Intel Core Ultra Arrow Lake Preview

[btarunr]

Hi W1zz or anyone in the know . I read most of the slides but it's just marketing gibberish. Does it finally have AVX-512 support without buts or ifs?

Nope, AVX512 isn't exposed. There's AVX2 and VNNI.

 

[MxPhenom 216]

Well I won't make it personal. +2 percent is not alright for AMD. -5 percent for Intel is even more not alright.

We are in the midst of the biannual marketing cycle but they are pushing useless products because the CPU and GPU market is an "upgrader's market". What that means is we all already have access to commodity pricing in compute perfomance. The only reason to spend more money is for upgraded performance and so, there is no real market for new product at all right now.

Just buy the 8600G for $145 USD for a cheap PC and the 7950X for rendering and the 9800X3D for highest refresh gaming when it comes out. Otherwise probably best to sit tight with older product right now.

I just checked, the new Arrow Lake i5 product and motherboard is $806.40 Canadian bucks after tax. Nope. Just buy an older <$100 motherboard and a faster older CPU instead.



there's no advantage to raising the memory speed at the cost of latency, it is just marketing fluff

Not really, and even then you act like i dont know that. If memory makers start pushing frequencies then latency will be next to come down too. It always does.

Whats your response when AMD says their sweet spot is some x amount higher than their default? Just "marketting fluff"??... because you'd be wrong. Their platforms always operate better at that sweet spot and it can be better if you can push it above while maintaining 1:1 uclk to mclk ratio. And there is a ton you can do with ddr5 to mitigate latency on these dimms beyond just CAS.

I like tweaking ram. I work on layout of memory PHYs in VLSI as my job. Its fun for me.

 

[regs]

Why not combine the two designs into one? I expect this to happen in the future.

Better stick AVX-512 and AVX10 to P-cores only. There just need to make all AVX-512+ call be forwarded to P-cores. I don't know why they still can't make it.

 

[Dr. Dro]

Nope, AVX512 isn't exposed. There's AVX2 and VNNI.

Yeah, from what I could gather they retained the same instruction sets as Raptor Lake despite the new microarchitectures. That's really a tough sell, considering AMD's gone all out with their instruction set support on Ryzen this generation, including competent AVX-512 of the likes Intel can't offer even on Xeon 6. Shame

 

[Nhonho]

I think it's unlikely that Intel will release multiple generations of CPUs for a single socket, as AMD does with the AM4 and AM5 sockets.

Chipsets are much easier and cheaper to design and manufacture, and they generate a huge profit for the company.

Better stick AVX-512 and AVX10 to P-cores only. There just need to make all AVX-512+ call be forwarded to P-cores. I don't know why they still can't make it.

I think Intel should abandon this P and E core approach in consumer CPUs and launch them with only P cores. Consumer applications do not use this exaggerated amount of cores.

Intel could make some of the P cores work at low clock speeds so that they have low power consumption, and the CPU itself would increase the clock speed of all cores only in situations when there is a real need for it.

 

[Minus Infinity]

I think it's unlikely that Intel will release multiple generations of CPUs for a single socket, as AMD does with the AM4 and AM5 sockets.

Chipsets are much easier and cheaper to design and manufacture, and they generate a huge profit for the company.



I think Intel should abandon this P and E core approach in consumer CPUs and launch them with only P cores. Consumer applications do not use this exaggerated amount of cores.

Intel could make some of the P cores work at low clock speeds so that they have low power consumption, and the CPU itself would increase the clock speed of all cores only in situations when there is a real need for it.

There will be no such thing a P and E cores when what left of Jim Keller's Royal cores come along as rentable units. The core will be able to seemlessly switch to whatever mode is required, high performance, ultralow power etc. No more problems with schedulers.

 

[Nhonho]

Jim Keller's Royal cores

In this approach, would there be a "P core block" and another "E core block", with the "P core block" operating only when there is a high demand from the software? I don't think this is a good idea. This way, a large and expensive amount of silicon would always be idle (not used).

I think the best solution is CPUs have only P cores, with 1 "extra big core" for every 4 P cores.

And see it:

Intel reportedly dismantles Jim Keller's revolutionary Royal Core project and cancels Beast Lake

Intel has reportedly canceled the upcoming Beast Lake architecture according to a new leak. According to past reports, Beast Lake had a considerable focus on single-core performance. In addition to canceling Beast Lake, the report from Moore’s Law Is Dead also claims that we may not see the...

www.notebookcheck.net

 

[sn2x]

Pretty impressive the E-Cores have slightly higher IPC than Raptor Lake P-Cores.

I can't wrap my head around how the P-Cores only have 9% IPC higher than last gen, when the E-Cores are 2% ahead. How can the big cores only be ~7% higher IPC than E-Cores? Those E-Cores must be really strong for this to happen, or am I reading it wrong?

I would say the P-Cores are weak, but those also appear to be on par or maybe better than Zen5 Cores...

 

[N/A]

Pretty impressive the E-Cores have slightly higher IPC than Raptor Lake P-Cores.

I can't wrap my head around how the P-Cores only have 9% IPC higher than last gen, when the E-Cores are 2% ahead.

The P core integer units have increased from 5 to 6 ALUs. The E core received one additional integer on each of the 4 pipelines for a total of 8 ALUs among other things that were beefed up.

 

[danc]

The "Arrow Lake-S" SoC puts out a dual-channel DDR5 memory interface, which we described in the previous pages. It also features a massive 48 platform PCIe lanes (CPU + chipset). Intel has increased the number of PCIe Gen 5 lanes put out by the CPU to 20—that's 16 lanes meant for the PEG (x16 slot for graphics cards), and one CPU-attached M.2 NVMe slot that runs at Gen 5 speeds without eating into the 16 PEG lanes. The CPU puts out a second set of Gen 4 x4, which can be wired out as an M.2 slot, or be used to drive high-bandwidth onboard devices, such as a discrete Thunderbolt 5 controller. The processor itself fully integrates a Thunderbolt 4 controller, which puts out a couple of 40 Gbps ports.





The processor connects to the Z890 chipset over a DMI 4.0 x8 chipset bus (bandwidth comparable to PCI-Express 4.0 x8). It puts out 24 PCI-Express Gen 4 downstream lanes. This is a massive increase from the Z790, which put out 16 Gen 4 and 8 Gen 3 lanes. The integrated USB complex consists of 32 USB 3.2 5 Gbps serial-deserializers, which can be configured by motherboard designers into five 20 Gbps ports, ten 10 Gbps ports, and ten 5 Gbps ports. There's also a 14-port USB 2.0 hub. Intel has retired the HDA "Azalia" audio interface with Z890, which means onboard audio CODECs will have to use the newer MIPI SoundWire and USB 3.2 interfaces (which CODECs like the Realtek ALC4080 and ALC4082 already do).

I am lost about z890 I/O

In total, Intel has...

16X gen5 PEG
4X gen5 nvme
4X gen4 nvme or motherboard vendor choice
4X gen4 for integrated TB4
8X gen4 downstream to chipset.

If so, it is better than AM5 no?

 

[oxrufiioxo]

I am lost about z890 I/O

In total, Intel has...

16X gen5 PEG
4X gen5 nvme
4X gen4 nvme or motherboard vendor choice
4X gen4 for integrated TB4
8X gen4 downstream to chipset.

If so, it is better than AM5 no?

depends I guess.

 

[Jayn2]

Wow, rip gamers..... 2024 has been rough for gamers.

Hopefully the 9800X3D can save the day.

Kinda disappointing considering intel has a node advantage with this.


But it does have AI to save the day said nobody in their right mind ever lol.

No xmx, but it does have pcie5 for an external GPU card. Should we expect a pcie5 Battlemage? That would dwarf the TOPS of any NPU to date.

 

[Jayn2]

Interesting how tables have turned. When Ryzen launched, it had better MT performance and worse gaming performance.

Also, no word on platform longetivy. Guess Arrow Lake's customers will be throwing their MB away after 1 or best case 2 generations.

Next.

Looks to me like CUDIMM support can be an upgrade path as they become cheaper.

 

[Bwaze]

The problem is, these CPUs are aimed mainly to people who haven't upgraded in a generation or two. Releasing the generation that barely, or not even, surpasses last generation, and trails competition, is a hard sell right now - because owners of two, three generations old PC are just shocked at prices of motherboards, comparable to the ones they bought the last time...

 

[danc]

depends I guess.

View attachment 367047

View attachment 367048

X670E and X870E i/o are pretty much "the same"

Here is AM5

16X gen5 PEG
4X gen5 nvme
4X gen5 nvme (X670E) OR integrated TB4 (X870E)
4X gen4 downstream to chipset.

 

[Jayn2]

Intel's plan is avx10.2, where they expand the kinds of avx operations. They'll extend that support to the e-core xeon products. That will bring along semi-custom customers who rely on those simd operations.

The question then will be if AMD can afford to duplicate those simd operations without also reverting to avx2.

 

[iceboy]

This has got to be the most nonsensical slide in the universe.

View attachment 366974

Look, 250W has 1 pixel less performance than 125W and 175W.

 

[DavidC1]

Pretty impressive the E-Cores have slightly higher IPC than Raptor Lake P-Cores.

I can't wrap my head around how the P-Cores only have 9% IPC higher than last gen, when the E-Cores are 2% ahead. How can the big cores only be ~7% higher IPC than E-Cores? Those E-Cores must be really strong for this to happen, or am I reading it wrong?

I would say the P-Cores are weak, but those also appear to be on par or maybe better than Zen5 Cores...

Because the P core design and the team sucks, while the E core team has been executing near flawlessly since it was known as the Atom chip. They struggled in the beginning, did good with first OoOE chip Silvermont, trailed a bit with 14nm shrink, but after 2015 it was smooth sailing.

I don't think the P cores are even 7% ahead. It's possible it's under 5%. Basically the clock frequency is the biggest difference now.

Think about it. between 2015 and 2024, the P core got 53% gains. The E core got 30% in 2015, 30% in 2017, 30% in 2020, 30% in 2022, and 30/70% in 2024. The current P core team is Haifa Israel based, the same one that created Pentium M, Core Duo and Merom/Conroe Core 2. That team/design is going to have to be completely overhauled or even dismantled and replaced by the much better E core one.

I think Intel should abandon this P and E core approach in consumer CPUs and launch them with only P cores. Consumer applications do not use this exaggerated amount of cores.

Arrowlake eliminates Hyperthreading, so it only has 24 threads, versus 32 for AMD. Even the previous gen E cores were stronger than logical threads from HT anyway. With the gap between E and P being only about 30% in both ST and MT versus 60% and 2x last generation, there will be a lot less workloads where it'll slow down.

 

[Ayhamb99]

We are in the midst of the biannual marketing cycle but they are pushing useless products because the CPU and GPU market is an "upgrader's market". What that means is we all already have access to commodity pricing in compute perfomance. The only reason to spend more money is for upgraded performance and so, there is no real market for new product at all right now.

Agreed, which is why I am holding off until the 9000X3D chips out come to then have my final opinion on this gen, if it falls short as well then it just means this generation is mediocre in general.

Arrow lake is Intel's first attempt at a tile based desktop CPU similar to what AMD did first with Ryzen 3000s, so to be honest I was expecting it to be rocky or even have performance regression. Ryzen 3000s suffered from high latency due to the new chiplet design at the time if my memory is correct, which was rectified with Ryzen 5000s which had massive performance gains over 3000s. If anything I think Arrow Lake is Intel's Zen 2 moment in the sense of them ditching their traditional monolithic design that they have used for their consumer desktop CPUs for the last decade to a tile based design, I guess they need another generation in order to properly optimise their new tile design just like how AMD did with their 5000s chips.

Intel's only possible saving grace for this gen from what I would predict is Bartlett Lake if they actually go through with it. The question is if they would backport their lion cove cores back to LGA 1700 or would they just optimise and reuse Raptor Cove. In any case, I just hope they recognise that they need to start releasing their own SKU or SKUs with stacked cache just like AMD's X3D chips because at this point it has to be apparent to them that most games nowadays benefit more from more cache compared to higher clock speeds.

 

[N/A]

Look, 250W has 1 pixel less performance than 125W and 175W.

What all that means is that the CPU doesn't need more than 125 W to play any game you throw at it. If you set to 250W in the BIOS it does nothing, because it still uses less than 125W in all the games tested.
But then again the PC is supposedly running a 4090 and only 360W on the wall. no way.

 

[BoggledBeagle]

... Releasing the generation that barely, or not even, surpasses last generation, and trails competition, is a hard sell right now - because owners of two, three generations old PC are just shocked at prices of motherboards, comparable to the ones they bought the last time...

In what exactly are Arrow lake CPUs trailing Zen 5 CPUs except in gaming compared to special 3D variants of the Zen 5 CPUs?

245K will be a better choice for mid range PCs than anything AMD makes. Mid range PCs do not contain 4090 or 5090 GPUs and in gaming people are GPU limited in 99 cases of 100.

265K and 285K will be a viable choice for higher end PCs, except for people with 4090 or 5090 GPUs who game in lower resolution and can become CPU limited and gaming performance is the most important aspect of PC performance for them.

 

[Bwaze]

In what exactly are Arrow lake CPUs trailing Zen 5 CPUs except in gaming compared to special 3D variants of the Zen 5 CPUs?

Reviews aren't out yet, so it's a bit early to call. But Intel itself puts gaming on pa with previous generation, which already trailed AMD's previous generation - so the difference with Zen 5 could be significant. Could, of course, 3D variants of Zen 5 could be as underwhelming as regular variants. And of course that wasn't aimed solely on Intel, AMD will also have a hard time selling CPUs that are basically the same as last generation.

And it's funny how gaming vs. productivity debate can shift - when Zen 1 and 2 trumped everything Intel had in productivity but trailed in gaming, a lot of people were pointing out that gaming is really one of the rare things where small differences (10, 20%...) in performance are even measured and significant - productivity is quickly "fast enough", and home CPU's aren't really a best choice for rendering, photo and video editing, if that's your job and you depend on fast PC.

 

[Iain Saturn]

Any guesses to Davinci Resolve performance? Please

 

[Wirko]

Better stick AVX-512 and AVX10 to P-cores only. There just need to make all AVX-512+ call be forwarded to P-cores. I don't know why they still can't make it.

I don't understand what you mean by forwarding but yes, any thread can run on an E-core until an AVX512 instruction turns up, and then there are ways to save it from crashing. One, freeze the thread, save the state and schedule it to continue on a P-core. Two, emulate AVX512 in software, which would be extremely slow, then move the thread to a P-core ASAP.

 

[RogueSix]

The problem is, these CPUs are aimed mainly to people who haven't upgraded in a generation or two. Releasing the generation that barely, or not even, surpasses last generation, and trails competition, is a hard sell right now

It's an easy sell because Intel have conveniently ruined the last TWO generations by letting 13th and 14th gen CPUs degrade to the bone.