like Intel can make 5.4 GHz chips for days on. It doesn't work that way. Ofcourse they can manufacture processors capable of some ridiculous 6.0 GHz (not accounting for power consumption)... but that would be 0.1% yield. You cannot run a factory like this.

[1] https://www.techspot.com/article/2039-chip-binning/

I'm after an 8/16 chip myself so that would work for me but Intel chips are DOA at this point.

5960X would be a Threadripper part (at least based on the releases of the 3000-series Ryzen desktop processors). Threadripper has typically lagged the Ryzen announcement. I think we can expect some insane TR parts soon.

The fact that they made of point of supporting threadripper in their announcement gives me a certain level of confidence there will be Zen 3 thread rippers. I mainly say this because people were curious if the 3950X would mean the end of the threadripper line, which was of course later proven wrong when the 24, 32, and 64 core TR parts came out.

AMD is clearly moving TR into the "Pro" / Workstation market. As such, "soon" means "the next Workstation trade event/show".

With CES 2021 cancelled/postponed for Jack knows how long, "soon" might mean Q2 2021.

Threadripper has always been for workstations, so they're not moving it anywhere.

Well, they have a “thread ripper pro” line which supports more RAM and official support for ECC.

Yeah, with no industry events, it sure was impossible for them to spawn a hundreds-of-comments HN thread with their announcement video yesterday....

5960X was also an Intel processor in 2014 (Intel Core i7-5960X @ 3.00GHz).

> 20 % IPC increase, 20 % higher perf/W on the same process, effectively double the cache size, and reduced memory latency. Absolute insanity from AMD.

An increase like this wasn't something unexpected from a generational update back in nineties.

But this is an architectural tweak on the same node. Zen2 was 7nm TSMC, while Zen3 is 7+nm TSMC.

TSMC must have optimized the heck out of their transistors to deliver such a large benefit. Or the AMD Zen 3 architecture really found some low-hanging fruit or something to grossly improve performance.

> low-hanging fruit or something to grossly improve performance.

If there’s such a thing anymore I’d be fascinated to hear about it.

Anandtech will have one of their detailed microarchitecture breakdowns at product availability day (November 5th).

The IPC increase seems to be coming from the microarchitecture, not the process. The Anandtech article includes a breakdown:

https://images.anandtech.com/doci/16148/1%20Slide%206_575px....

    +2.7% Cache Prefetching
    +3.3% Execution Engine
    +1.3% Branch Predictor
    +2.7% Micro-op Cache
    +4.6% Front End
    +4.6% Load/Store

No silver bullet there, just small gains in a bunch of areas. Some of those may have been hard wins too.

In the slides it was still marked as 7nm without the +.

It's much harder to do now

Plus 15% IPC and 10% more max boost, this checks out. I'd expect the 8-core part to have a higher multi-threading performance increase than the 12-core part, because it has just one CCX.

I think it's been neat from an architecture perspective for AMD. TSMC (and AMD) has had trouble matching the frequencies that Intel has been attaining on their 14 nm node, so the only way they were going to beat Intel's single threaded performance was to increase IPC enough over Intel that the frequency gap didn't matter. Seems like they finally got there (the point where their IPC is high enough that a 4 point something GHz Zen part can exceed the single threaded performance of a 5 point something GHz Intel part).

Intel is toast

Maybe so, but could you please stop posting unsubstantive comments to Hacker News?

You are right, I was not in the best state of mind.

It happens to all of us.

Reply on news.ycombinator.com