News Intel's lackluster Arrow Lake appears to have a refresh inbound — Arrow Lake Refresh appears in reference document

[bit_user]

Which isn't even a DIY consensus. I'm a DIY'er and I replace my (desktop) PCs only after 5+ years. Socket longevity is irrelevant.

In the past, I've always had to buy a new motherboard, each time I decided it was time for an upgrade. My typical upgrade cycle has been 5-10 years, BTW.

However, I currently have a 12th gen i5, and would consider replacing it with a Bartlett Lake i7 or i9, if that proves decent. The RAM I have is DDR5-5600 (ECC), which is the fastest ECC UDIMM absent a CKD that almost certainly won't be supported by Bartlett Lake's memory controller. Since I use ECC RAM, I'm limited to the W680 motherboard chipset, and there's no better W680 board out there, for my purposes. So, if I do it, then it'd indeed be a CPU-only upgrade.

Also, I recently picked up a cheap 9600X + mobo + memory bundle. I really just wanted a cheap way to fiddle around with AVX-512 and I already had all the other parts needed to create another Linux machine. I had tried using an Amazon cloud instance, but I got weird results when trying to run benchmarks on it, and it was generally more of a hassle than I prefer to deal with. Now that I've bought into the AM5 ecosystem, I can definitely see myself continuing to build up that machine, according to circumstances and my needs. As you can imagine, the board is a basic (but not bad) B650 and the RAM was only a single 16 GB DIMM. I could even see myself doing a mobo-only upgrade, if I end up buying a PCIe 5.0 dGPU to put in that machine.

TL;DR: while I've always been in the same boat as you, I might now have two examples of where I'd be doing CPU-only or mobo-only upgrades!

 

[halfcharlie]

I'll point it out as often as it needs to be. There is no real world impact on gaming from your CPU if it's from any time in the last 10 years. This has been proven with recent benchmarks going back to 10 year old i5's losing less than 10 FPS compared to the current gen CPU's both matched with a powerful modern card. If what you want is gaming performance put your money into the GPU, lots of RAM (Oblivion Remastered destroys my 64GB), and a fast system drive. It's as simple as that. Anything associating gaming performance and CPU's is misleading at best, and gaming certainly had no effect on Arrow Lake's failure, the CPU is the least important aspect for gaming other than something benign like the case itself. If you also use your system for productivity like I do, that's when you put money into a CPU beyond entry level as well.

Anyway this would put me in a hard place. I'm in desperate need of an upgrade, mostly for USB ports and RAM, I skipped Arrow Lake because it still wasn't quite what I wanted and neither were the motherboards. I don't want to wait another year either, but with a 5 year save and upgrade cycle I have to get it right.

 

[thestryker]

gaming certainly had no effect on Arrow Lake's failure,

In the DIY space it's a reasonable assumption that this is the entirety of the reason. ARL is roughly equivalent to AMD offerings (not counting AVX512) and superior to RPL outside of gaming and it has a better platform than either.

Anyway this would put me in a hard place. I'm in desperate need of an upgrade, mostly for USB ports and RAM, I skipped Arrow Lake because it still wasn't quite what I wanted and neither were the motherboards. I don't want to wait another year either, but with a 5 year save and upgrade cycle I have to get it right.

There's no sign that AMD is changing the inherent flaw in their platform for connectivity so what's on offer now is probably as good as it gets from them. I'd say ARL has the best client platform currently for connectivity unless you need PCIe 5.0 x4 slots as some X670 boards have these. Integrated TB4 and the dedicated PCIe 5.0 x4 and PCIe 4.0 x4 for M.2 just give ARL a leg up and the PCIe 4.0 x8 DMI connection leads to more chipset connectivity.

 

[bit_user]

There's no sign that AMD is changing the inherent flaw in their platform for connectivity

Have you heard definite confirmation their next gen chipset won't upgrade the link to PCIe 5.0? As you point out, that's clearly a weak spot, but it's also one that's easily addressed (at least, to the extent of reaching parity with Intel).

Integrated TB4

This is the only place AMD definitely won't catch up, but it's also one where I have no current use case. It reminds me of back in the Pentium 4 days. I liked the idea that my motherboard had a Firewire port, but I wouldn't have paid extra for that one feature and never did end up using it.

I know that's a slightly flawed comparison, since USB 4 does indeed look set to become dominant, but I'm not going to use it for monitors and my USB 3 connectivity needs have always been so modest that I couldn't even tell you which generation or version of the standard which of my various boards support.

 

[thestryker]

This is the only place AMD definitely won't catch up, but it's also one where I have no current use case.

AMD's X870 series chipsets mandate USB4 which takes away from PCIe (I've seen some boards where it's shared, but this doesn't seem common) so that's where I see the advantage for overall platform whether you use it or not. Of course X670 based boards still seem to have decent availability, but I can't imagine that lasts through the next generation.

Have you heard definite confirmation their next gen chipset won't upgrade the link to PCIe 5.0? As you point out, that's clearly a weak spot, but it's also one that's easily addressed (at least, to the extent of reaching parity with Intel).

I don't know if they could do that and maintain backwards compatibility. I know for 11th Gen Intel moved to 8 lanes, but kept DMI data rate at PCIe 3.0 for Z590 and maintained backwards compatibility to 10th Gen which had 4 lanes. If compatibility was the reason Intel did that then AMD could of course follow suit and move to 8 lanes, but the question then becomes chipset strategy and backwards compatibility.

 

[bit_user]

I don't know if they could do that and maintain backwards compatibility.

AM4 spanned both PCIe 3.0 and 4.0.

​

I think newer AM5 chipsets could upgrade the link to PCIe 5.0 (when paired with a Zen 6 CPU), if they wanted.

 

[mac2213]

The year is nearly half done and last October Intel introduced Arrow Lake K skus. Does it make sense to refresh Arrow Lake this summer only to release Nova Lake K skus in the fall. MJ recently confirmed that Nova Lake Desktop skus will rely on TSMC 3nm indicating that the process is available to Intel today. There was also a comment regarding packaging from TSMC during her recent BofA call (and the lack of HBM cache due to packaging was Intel's real issue on Arrow Lake). Is it possible that this Arrow Lake refresh data is really a place holder for an early release of TSMC 3nm Nova Lake K skus.

 

[thestryker]

The year is nearly half done and last October Intel introduced Arrow Lake K skus. Does it make sense to refresh Arrow Lake this summer only to release Nova Lake K skus in the fall. MJ recently confirmed that Nova Lake Desktop skus will rely on TSMC 3nm indicating that the process is available to Intel today. There was also a comment regarding packaging from TSMC during her recent BofA call (and the lack of HBM cache due to packaging was Intel's real issue on Arrow Lake). Is it possible that this Arrow Lake refresh data is really a place holder for an early release of TSMC 3nm Nova Lake K skus.

NVL is a 2026 part and is almost certainly not using N3. There have been rumors about TSMC N2, but they're still rumors and nobody knows if it's whole stack or just part.

HBM isn't viable for any sort of caching in a consumer level part due to cost.

 

[TerryLaze]

The year is nearly half done and last October Intel introduced Arrow Lake K skus. Does it make sense to refresh Arrow Lake this summer only to release Nova Lake K skus in the fall.

It makes sense if they found a serious flaw or a decent performance improving design change.
Or if they want to move it over to 18A to reduce the cost of production.

Or the poor manual maker was just trying to be thorough and put in the refresh because many CPU gens do get refreshes and any refresh would still work in that mobo, doesn't mean that there will be one, but if there would be it would work.

 

[bit_user]

HBM isn't viable for any sort of caching in a consumer level part due to cost.

And supply shortage. I'm not sure if this is still true... but, until recently, HBM was the limiting factor of AI training chips. So, there's no spare HBM production capacity left for lower-margin consumer stuff.

Also, when Nvidia detailed their Grace CPU, they said:

"An HBM2e memory subsystem would have provided substantial memory bandwidth and good energy efficiency but at more than 3x the cost-per-gigabyte and only one-eighth the maximum capacity available with LPDDR5X."

Source: https://developer.nvidia.com/blog/nvidia-grace-cpu-superchip-architecture-in-depth/
​

So, I expect CPUs will not rely solely on HBM. They'll either go with some LPDDR memory, like Apple's M-series, or they'll use HBM as a fast "near" memory, paired with something like CXL memory to scale capacity.