Intel ARC GPUs, Xe Architecture for dGPUs [2018-2022]

[Cyan]

Intel Arc A370M Matches Nvidia RTX 3050 Mobile in Official Benchmarks | Tom's Hardware (tomshardware.com)

Intel's Arc A370M launched around two months ago and represented one of Intel's entry-level products in its Arc 3 series of GPUs, accompanied by the lower-end Arc A350M. The Arc A370M comes with 8 Xe cores and a graphics clock of 1550MHz. Memory maxes out at 4GB, featuring GDDR6 modules and a 64-bit wide bus. GPU power will be anywhere between 35-50W, depending on the laptop model.

In 3DMark Time Spy, the A370M notebook scored a rather impressive 4405 points, coming ahead of RTX 3050 laptops, including the HP Spectre notebook with 3772 points and the Vivobook Pro featuring a score of 4396 points.

In Shadow of the Tomb Raider, The A370M outperformed the HP Spectre RTX 3050 notebook, coming out with a 59 FPS average at 1080P on the highest settings. The RTX 3050 only managed a flat 50FPS.

Metro Exodus Enhanced Edition was another game PCWorld tested, with the A370M losing this time to the RTX 3050. The A370 scored 19 FPS while the 3050 gained 22 FPS. Without question, the unplayable frame rates can be attributed to the ray tracing requirements demanded in this specific version of Metro Exodus.

However, the good news is that Intel's A370M can play ray tracing games if that is something you really want. With FSR or Intel's future XeSS upscaling technologies, we wouldn't be surprised to see playable FPS on the A370M with ray tracing enabled.

 

[xpea]

Intel Arc A370M Matches Nvidia RTX 3050 Mobile in Official Benchmarks | Tom's Hardware (tomshardware.com)

Intel's Arc A370M launched around two months ago and represented one of Intel's entry-level products in its Arc 3 series of GPUs, accompanied by the lower-end Arc A350M. The Arc A370M comes with 8 Xe cores and a graphics clock of 1550MHz. Memory maxes out at 4GB, featuring GDDR6 modules and a 64-bit wide bus. GPU power will be anywhere between 35-50W, depending on the laptop model.

In 3DMark Time Spy, the A370M notebook scored a rather impressive 4405 points, coming ahead of RTX 3050 laptops, including the HP Spectre notebook with 3772 points and the Vivobook Pro featuring a score of 4396 points.

In Shadow of the Tomb Raider, The A370M outperformed the HP Spectre RTX 3050 notebook, coming out with a 59 FPS average at 1080P on the highest settings. The RTX 3050 only managed a flat 50FPS.

Metro Exodus Enhanced Edition was another game PCWorld tested, with the A370M losing this time to the RTX 3050. The A370 scored 19 FPS while the 3050 gained 22 FPS. Without question, the unplayable frame rates can be attributed to the ray tracing requirements demanded in this specific version of Metro Exodus.

However, the good news is that Intel's A370M can play ray tracing games if that is something you really want. With FSR or Intel's future XeSS upscaling technologies, we wouldn't be surprised to see playable FPS on the A370M with ray tracing enabled.

Well, 2 important remarks from PCWorld source article:
"Tests performed... at Intel’s Jones Farm campus"
"Intel reference platform and HP Spectre x360 were set to performance power management, while the Acer Nitro 5 was at its default setting"
Not apple vs apple scenario in this strong odor of Intel shenanigans...
And the number of tested games is way too low to discard a heavy cherry pick from Intel.
So, as usual, let's wait for fully independent reviews...

 

[trinibwoy]

Impressive performance from the A370M with those specs especially with half the bandwidth of the 3050. Surely good enough to steal OEM wins from Nvidia.

 

[Putas]

If they did not touch the Vivobook Pro 15 it is the best head to head comparison, since its 3050 uses the same TGP of 35-50 Watts.

 

[xpea]

Looks like we are still far away from ARC launch on the DIY market:
https://community.intel.com/t5/Blog.../Gaming/Engineering-Arc-5-9-2022/post/1383055

by Lisa Pearce, Vice President and General Manager for the Visual Compute Group
We will release our entry-level Intel Arc A-series products for desktops (A3) first in China through system builders and OEMs in Q2. Etail and retail component sales will follow shortly in China as well. Proximity to board components and strong demand for entry-level discrete products makes this a natural place to start. Our next step will be to scale these products globally.

Roll-out of Intel Arc A5 and A7 desktop cards will start worldwide with OEMs and system integrators later this summer, followed by component sales in worldwide channels.

Drivers and COVID are to blame.
Well I guess September is when we will start to see these cards on the shelves... Just in time to compete with Lovelace !
What a disaster of a launch...

 

[Jawed]

I honestly think Intel doesn't care about the competitive ability or timing of Arc - it's a beta test product with which Intel learns how to get serious.

If Intel does annual product refreshes to "catch up" over the following 2 or 3 cards, that will be more interesting. It seems to me that Intel can catch up simply by iterating more rapidly. AMD and NVidia are very cosy with their 2-year cycle. NVidia got to where it is now by originally iterating very rapidly, over 20 years ago...

 

[Putas]

We no longer have the big manufacturing leaps from 20 years ago, so there is little need to move faster.

 

[Jawed]

We no longer have the big manufacturing leaps from 20 years ago, so there is little need to move faster.

I disagree. Architecture and packaging are moving very rapidly.

Ray tracing has shifted the goalposts. The portion of the die being used for ray tracing is currently very small. There's a huge opportunity to re-balance GPUs based upon shifting the balance and adding the right kinds of dedicated hardware - whether that's for sorting rays or caching, for example.

 

[trinibwoy]

Looks like we are still far away from ARC launch on the DIY market:
https://community.intel.com/t5/Blog.../Gaming/Engineering-Arc-5-9-2022/post/1383055

Drivers and COVID are to blame.
Well I guess September is when we will start to see these cards on the shelves... Just in time to compete with Lovelace !
What a disaster of a launch...

Arc coming anywhere near Ampere is a huge positive sign and far from a disaster. It could be so much worse.

 

[DegustatoR]

Arc coming anywhere near Ampere is a huge positive sign and far from a disaster. It could be so much worse.

I mean, an N6 part coming at the same performance as a 8N part from a year ago isn't exactly great. Anything less than that would be a disaster leading to the absence of any reason to even launch it.

 

[trinibwoy]

I mean, an N6 part coming at the same performance as a 8N part from a year ago isn't exactly great. Anything less than that would be a disaster leading to the absence of any reason to even launch it.

I disagree. It’s actually better than expected. The process node is irrelevant. It’s drivers, marketing and pricing that matters. The bar for Intel is very low. If Xe is anywhere in striking distance they can always use pricing to compete. Nvidia is probably enjoying stupid high margins so Intel has a lot of wiggle room.

They also don’t need to match the 4090. Matching the 4060 with this first attempt would be more than enough to make a serious dent. Again it’s going to come down to pricing. Nvidia and AMD seem to be trying to sell midrange cards for $400+.

 

[digitalwanderer]

Is it fair that I find it extremely unlikely that Intel comes out of the gate with their first shiny consumer GPU and is competitive with nVidia and AMD? I mean I'm expecting them to be decent, but it'll be a bit before they get to parity with the other two I think.

Just my thoughts, nothing to really back them up 'cept logic and history....

 

[Albuquerque]

Well, let's be at least a little more fair to them... They've been building video chipsets for decades at this point, even to the point of this NOT being their first attempt at a consumer add-in card (remember the i740?) Intel also set out to deliver better performance with their Iris Pro line with the embedded 64 or 128MB of e-dram cache for all that sweet sweet IGP performance boost? The Broadwell 5775 got a lot of use from it, especially when you had the IGP turned off

Anyway, point being Intel has been dipping various toes into the graphics market for a long time, and some of that time they even had intentions of gaming. No part of me is going to tell you any of their gear was convincingly performant, yet at least they had some rigor around building those chips and the related drivers.

 

[Kaotik]

They've been building video chipsets for decades at this point, even to the point of this NOT being their first attempt at a consumer add-in card (remember the i740?)

i740 was Real3D product/tech really, though

 

[arandomguy]

My feeling based on everything is that Intel is going to be primarily focused on OEM/system integrator sales as opposed to DIY Retail. Their entire GPU strategy is likely focused on being able to offer a complete inhouse CPU and GPU solution to OEMs for both consumers and data centers.

The yearly release cadence is likely also largely influenced by matching their CPU cycle and OEM yearly refresh cycles. Naming and technical marketing aside I'm guessing those releases are going to be more conservative in terms of gains compared to what a longer 2 year cycle would be. I don't know if people should be going in with the expectation that the faster release cycle is to "catch up" with AMD/Nvidia.

With that in mind it also does mean despite the delays they'll still have some breathing room against RDNA3/Lovelace as AMD/Nvidia will likely do a retail focused launch first as per what they typically do. Unless things change also they'll do a top down launch, which means RDNA3/Lovelace GPUs in the same segments may not even hit retail until next year.

 

[DegustatoR]

The process node is irrelevant.

The process node is very relevant as it determines the ability to actually compete in pricing on the market.
Your expectation that Intel will be able to undercut Nvidia and AMD on a more expensive node while producing something like 1/10th of volume is weird. They likely won't.
Which in turn means that if their top GPU will be able to compete only with low/mid range of Nv/AMD they'll be burning money on producing them.

 

[trinibwoy]

The process node is very relevant as it determines the ability to actually compete in pricing on the market.
Your expectation that Intel will be able to undercut Nvidia and AMD on a more expensive node while producing something like 1/10th of volume is weird. They likely won't.
Which in turn means that if their top GPU will be able to compete only with low/mid range of Nv/AMD they'll be burning money on producing them.

What’s weird about undercutting Nvidia’s significant margins? It’s Econ 101 for entering an established market with players who are enjoying large profits. Intel doesn’t have to burn money to lower prices from their current ridiculous levels.

 

[DegustatoR]

What’s weird about undercutting Nvidia’s significant margins? It’s Econ 101 for entering an established market with players who are enjoying large profits. Intel doesn’t have to burn money to lower prices from their current ridiculous levels.

I'll give you two reasons:
1. Assuming that Nvidia's margins in the low end are "significant".
2. Not seeing that these are directly linked to the fact that Nv is using a much cheaper production process with a relatively old architecture.

 

[trinibwoy]

I'll give you two reasons:
1. Assuming that Nvidia's margins in the low end are "significant".
2. Not seeing that these are directly linked to the fact that Nv is using a much cheaper production process with a relatively old architecture.

Yes I’m assuming Nvidia is making good money on 3060’s and 70’s.

We don’t know the economics of TSMC 6nm vs Samsung 8nm. Either way 6nm is an old mature process by now. Given the leaked performance Intel seems to be doing more with less so I’m not sure why you’re so skeptical.

 

[DegustatoR]

Yes I’m assuming Nvidia is making good money on 3060’s and 70’s.

I wouldn't be too sure about that considering that 3060 has a "weird" MSRP and there presumably was some push back on it back at launch from AIBs who were left with very slim margins.

We don’t know the economics of TSMC 6nm vs Samsung 8nm.

We know that Samsung's 8nm was considerably cheaper than N7, no reason to suspect that this has changed with N6 so far.

Either way 6nm is an old mature process by now.

It's been in active use for less than a year. The fact that it is an upgraded N7 doesn't mean much for its costs.

Given the leaked performance Intel seems to be doing more with less so I’m not sure why you’re so skeptical.

I'm not skeptical about that. I'm just not seeing this as a success for Intel as anything less than what we see in leaks would be a disaster. And it is looking increasingly likely that Alchemist will launch against Lovelace/RDNA3 which will change this "leaked" picture somewhat.
What I am skeptical about though is the proposal that Intel will be able to iterate faster than Nv/AMD - don't really see how that would be possible unless the meaning of iterating faster here is launching partial lineups each year (a bit like AMD did with GCN3/Polaris/Vega/RDNA1). In that sense yes, they may launch low end this year, mid range in 2023 and high end in 2024 with a new low end gen coming in 2025.