Arrow Lake 2024 (and beyond)

hc=8IW2fmswFX8Q7kNvgG-Pfr9&_nc_ht=scontent.fymq1-1.jpg


https://wccftech.com/intel-core-ult...ercent-faster-core-i9-14900k-v-ray-benchmark/

Considering that a benchmark that respond well to Hyperthread (+19/20%), the gap between the 285 and the 14900k would be really good.
 
Blackstone said:
View attachment 681558

But do I need to upgrade from this CPU lol:
View attachment 681559
Click to expand...
That's impressive, given it doesn't use Hyperthreading. It does more with less... I am getting AMD FX series / Bulldozer (that actually work) kind of vibes off of the new chips. Their GPUs supposedly outrun AMD Strix point APUs at lower wattage. But they get crushed as they don't scale with power draw and the AMD APUs do. Still, very impressive that Intel can match AMD's graphics and outstrip them at all in 2 generations (technically 3).

Should be real interesting seeing as AMD isn't going after the top end for graphics this generation... Wonder if Intel might be able to leap frog their product stack on discreet GPUs. I think AMD made some tactical errors.... Intel may be poised to rebound.

EDIT - I misspoke, Battlemage doesn't hit until Lunar Lake (looking at the wrong slides). Looks like these have Alchemist GPUs
 
I just hope Arrow Lake will kick ass at gaming. Although beating the 7800X3D by any significant margin may be tough since the new Windows update boosted it's performance by almost 10% on average. Still, fingers crossed for a big leap in gaming performance for ARL.
 
Looks like Intel still needs a lot of power for full multicore.

On the flipside, high single core performance and no HT, could mean pretty good gaming efficiency. I think they will probably end up around Zen 4/5 power usage in gaming, with pretty close to X3D performance (if you have really fast DDR5). But, the X3D will still have a marked efficiency advantage in gaming.


Funny thing is, I swear like a year ago, Intel was working on a sort of "big L4 cache" tile.
 
LukeTbk said:
Do not get the potentially 2 time the bandwidth of zen5 talk, I could be missing the point of the 7 minute video format with random images, I feel this is a old video re-released, with percentage just out of a bit nowhere:

3 weeks ago:

View: https://www.youtube.com/watch?v=rmiujWgfOfQ&t=297s
Click to expand...

The claim is that the Infinity fabric that connects the CCDs to the IO die can only manage 64GB/s, so no matter what ram you hook up to an AM5 chip that's all you'll get. Meanwhile Intel chips don't have this limitation, and can take full advantage of faster ram. Or maybe not, as some of the articles about CUDIMMs (clocked unbuffered DIMMs, can run at 9000+) say they'll have to run in gear 2, so memory controller at half speed. That could result in a situation similar to what we see with AMD where there's a sweet spot for ram speed. Still, the i9-14900K is already ahead on DRAM bandwidth, and one would expect Arrow Lake to improve on that.
 
I think this a bit going over my head as we often see gain to zen5 going over ddr5-4000 in dual channel (that when you hit the 64GBps limit).
 
LukeTbk said:
I think this a bit going over my head as we often see gain to zen5 going over ddr5-4000 in dual channel (that when you hit the 64GBps limit).
Click to expand...
On AM5 the sweet spot for performance is ~6000 and CL as low as you can get. A little more MHz if you have a good chip. Go too fast and you have to lower the clock speed on the memory controller. It's not just about how many GB/s it can move, it's also how long between the CPU wanting whatever is at X address and the ram & memory controller delivering it. Assuming that guy is right about the 64GB/s limit between a CCD and the I/O die (I have no personal knowledge of this, though I am now curious to read up on it), that means that on AMD latency is basically everything... unless you have a multithreaded workload on a CPU with more than one CCD. It's 64GB/s per CCD, not 64GB/s for the entire CPU package. Hopefully this makes sense.

Now back to Arrow Lake... I'm really pretty interested in this CAMM2 thing. Unless they come out with denser DIMMs it's looking like my best way to get 64+GB without using dual rank DIMMs or more than one DIMM per channel and taking the associated speed hit. CUDIMMs could be nice too, but they'll need to either be bigger than 24GB single rank or support higher speeds with dual ranks and/or 2 CUDIMMs per channel. Funny thing is the launch press release on the Xeon 6900 p-core models got be drooling over server ram yesterday. The new Xeons support memory speeds up to 6400 using RDIMMs and 8800 using MRDIMMs. Multiplexed rank DIMMs or something line that. 8800 in server sizes? If they're doing that with Xeons I'm wondering what we're going to see on desktops.
 
zandor said:
The claim is that the Infinity fabric that connects the CCDs to the IO die can only manage 64GB/s, so no matter what ram you hook up to an AM5 chip that's all you'll get. Meanwhile Intel chips don't have this limitation, and can take full advantage of faster ram. Or maybe not, as some of the articles about CUDIMMs (clocked unbuffered DIMMs, can run at 9000+) say they'll have to run in gear 2, so memory controller at half speed. That could result in a situation similar to what we see with AMD where there's a sweet spot for ram speed. Still, the i9-14900K is already ahead on DRAM bandwidth, and one would expect Arrow Lake to improve on that.
Click to expand...
12th, 13th, and 14th gen intel also run their DDR5 in Gear 2.
 
If I'm not mistaken the Fabric Clock in AMD for true stability is 2000 in 1 to 1. You can push past that but you risk having issues. General consensus is it's kinda crazy to waste your money past DDR5-6000 when the Fclk is only running at 2000. Using DDR5-8000 is double the cost and not really worth it.

This is from the 13900K:
1727357664062.png

Going off of Intel spec, it would seem that their architecture was designed for Gear 1 Speeds of 1600 in DDR4 and 2400 in DDR5.

In practice, I've never been able to get a stable Gear 1 on 32 Gigs DDR4 over 4366 (2183)... It was stable for about 6 months and then the processor started to fail (13900K)... I run at 4000 in gear 1 now with 64 Gigs (12900KF). I see a lot of high performance kits in DDR5 dial the RAM down to 4000 (2000 Gear 1), so theoretically you can push over 2000 but it becomes less and less stable. I suspect I'm not the only person to draw these conclusions.

I don't think Intel's fabric is superior to AMDs. AMD is just more dependent upon it as their CPU dies communicated at fabric Speeds almost exclusively, though I think 9000 series has faster communication somehow and I'm no engineer but that doesn't seem to have translated into better performance than the 7000 series unless it's specific server or datacenter workloads.

DDR5 on Intel is generally run Gear 2 unless you are manually controlling your ram and "fabric" / bus. Gear 2 became the norm 12th Gen up. Before that everything was Gear 1.
 
zandor said:
On AM5 the sweet spot for performance is ~6000 and CL as low as you can get.
Click to expand...
Which is significantly more bandwith (96gbps) than 64gbs in dual channel, thus my point. Many ryzen cpu are single CCD and still take advantage of faster than 4000 MT/s ram (the frequency needed to reach 64GBps).
 
chameleoneel said:
12th, 13th, and 14th gen intel also run their DDR5 in Gear 2.
Click to expand...
Sorry, the leak/rumor is about Arrow Lake. If it can run 9000+ in gear 2, we could end up with a situation where you can run 4800 or 5000 or something like that in gear 1, thus potentially creating a sweet spot, at least as far as cost goes.
 
zandor said:
Sorry, the leak/rumor is about Arrow Lake. If it can run 9000+ in gear 2, we could end up with a situation where you can run 4800 or 5000 or something like that in gear 1, thus potentially creating a sweet spot, at least as far as cost goes.
Click to expand...
Yes and my point is that Intel chips already run in Gear 2, and still benefit a lot from faster RAM. Faster the better. (Alrhough 6000 with optimized timings competes well with 7200).

I'm sure Arrow Lake will behave the same and maybe even scale better.

With AMD's design, they do not benefit from faster RAM in Gear 2. In fact, they require DDR5 8000 in gear 2, to more/less equal 6000 in gear 1.
 
MistaSparkul said:
Less than a month to go now. Hoping for the best from Arrow Lake.
Click to expand...
Me too, but the latest benchmark I saw indicates weak performance. In the benchmark it performs worse than a 14900 (non K), worse than the AMD 9700X by a wide margin (especially in multithreading)... It could be a step back before leaping forward. I suspect any gains in this generation should (theoretically) be from efficiency gains. Can't find the article anymore so hopefully it was misrepresented.
 
The leaked results do seem to vary depending on the tests being shown. I'm not really too concerned about synthetic performance though, I just want ARL to destroy in gaming.
 
I’m just going with the biggest and baddest Noctura tower. I hve a dual tower with 2 120 mm fans, the DH15 and it has been an absolute beast on Coffee Lake 8700k.
 
You must log in or register to reply here.
Back
Top