Intel Meteor Lake Arc Graphics: A Fantastic Upgrade, Battles AMD RDNA3 Integrated Graphics

[Jumbotron]

I am duty and honor bound to amend my earlier thoughts and statements concerning Michael's earlier benchmarks of the CPU itself. Intel's iGPU makes up for a somewhat disappointing showing vs AMD's Zen 4. And even more curious was the power usage figures here are more competitive, indeed, impressive than what was shown in the strictly CPU benchmarks.

Maybe, Pat Gelsinger is steering his crew at Intel in a good direction. Now let's see if Intel can continue to execute going forward. After all...Zen 5 is waiting in the wings and soon, along with RDNA 4.

[blackshard]

Notebookcheck did also some tests und luckily there some bandwidth numbers for the (LP)DDR5 RAM:
- Acer-Swift-Go-14: https://www.notebookcheck.net/Acer-S....783349.0.html
- Framework Laptop: https://www.notebookcheck.net/Framew....756613.0.html
- HP-EliteBook-840 with 1280P: https://www.notebookcheck.net/HP-Eli....674583.0.html

----

Laptop / CPU	Acer / 155H	Framework / 7840U	HP / 1280P
RAM Type	LPDDR5-6600	DDR5-5600	DDR5-4800
Copy Bandwidth (AIDA64 / Notebookcheck)	80515 MB/s	62500 MB/s	56337 MB/s
Relative Bandwidth	100%	~77.6%	~70%

---
Edit, the RAM type/speed of the Acer could be wrong, imho is an indication that it could be LPDDR5-6600 based on the CPU-Z Screenshot from Notebookcheck, see: https://www.notebookcheck.net/filead...-14-120008.png

This makes the thing a bit more clear. Looking again at the benchmarks, it turns out that:

- synthetic tests (like vkpeak) mostly favour intel - interestingly they show very different architectures behind the scene
- real games show often tied results, with similar performance and similar efficiency (slightly favouring intel by a small margin)

DDR5 vs. LPDDR5 can be the real game changer here: 25% and more bandwidth for LPDDR5 may result in 25% more in-game performance and a bunch of saved watts, expecially during demanding game benchmarks.

[aufkrawall]

Can it now run Unreal Engine 5 on Windows, or even on Linux VKD3D-Proton? Unfortunately, the picked test cases for the article tell almost nothing about real-world performance with Wine/Proton gaming.

[mikk]

DDR5 vs. LPDDR5 can be the real game changer here: 25% and more bandwidth for LPDDR5 may result in 25% more in-game performance and a bunch of saved watts, expecially during demanding game benchmarks.

It is not a game changer, it is good enough for 5-10% and that's it. You are way over the top with your expectation based on just the theoretical bandwidth gain. Also the power saving is irrelevant for the Soc performance and efficiency.

[kiffmet]

aufkrawall AFAIK Intel's Linux GPU drivers still lack support for some performance-critical Vulkan extensions for VKD3D. This should change when the new Arc kernel driver releases and Mesa wires up the new interfaces.

Thread: Interestingly enough, the Windows GPU results for the Core 7 155H (see the notebookcheck link from blackshard) weren't quite as favorable, with the Ryzen 7840U in the Ayaneo Air 1S Retro Power (max. sustained pwr. draw allowed 25W) being 10% faster on average.

In 2 of the tested games (Dota 2 and X-Plane 11.11), the Arc iGPU managed to be a whopping 30% faster though. It seems like Meteor Lake's GFX performance is still rather workload dependent, but much less so than the Alchemist desktop cards.

Sites that got a review sample of laptops with the new Metor Lake chips did often receive one with LPDDR5x-7567 RAM - these configurations are said to not hit retail shelves in significant quantities and/or only at a significant mark-up. Be aware of that when reading/watching tests.

[andre30correia]

some proton games to see if intel works well or not is a good idea for a next benchmark, but this is was expected we will see the response of amd to this ones since the last igpu is the same with only higher clocks 680/780m are the same thing no upgrades from amd from some time in this area

[coder]

Notebookcheck did also some tests und luckily there some bandwidth numbers for the (LP)DDR5 RAM:
- Acer-Swift-Go-14: https://www.notebookcheck.net/Acer-S....783349.0.html
- Framework Laptop: https://www.notebookcheck.net/Framew....756613.0.html
- HP-EliteBook-840 with 1280P: https://www.notebookcheck.net/HP-Eli....674583.0.html

----

Laptop / CPU	Acer / 155H	Framework / 7840U	HP / 1280P
RAM Type	LPDDR5-6600	DDR5-5600	DDR5-4800
Copy Bandwidth (AIDA64 / Notebookcheck)	80515 MB/s	62500 MB/s	56337 MB/s
Relative Bandwidth	100%	~77.6%	~70%

---
Edit, the RAM type/speed of the Acer could be wrong, imho is an indication that it could be LPDDR5-6600 based on the CPU-Z Screenshot from Notebookcheck, see: https://www.notebookcheck.net/filead...-14-120008.png

Thanks! I had the same thought and ended up at the same place.

I did go one step further and actually check the AIDA64 memory benchmarks. I compared it to the results from the HP EliteBook 840 G9, according to the same site, which they also claim has DDR5-4800 (as you stated).

What I found was very intriguing.

Benchmark	HP EliteBook 840 G9	Acer Swift Go 14	Improvement	Est. Speed
Copy	56337	80515	42.9%	DDR5-6860
Read	65708	72244	9.9%	DDR5-5277
Write	58729	75291	28.2%	DDR5-6154
Latency	147.1	94.5	-35.8%	CAS@4800=62

We should keep in mind that these benchmarks probably measure the bandwidth as seen by a single core, yet it might not have access to the entire DRAM bandwidth of the SoC. Sometimes, multiple threads are needed to achieve maximum bandwidth. Even then, you often don't see the entire DRAM bandwidth without also engaging the iGPU and/or NPU blocks.

Also, let's not forget that Meteor Lake changed up the SoC interconnect, rather substantially. The CPU cores now have to jump across to the SoC tile, where the memory controller is located. I don't know if the topology is still a ring, or if maybe they went with the mesh they use in their server CPUs.

The large jump in latency would seem to confirm it's LPDDR5. I'm not sure if LPDDR5x adds yet more latency.

[coder]

It is not a game changer, it is good enough for 5-10% and that's it. You are way over the top with your expectation

Whatever the reasons, you can see from my above post (# 27) that there are indeed some massive jumps in bandwidth seen by the CPU cores. I'm somewhat at a loss to explain why read performance improved so much less than the others. Maybe it's something to do with the huge bump in latency.

Ultimately, what counts for GPU performance is the bandwidth available to the GPU, and that's best measured by synthetic benchmarks running on the GPU, itself. Sadly, the SHOC Texture Read Bandwidth benchmark is clearly showing us cache performance, rather than actual DRAM bandwidth.

Michael , could you please run the CLPeak: Global Memory Bandwidth test on at least the two Intel laptops?

https://openbenchmarking.org/test/pts/clpeak&eval=e05e38fa90546ff4d13bc1f0e8aa16d1344116 a0#metrics

Also the power saving is irrelevant for the Soc performance and efficiency.

Maybe the relative power savings isn't too significant at full load, but it's a bigger deal for battery life at low-utilization, which is how most people use laptops.

[blackshard]

It is not a game changer, it is good enough for 5-10% and that's it. You are way over the top with your expectation based on just the theoretical bandwidth gain. Also the power saving is irrelevant for the Soc performance and efficiency.

Depends on the workload. If the gpu is powerful enough, dram is the bottleneck. Consider a small increment in performance due to more bandwidth, a small decrement in power consumption due to low power parts and performance per watt will be reversed or parified

[qarium]

"battles AMD RDNA3", such an euphemism. The results say is bests AMD RDNA3 in almost every benchmark, AMD's wins are mostly on Tesseract. And then you realize AMD is also a no show for compute...
I expect the graphics/gaming aspect will not even have that big an impact, the IGP will still fall into the "light gaming" (whatever that means) category. But compute tends to be increasingly more present, things like Photoshop or Darktable will put OpenCL to good use when available, actually speeding up stuff that needs to be sped up.

the amd cpu is 20% faster why don't you do compute on the cpu ? all relevant compute projects like PyTorch support the CPU target.

right now i see it is like this: you can choose between AMD who is 20% faster on CPU and the Intel option who is 8% faster in the GPU part.

the freaks who buy apple m1/2/3 will love the new intel device because of the idle consumtion of 4watt for AMD and only 1watt for the intel one.
but i am sure these people who do like the new idle consumtion of only 1 watt they buy apple m1/2/3

"But compute tends to be increasingly more present, things like Photoshop or Darktable will put OpenCL to good use when available, actually speeding up stuff that needs to be sped up."

in many cases the stuff run faster on the CPU than per compute on the GPU....