AIDA64 Memory Bandwidth & AES
Now we will look at some specific AIDA64 CPU Benchmarks that test specific parts of the CPUs. The AIDA64 AES benchmark measures CPU performance using AES (Advanced Encryption Standard) data encryption. In cryptography, AES is a symmetric-key encryption standard. AES is used in several compression tools today, like 7z, RAR, and WinZip, and also in disk encryption solutions like BitLocker.
We will also look at Memory Bandwidth. The LGA1700 and LGA1851 platforms (Raptor Lake and Arrow Lake) are using identical memory, set at XMP 1 timings of CL34 and 6400MT/s in Gear 2. The AMD AM5 platform is running at EXPO 1 timings of CL32 at 6000MT/s.
In terms of memory bandwidth, the important factor here is that the Raptor Lake-R and Arrow Lake-S CPUs are running with the same memory configuration, timings, and MT/s. Therefore, we find that memory bandwidth is rather similar, but interestingly a little bit lower on the new LGA1851 Arrow Lake platform.
We do see about a 1,100MB/s read difference (1%) between the Intel Core Ultra 9 285K and the Intel Core i9-14900K. On the write performance there’s a more significant 1,500MB/s (2%) bandwidth difference. The Intel Core Ultra 5 245K, however, is a bit faster than the 14600K on read bandwidth, but a bit slower on write performance, but not a large degree. You can’t really compare this to the AMD configuration, they are completely different architectures, memory controllers, and of course different DDR5 speeds. We do see in terms of raw bandwidth, Arrow Lake (similar to Raptor Lake) has about 28% more raw bandwidth on read and a closer 10% difference on write bandwidth.
The CPU AES benchmark is good for showing multi-core performance, as well as cache performance, as we can see the 7950X3D and 7800X3D doing quite more than usual, thanks to their larger L3 cache. Because this is a more multi-core focused test, we also don’t see much difference between “Balanced” or “Best Performance” power profile modes on the Intel Core Ultra 9 285K or Ultra 5 245K.
In terms of performance, the new Intel Core Ultra 9 285K ends up being 8% slower than the previous generation Intel Core i9-14900K. Compared to the competition, the 285K ends up being 17% slower than the AMD Ryzen 9 9950X. The Intel Core Ultra 5 245K ends up being 13% slower than the previous generation Intel Core i5-14600K, and just on par or parity with the AMD Ryzen 7 9700X.
Discussion (19 replies)
Join Discussion →Those are some dissapointing results for gaming, kinda leaves me at an annoying point, 9800X3D is only an 8 core providing it launches when it's rumoured to (early november) rest is early next year and I don't realy want to go previous gen.
Also like I said in the other thread, the ultra 9 and 5 are already sold out at some outlets around here. intel still going strong (or no stock to speak off)
There are people who will buy Intel even when they have no logical reason to do so. I work with a guy like that. He won't buy anything else because of the reputation non-Intel CPU's had in the 1990's.
Gotta know supply to gauge demand...
My take: This is a transitional product that lays the ground work for future releases. This isn't remotely the first time Intel has done this. The Wilamette Socket 423 Pentium IV's were such a product. It introduced RAMBUS memory and they were often out performed by their Pentium III counterparts, especially Tualatin CPU's. The socket Northwood Socket 478 stuff was a lot better. The Prescott core CPU's were also kind of transitional as these were built with even deeper pipeline stages to ramp the clocks up even more. Their EM64T extensions were present on the core but weren't enabled (excluding one model) until the transition to LGA sockets.
Unfortunately, the only real benefit to these is that they shouldn't have any degradation and stability issues in games where as the 13th and 14th generation CPU's did prior to microcode updates and BIOS updates were created to mitigate them. Assuming your CPU wasn't degrading to start with.
Not competitive, high price, a short lived platform and Win 11 is a mess. Nothing ever works as intended at launch. What is wrong with these companies? Fumbling the ball at the 1-yd line seems to be the norm. Bush league overall...
@Brent_Justice - what I'm seeing about is that by decoupling the memory controller from the compute die, Intel has added latency to main memory access. Do we know if this is reflected in say AIDA64?
@Dan_D - looks like we're going to have to see if Intel can overcome the above latency issue the same way AMD did when they went to chiplets with Zen.
I think they can. How they would be able to mitigate this is unclear to me but this is the type of thing Intel is typically good at engineering their way through historically. Generally speaking, Intel's CPU's have not been effected by latency to the same degree as AMD's. AMD for one reason or another has always dealt with that to some degree. AMD's answer for the most part was simply to throw L3 cache at the problem which doesn't work in all scenarios.
Though throwing L3 cache at the problem is a simple fix, but its a costly one.
AMDs use of stacking for the L3 cache on a CCD, which in turn they can use for gaming SKUs as well as enterprise (which they originally created them for), is genius.
Imagine slow(er) Zen 3 and Zen 4 cores (both clockspeed and IPC vs. Alder/Raptor P-cores) not only outrunning them but also doing so at 1/2 the power draw for gaming.
Now imagine Intel P-cores backed up by enough L3 cache to make memory latency irrelevant.
I think Intel needs to copy AMD here (or emulate? imitate?) and find a way to get that cache in a die with broader market appeal (i.e. also usable in Xeons). Their chiplet setup should enable them to do so.
One thing that's been sticking in my mind with the Arrow Lake release is that this is the first time Intel has decoupled their memory controller from their compute dies since... Core 2? At least in the consumer space (so not talking about buffered DIMMs). AMD did with Zen 2 IIRC. Intel has been using IMCs since Lynnfield in the i7 870 from 2009, 15 years ago!
I'm not sure that this is even possible. I think there are probably scenarios where simply adding cache doesn't help.
Well, AMD had an integrated memory controller since the Athlon 64 days. However, Intel's Core 2 Duo/Core 2 Quad were the last CPU's to have an off-die memory controller. Though it was in the chipset at the time and not anywhere on the CPU itself.
I believe its two things:
1. Consumers buy products that suck on the promise that they will be fixed later
2. Development practices that involve the term: MVP (Minimum Viable Product)
Number 2 combined with number 1 just continuously lowers the bar for number 2.
Thanks, @Brent_Justice for the very detailed review of these CPUs and extra testing with both power modes. It's a shame that APO isn't supported for the games reviewed, especially since they are popular titles that users are likely to have in their libraries. I wonder if Intel will be able to provide updates later on which might improve those results. Really interesting how CB2077 has a prioritize P-core option that only further proved there are other potential gains to be had.
A lot to unpack with these and obviously Intel is experimenting with a number of strategies with them.
I will have more information in an upcoming review, but installing some DDR5 CUDIMMs 8400, but running at 8200 MT/s currently, AIDA64's Cache Mem latency result is 88ns this is at 8200MT/s CL40-50-50-128 CR2. I will have some memory comparison data in an upcoming article at 6400, 8200, 8400 and 8600 on this motherboard with this RAM kit, so stay tuned. But my first reaction is that this isn't bad latency, I think it was worse at the beginning of all this, but as new BIOS's have come out, latency has improved. MSI's fix notes state XMP and memory performance optimization with the new BIOS releases I'm using.
I really want an all P-Core option, an 8-Pcore only or 12-Pcore only CPU would be a great option for gaming.
I am NOT a fan of big.LITTLE / hybrid architecture on x86 desktops. Intel please give us P-core only options!
I would prefer only p-cores too, but due to the lack of HT would we not require like at least a 10 or12 core if you need 8 for games already.
I mean testing conditions is one thing, but a real everyday use system usually has more running in the background then just the bare minimum.