Intel Integrated Graphics Performance From Gen9 To Meteor Lake Arc Graphics

[blackiwid]

The 155H's GPU tile is on TSMC's N5 process, it's SOC and IO tiles are on N6, and it's CPU tile is on the Intel 4 process. The 7840U is one monolithic die that uses TSMC's N4 process. In other words AMD's iGPU has the advantage completely when it comes to it's manufacturing node.

You can't just compare the numbers N6 TSMC is not 1:1 comparable to intel N6.

Both can support memory up to 7400MT when using LPDDR5X so both can be made faster. If the 7840U used higher clocked RAM then it would close the performance gap but would also increase the power used by the memory controller and increase the actual GPU utilisation.

What Performance Gap the gap is much smaller than the RAM % difference and the performance difference could even scale MORE than 1:1 with the ram because let's assume the cpu and gpu uses 50% of the ram with the slower ram, if the additional speed would be used only by the gpu that could mean with let's say 20% more ram speed that the gpu has 40% more ram speed for it's work, but even if it were only 20% and news flash igps / apus scale nearly 1:1 with ram speed because that is with better ones the bottleneck:

it got an average of 64.70 FPS with a 93.77 FPS maximum, compared to the Core Ultra 7 155H's 67.7 FPS average and 93.7 FPS maximum.

And that is the "centrino moment" from Intel they just shot all their powder for a long time, AMD will release soon a big own thing, and still with the same ram according to this posted numbers from the 1st site of this comments would be ahead with their old systems. Speed wise at least.

[Myownfriend]

You can't just compare the numbers N6 TSMC is not 1:1 comparable to intel N6.

I'm not. I'm comparing TSMC N5 and N4. Intel's GPU tile uses the former and AMD's entire APU uses the latter. Only the 155H's CPU tile isn't made at TSMC.

What Performance Gap the gap is much smaller than the RAM % difference and the performance difference could even scale MORE than 1:1 with the ram because let's assume the cpu and gpu uses 50% of the ram with the slower ram, if the additional speed would be used only by the gpu that could mean with let's say 20% more ram speed that the gpu has 40% more ram speed for it's work, but even if it were only 20% and news flash igps / apus scale nearly 1:1 with ram speed because that is with better ones the bottleneck:

Sure it could scale more than linearly if the GPU is being bottlenecked by memory bandwidth but it's not realistic that the additional bandwidth would be used only by the GPU. If the GPU is completing more frames then that means the CPU has to set up, calculate collisions, etc. for more frames. The majority of the additional bandwidth would go to the GPU, sure, but not all of it.

The bandwidth difference is 14% and the gap in theoretical compute performance (clock speed x ALUs) is 11%. That would put both iGPUs (by this simple math which ignores architectural differences) at about as equally bandwidth starved. The first benchmark in the article puts the 155H almost 18% ahead of the 7840U so how is the performance gap much smaller than the RAM speed difference?

I feel like people are trying really hard to make Xe seem less impressive. It's matching or beating RDNA3 while using quite a bit less power despite being on an older fab node. That's impressive.

And that is the "centrino moment" from Intel they just shot all their powder for a long time, AMD will release soon a big own thing, and still with the same ram according to this posted numbers from the 1st site of this comments would be ahead with their old systems. Speed wise at least.

Literally have no idea what you're talking about.

[drakonas777]

I feel like people are trying really hard to make Xe seem less impressive.

Depends on what exactly people find to be impressive. For me personally achieving the parity with competitor is not impressive, it's an expected improvement. A bare minimum to be honest.

[coder]

For me personally achieving the parity with competitor is not impressive, it's an expected improvement. A bare minimum to be honest.

AMD's Radeon 780M was, itself, a big generational jump in performance. For Intel to catch up to it would be like AMD releasing a RX 8900 XT that matches Nvidia's RTX 5090. I think that would be pretty impressive.

[drakonas777]

AMD's Radeon 780M was, itself, a big generational jump in performance. For Intel to catch up to it would be like AMD releasing a RX 8900 XT that matches Nvidia's RTX 5090. I think that would be pretty impressive.

780M is not at all a big generational jump compared to 680M. Analogy is also invalid, since Intel is responding to the nearly year old competition. It's more like 8900 XT matches 4090, though I see no point in hypothetical analogies.

[coder]

780M is not at all a big generational jump compared to 680M.

It's pretty consistently about 15% faster.

Analogy is also invalid, since Intel is responding to the nearly year old competition.

Phoenix laptops didn't start shipping in volume until about 6 months ago. The theoretical launch date was only about 8 months ago.

The fact of the matter is that it's what AMD has on the market, and Hawk Point won't change that.

As mentioned above, all of this is in spite of Intel being at a half-node disadvantage (i.e. N5 vs. N4, which is an optimized N5-derivative).

[drakonas777]

It's pretty consistently about 15% faster.

Well, that's not a big generational jump for me. Especially in the context of historical iGPU generational jumps where 50+% improvements were not that uncommon.

But that's not my main point. I just don't subscribe to the way of thinking where improvements are analyzed in the context of the same vendor existing products. I always analyze them in the broader competitive landscape. This is the reason why I don't find original ZEN core that impressive for example. Sure, 50+% IPC improvement over bulldozer sounds impressive on it's own, until you realize that ZEN core was a rough equivalent to the Broadwell, which was already dated at the time. It was a good core, not impressive though. I'm not saying this is comparable to Mateor Lake GPU situation, just explaining my position.

[coder]

I always analyze them in the broader competitive landscape.

That's what we're doing.

Intel's previous generation was behind AMD, and now they've leapfrogged AMD, in spite of being at a slight node disadvantage and using less power!

[qarium]

It's 5600MT for the 7840U and 6400MT for the 155H so yea, that could be factor but both are sharing their RAM with the CPU and other blocks in the chip so they're factor, too.
Yes, the 155H's GPU has 33% more shader cores but the 7840U's GPU runs 20% higher clock speeds. If normalize there would be an 11% gap in theoretical performance. That would explain the performance gal in some tests but not the power gap. If the 7840U were clocked higher to match the 155H's performance than it's efficiency would drop and the power gap would increase.
But also...
...they're different architectures so the size of the cores, their layout, cache configuration, IPC, etc, is all very different. Individual cores going to use a different amount of transistors depending on the architecture.
The 155H's GPU tile is on TSMC's N5 process, it's SOC and IO tiles are on N6, and it's CPU tile is on the Intel 4 process. The 7840U is one monolithic die that uses TSMC's N4 process. In other words AMD's iGPU has the advantage completely when it comes to it's manufacturing node.
Both can support memory up to 7400MT when using LPDDR5X so both can be made faster. If the 7840U used higher clocked RAM then it would close the performance gap but would also increase the power used by the memory controller and increase the actual GPU utilisation.
We don't know that AMD's GPU uses less transistors. We just know its transistors are tinier and better packed because it uses a new manufacturing process. If Intel's GPU used TSMC 4nm then it's iGPU would be smaller than it is, it could be clocked higher, and it would be even more efficient.
Not necessarily. Older processes are cheaper to manufacturer and have higher yields. Intel is also using tiles/chiplets here while AMD is using a monolithic die.
We don't know that AMD's iGPU is using few transistors. Intel's GPU tile is pretty small. In terms of execution units, it's the same as an A380 but the memory controller, PCI-E, media engines, audio engines, and display engines have been moved into the SOC tile.
Why? What does that even mean?

thank you for the memory update. its relative simple to make a die shot for the AMD chip die and on the die shot you see the party what is gpu and what is CPU...

i don't think that for this type of SOC there is a need for a chiplet design. a monolitic design looks better to me.

if you say it is the same shader numbers and same architecture than the Intel ARC A380
"Process Size: 6 nm. Transistors: 7,200 million. Density: 45.9M / mm². Die Size: 157 mm²."

so we have a die shrink to 5nm but its save to say that this iGPU has something like 7,2 million transistors.

does anyone know any numbers for the AMD igpu ?

[qarium]

That's what we're doing.
Intel's previous generation was behind AMD, and now they've leapfrogged AMD, in spite of being at a slight node disadvantage and using less power!

true but amd will just put in faster ram and will increase the clock speed a little and then they have the same performance

and intel will use a little less power. but most of the people will not care about this if the price is cheaper.

and then in less than 5 month we have RDNA4 gpu on the AMD side