![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
I do kind of think CAS Latency is going to shit, permanently, unless you are willing to pay ungodly sums of money to get something in the the 30's. Even then the performance gains will be negligible. The focus will be on the frequency gains. It seemingly already has been like that on DDR4 for a while now. I imagine DDR6 will have CAS latency of 80 or something... lol
Latency has taken a hit on every generation from SDRAM to DDR. RAMBUS was brutal on the latency issues IIRC. It's always been a thing as the RAM has advanced architecturally. I remember SD RAM being incredibly low latency
I noticed the higher speed DDR5 sticks are actually lower timings. Normally they go up with speeds.
I think it's a symptom of quality of the memory required to sustain those frequencies. We will probably hit a point where everyone and their cousin is manufacturing DDR5 and then the latencies will be all over the place for a while until the process matures. I think we saw stuff like this with high end DDR4 from Samsung and Micron (with all the cheapo brands offering similar frequencies at higher latencies).
You can disable the Intel and go exclusively Nvidia, it destroys battery life though but you can do it, it is in the windows settings.
I appreciate the breakdown.
I suspect that 13th Gen Intel will be beastly. They are doubling their L2 Caches and adding another 4MB to L3 I believe. If AMD's 7000 Series is just now trading blows with 12th Gen Parts, they're gonna be in serious trouble with the 13th Gen Intel Parts. If that wasn't enough, they allegedly have clock speeds of up to 5.8 Ghz on a single core and their efficiency is back-ish. No small feat on the same node.
TSMC can't physically do the 3D stacking chips yet for their 5nm process, The BESI 8800 Ultra is not as stable or accurate as is needed at that size and they are physically failing at lining up the two chips more than 60% of the time, TSMC and BESI are working hard on a solution but they don't have one yet. TSMC though has supposedly developed a new bonding agent for the stacked chips that has better thermal properties than the one used for the 5000 series. The copper-based bonding agent used in the 5800x3d has too high a resistance so when more than a few volts are pumped through it, it would begin to physically melt, this is why AMD had to lock that down hard. The new bonding agent has a lower resistance though still higher than they like, but they can't change the melting point of it because much higher than it currently is and the application of it would damage the chip in the process.
Once you get to DDR5 5600, the benefits can be non-trivial, in CPU limited situations. Hardware Unboxed has been using DDR5-6400 for Intel 12th gen and its pretty significant in CPU limited games:
I'm not sure where you are getting this narrative from.
Exactly. See ryzen 5000 pricing and product stack for a clear example of this. Maybe an awesome showing from raptor lake will drive down the 5800x3d pricing some more. I wouldn’t mind that at all
All these top end CPUs are overkill for their matched GPU tier. The differences are in multithreaded stuff and power usage. For most of us, they’re basically interchangeable (with Intel just needing a 240v outlet)Sadly if anything I expect what I am seeing to do the opposite and drive the costs of the 5800x3d up, if it turns out it still competes well against the new generation it gives much further incentive for existing AM4 users to just get that and hold off on a new system, the demand spike would send the pricing back up not further down.Exactly. See ryzen 5000 pricing and product stack for a clear example of this. Maybe an awesome showing from raptor lake will drive down the 5800x3d pricing some more. I wouldn’t mind that at all
With all the new attention that 79XX reviews are ending up bringing to 5800X3D since it's so strong on benchmarks, I see the $399 price staying put for quite a while. The $409 13700K is going to edge out 5800X3D for gaming, but for anyone with an existing AM4 investment that primarily games, the latter is a no brainer drop-in replacement.Exactly. See ryzen 5000 pricing and product stack for a clear example of this. Maybe an awesome showing from raptor lake will drive down the 5800x3d pricing some more. I wouldn’t mind that at all
If you look at their 12100 video for memory testing: DDR4 4000 CL19 single rank, was barely better than single and dual rank 3200 CL14, in maybe 4 games. And if you look at the 12 game average, both single and dual rank DDR4 3200 CL14 beat DDR4 4000 CL19.
You joke but AMD it does not multi-task that well to that degree when all cores are in use, to date AMD's memory and task scheduling has left a lot to desire, and I am hoping that the new IO die on the 7000 series improves it but I am not super hopeful on that front.Does ryzen have something similar? Oh, I forgot, ryzen is all P core
Thread director initially made the decision based on workload size and system stress I don't think it took active focus into consideration, so it would put things like your email and excel windows onto the e cores and put Blender on the P cores, but then if you were running Unreal in another task they would try to share the P cores while your web browser and other stuff took up what was left of the E cores.
I'll be curious how intel handles it when they start gluing CPUs together. the CCX to CCX hit is real for sure.
The Sapphire Rapids leaks to date have been promising, Intel has a lot of experience with multi-chip in the form of their multi-socket systems I expect much of that can be scaled down and remain relevant. Intel does use an MCM package for their Ponte Vecchio GPU and from what I have read about that it performs very well at its intended tasks. The hardest part for MCM designs is the interconnect and the scheduling, Intel seems to have a firm grasp on both so I am not worried about it being bad, how it stacks up to TSMC and AMD's process will be interesting though as it's essentially going to be Intel's 1st generation packaging vs TSMC's 4th.
I see what you are attempting to prove. It's highly game specific and the DDR4 4000 kit matches or edges out the DDR5 4800 in many benchmarks that outstrip the DDR4 CL14 3200 memory. The CL14 3200 CAN beat the DDR5 4800 and the DDR4 4000 but it's not in every scenario. Clearly, some games are more sensitive to the CAS Latency and others are more interested in frequency. In many of the benchmarks the DDR4 4K outstrips the DDR5 4800, or matches it exactly. This would effectively prove my point. The CL14 3200 Mhz stuff can beat them both sometimes (in a good amount of titles).
DDR5 has some incredibly cool features, the channel splitting, EEC, and blah blah blah, it is a very big shift in how consumer memory works. But until the OSs and Game engines are redesigned in a way that takes advantage of those features then the latency will hurt it for a number of use cases and is then going to rely on pure speed to catch it back up. Some years from now when DDR5 is the majority and the next generation of OSs and game engines is out using those feature sets then DDR5 will be the clear winner. Until then you take what you can get, but if you can spring for DDR5 then I expect it will serve you better in the long run.
It's like a rinse and repeat scenario ever new generation, until things get normalized and widely accepted. Then the new tech really starts pulling ahead.
Some of it is DDR5, some of it is the OSs doing better at memory management, drivers, blah blah blah, DDR5 is the future and it is better and mass adoption will make it cheaper. I am just hoping that having more than 2 memory channels stops being a bad thing on the consumer side, I would love at least 4 channels for DDR5 operating at full speed on the consumer side. Doubt I'll see it this generation but hopefully next.
It's not the act of minimizing it that freed up the P cores it was the launching of another high-resource application, ensuring performance for the active task. It's actually pretty useful for a working environment and office setting, the application load they displayed there was a little over the top in the Demo but it showed some good points. I mean very few people are going to be doing large compiles in Blender while simultaneously building scenes in UT5, and gaming. But the fact the game was running smoothly while it was still doing those other tasks is kind of a big deal, My Threadrippers, Xeons, Epics, Ryzen's, and other Intel CPUs couldn't pull that off.
I get that, but that's not what they showed. The P cores weren't relegated to a foreground task, they just went dead before another application was launched.
Not even with task priority?
I'm fairly certain you had to manually set this before and it was a huge PITA. The Scheduler in the OS is handling everything seamlessly and all without user input. If I understood the video correctly the E cores completely offloaded the tasks from the P cores without a hit to performance in any way. I could be wrong.
Prior to this P/E-core nonsense for PC CPUs - yes - this was entirely possible just by virtue of the OS load balancing. On Linux on my 3975WX I can compile something like LLVM with 48 threads ("vcpu"), have other uses logged into the machine doing smaller compiles, and still play a game in the foreground - all concurrently. If I look at something like htop I see most cores pegged to 100% by their respective processes. This doesn't require assigning affinity manually or anything more complicated like that.
I can't argue with the ability to multitask all being there before. The behavior of the Windows 11 Scheduler is more advanced than prior iterations of Windows. I'm not personally a bit fan of P/E cores but it sure looks pretty cool under Windows 11. I suspect that something similar happens with AMD processors in Windows 11, when you have a shitload of cores. IIRC in the past all the cores would just carry a part of the load. What's happening here seems specific to the Intel Architecture.