raptor lake Intel Core i9-13900K 'Raptor Lake' CPU breaks 40,000 points in Cinebench R23 with unlimited power and 5.8GHz clock

[Ydfhlx]

On 8/8/2022 at 12:33 AM, PDifolco said:

PPT 350W ?? Who will use that  ?? And I don't even imagine the cooling you need ...

360mm radiator on an open air bench and it hit 90°, presumably in room with A/C

 

This is impossible to cool under normal circumstances.

 

And it reminds me of FX CPUs, some of them were sold together with AIOs.

[SeriousDad69]

3 hours ago, LAwLz said:

Why are you going Zen 4 if Raptor Lake doesn't enable AVX-512?

 

Zen4 doesn't have AVX-512 either. 

Literally every leak and rumor has stated Zen 4 as having AVX-512 support. Probably not complete support, as AVX-512 actually encompasses a lot of instructions, but support for most of it. Everyone already knows Raptor Lake is going to have support for it, just waiting to see if Intel is going to disable it on Raptor Lake-S so there's no potential compatibility problems for their baby cores.

[LAwLz]

1 hour ago, SeriousDad69 said:

Literally every leak and rumor has stated Zen 4 as having AVX-512 support. Probably not complete support, as AVX-512 actually encompasses a lot of instructions, but support for most of it. Everyone already knows Raptor Lake is going to have support for it, just waiting to see if Intel is going to disable it on Raptor Lake-S so there's no potential compatibility problems for their baby cores.

As far as I know, AMD has not confirmed AVX-512 support, at least not AVX-512 Foundation.

When TechPowerUp interviewed Robert Hallock, he specifically said "AVX 512 VNNI for neural networking and AVX 512 bfloat16 for inferencing" and "we're not using a fixed-function acceleration".

 

I would not be surprised if Zen 4 ends up not supporting AVX-512 outside of the specific VNNI and bfloat16 instructions.

Hopefully it does, but I wouldn't bet on it unless AMD confirms it. I wouldn't be so sure on Raptor Lake supporting AVX-512 either.

 

 

I think a lot of the rumors are just spread by people that just heard "AVX-512" mentioned and ignored/didn't understand the "VNNI" and "bfloat16" that were mentioned after.

[leadeater]

2 hours ago, LAwLz said:

As far as I know, AMD has not confirmed AVX-512 support, at least not AVX-512 Foundation.

When TechPowerUp interviewed Robert Hallock, he specifically said "AVX 512 VNNI for neural networking and AVX 512 bfloat16 for inferencing" and "we're not using a fixed-function acceleration".

 

I would not be surprised if Zen 4 ends up not supporting AVX-512 outside of the specific VNNI and bfloat16 instructions.

Hopefully it does, but I wouldn't bet on it unless AMD confirms it. I wouldn't be so sure on Raptor Lake supporting AVX-512 either.

 

 

I think a lot of the rumors are just spread by people that just heard "AVX-512" mentioned and ignored/didn't understand the "VNNI" and "bfloat16" that were mentioned after.

And even if Zen 4 supports AVX512F that may well be microcode locked out on consumer and only on EPYC and maybe TR.

[LAwLz]

9 minutes ago, leadeater said:

And even if Zen 4 supports AVX512F that may well be microcode locked out on consumer and only on EPYC and maybe TR.

Or they might only implement AVX-512F in zen 4c, their other server architecture. 

 

Or they might do what they did with AVX2 and implement AVX-512 by splitting it into two 256 bit wide instructions. Not sure if that's possible this time around and what the performance is like though. 

 

Anyway, best to not assume anything until AMD confirms it in non-vague ways. 

[leadeater]

12 minutes ago, LAwLz said:

Or they might only implement AVX-512F in zen 4c, their other server architecture. 

I doubt that, 4c is their Cloud optimized low power area reduced sub-architecture so AVX512 simply doesn't align with that at all. It'll be in Zen 4 or not at all.

[porina]

15 minutes ago, LAwLz said:

Or they might do what they did with AVX2 and implement AVX-512 by splitting it into two 256 bit wide instructions. Not sure if that's possible this time around and what the performance is like though. 

That would be like Intel's "one unit" 512F implementations. It still provides a performance uplift over AVX2. "two unit" would of course be better but I wouldn't expect that from AMD.

[SeriousDad69]

9 hours ago, LAwLz said:

As far as I know, AMD has not confirmed AVX-512 support, at least not AVX-512 Foundation.

When TechPowerUp interviewed Robert Hallock, he specifically said "AVX 512 VNNI for neural networking and AVX 512 bfloat16 for inferencing" and "we're not using a fixed-function acceleration".

 

I would not be surprised if Zen 4 ends up not supporting AVX-512 outside of the specific VNNI and bfloat16 instructions.

Hopefully it does, but I wouldn't bet on it unless AMD confirms it. I wouldn't be so sure on Raptor Lake supporting AVX-512 either.

 

 

I think a lot of the rumors are just spread by people that just heard "AVX-512" mentioned and ignored/didn't understand the "VNNI" and "bfloat16" that were mentioned after.

Well, if Zen 4 desktop AND Raptor Lake-S both lack AVX-512 support, it at least makes my decision easier. I'll see if I can get a used 11700K and board for cheap, and if I cant I'll just get whatever has the highest single thread at the lowest cost(Probably a 13600K). Really wish I could justify stretching my Skylake system out for another year and just getting a Meteor Lake or Zen 5 CPU, but it is what it is.

[igormp]

6 hours ago, LAwLz said:

Or they might only implement AVX-512F in zen 4c, their other server architecture. 

 

Or they might do what they did with AVX2 and implement AVX-512 by splitting it into two 256 bit wide instructions. Not sure if that's possible this time around and what the performance is like though. 

 

Anyway, best to not assume anything until AMD confirms it in non-vague ways. 

I was going to comment on that. It is possible, but performance wouldn't be that great over regular AVX2.

[KaitouX]

Another Raptor Lake lineup leak, it's basically the same as the previous one, but this time with some interesting information on the lower tier SKUs.

The lineup itself is basically the same as this leak from one year ago:

 

Everything below the 13600K(F) is supposedly just an Alder Lake refresh, while the 13600K(F) and above are Raptor Lake.

I'm curious to know if this will increase in any significant way the performance difference between the 13600/13500 and the 13600K.

[LAwLz]

9 hours ago, KaitouX said:

Everything below the 13600K(F) is supposedly just an Alder Lake refresh, while the 13600K(F) and above are Raptor Lake.

I'm curious to know if this will increase in any significant way the performance difference between the 13600/13500 and the 13600K.

That's a shame. I was hoping for a 4+4 core i3. That would have been amazing. 

 

I hope the 13700 (non-K) gets released in the near future. I feel like overclocking is basically dead, and the extra base frequency doesn't seem to matter because of all-core boost anyway, so the K variants seems like a pretty bad deal most of the time. 

[StDragon]

On 8/9/2022 at 2:00 PM, leadeater said:

I doubt that, 4c is their Cloud optimized low power area reduced sub-architecture so AVX512 simply doesn't align with that at all. It'll be in Zen 4 or not at all.

AVX512 is broken down into subsets. An important feature is VAES which is supported on newer Intel CPUs (except Alder Lake) and Zen 4. At the moment there's a bug with Windows 11 implementation using this instruction set for BitLocker, but performance should be an improvement over Windows 10.

 

Utilizing the existing AES-NI for storage encryption is already limited and doesn't scale with PCIe 4.0 (let alone 5.0) NVMe drives. And while technically Opal or eDrive will leverage hardware based encryption right on the SSD controller, it poses other security challenges in addition to inhibiting BMR restoration. Many would prefer BitLocker if the performance can keep up when needed.

 

[leadeater]

19 hours ago, StDragon said:

AVX512 is broken down into subsets.

I know, we were talking about AVX512F which is actually AVX512 when talking about instruction sets as a functional thing in comparison to AVX2/AVX/SSE etc. The encryption engines and AI accelerator specifications as part of AVX512 can and have been implemented without the entirety of AVX512F which is only the most basic set of instructions required to say you have "AVX512" and not "AVX512-VNNI" etc.

 

Zen 4c will have the encryption engines part as SSL/TLS offload is a big requirement for the use cases it'll get used for. Zen 4 in EPYC which may or may not actually have AVX512F is where you'd see that rather than Zen 4c. Maybe both could have it but with Zen 4c being a die space and power optimized configuration to increase core counts AVX512F getting the chop would be right up there on the list to remove, along with extra large caches which you'd need to effectively use AVX512F.

[Shzzit]

On 8/7/2022 at 9:32 PM, BlueChinchillaEatingDorito said:

Should've done my research when I swapped out my 1100T for the FX-6300 from my friend who upgraded to Ryzen. Didn't know it was this bad of a regression... 

I'm glad back then i made the choice of getting the i7-980x, it lasted me a LONG TIME.  In Fact it's still in a build my cousin is using with a gtx 1060 lol. 

[Shzzit]

On 8/9/2022 at 6:46 AM, Ydfhlx said:

360mm radiator on an open air bench and it hit 90°, presumably in room with A/C

 

This is impossible to cool under normal circumstances.

 

And it reminds me of FX CPUs, some of them were sold together with AIOs.

Its overclocked to 5.5ghz with alot of voltage.  Stock power im guessing will be alot less.  Even then i would still OC it and have it suck as much watts as i can get hahaah. 

[BlueChinchillaEatingDorito]

1 minute ago, Shzzit said:

I'm glad back then i made the choice of getting the i7-980x, it lasted me a LONG TIME.  In Fact it's still in a build my cousin is using with a gtx 1060 lol. 

To be fair... AMD had nothing in the price range of the 980x. The 1100T was basically the best chip they can come up with and was half as much of the 980x. So if you had that much money to spend... AMD was never going to be on your radar. 

[ouroesa]

On 8/8/2022 at 7:47 AM, BiG StroOnZ said:

It's only 4% slower than a Threadripper 3975WX

Worth noting that the Threadripper 3975WX is a 2 year old CPU running at ~half the clockspeed.

[BiG StroOnZ]

1 hour ago, ouroesa said:

Worth noting that the Threadripper 3975WX is a 2 year old CPU running at ~half the clockspeed.

 

Right, but the 3975WX also has 100% more threads than the 13900K:  32t vs 64t. It also has 33.33% more cores than the 13900k: 24c vs 32c. Which is ever important in this particular benchmark. This is why I also posted the 5975WX results, as the 5975WX is much newer and based on Zen 3 (better IPC and higher clocks) rather than Zen 2. Keeping this all in mind, the 13900k was only 16% slower than the 5975WX. IMO, this is an exceptional showing for the Raptor Lake processor. Lastly, it should be noted it's not quite half the clock speed. The 3975WX is capable of a 3.75GHz 32c Turbo Frequency. Half would be 2.75GHz. Which even its base clock is 3.5GHz. It's also important to keep in mind that the P-Cores are the only cores running at 5.5GHz. While the E-Cores are running at 4.3GHz. That is only 15% faster than the 3975WX all core.

[LargeMan]

I WISH I COULD AFFORD ONE. 

[sounds]

While it's not a leak, probably Intel-approved, this is interesting: chipsandcheese.com posted cache benchmarks of a 13900K Raptor Lake engineering sample.

 

P-Core Cache Latency (lower is better):

 

E-Core Cache Latency (lower is better):

 

The graphs show only a small reduction in latency, but that translates to better performance per watt.

[bcredeur97]

On 8/7/2022 at 10:25 PM, AndreiArgeanu said:

300W CPU paired with a 600-800W GPU. Guaranteed 1kW space heater. I somewhat expected this though but I thought it wouldn't happen since Intel stepped up their game with e cores and what not. They've quite literally reached bulldozer levels of power consumption and that's when power limited. And when not power limited they beat even the extreme edition cpu's of the day that chugged power like there was no tomorrow. I'm curious what kind of cooling solutions will be required for this CPU.

-snip-

at this point you HAVE TO put your PC in a dedicated and remote room/closet to not heat up your space. Lol.

That really is the best way to do it though. Just have a desk with monitor and accessories. No PC to worry about, that's "hosted" in the other room lol

 

Everyone's laundry room will double as a server room now!