Monday, August 22nd 2022
Intel 14th Gen "Meteor Lake" APUs Reportedly Feature Ray Tracing, May Lack XeSS
Intel's future Meteor Lake APUs seem to be playing catch-up to AMD's integrated graphics in more ways than one. Twitter user Coelacanth's Dream has dug up information that indicates Intel's commitment to bring ray tracing support to even its IGP (Integrated Graphics Processing) tiles. According to bits and pieces from Intel Graphics Compiler (IGC) code patches, it seems to be confirmed that ray tracing support is indeed coming to the TSMC-made, 3 nm GPU tiles in Meteor Lake. The kicker here is the presence of flags that detect whether the iGPU is of the "iGFX_meteorlake" type - if so, IGC sets ray tracing support to enabled.
Puzzlingly, Intel's upscaling technology, Xe SuperSampling (XeSS) could be out of the picture - at least for now. It seems that IGC patches for the upcoming APU family still don't allow for DPAS (Dot Product Accumulate Systolic) instructions - instructions that rely on XMX (Intel Xe Matrix Extensions), the AI engines responsible for executing 128 FP16/BF16, 256 INT8, or 512 INT4/INT2 operations per clock. These low-precision operations are the soul of algorithmic supersampling technologies such as XeSS.
It seems strange that Intel would go out of its way to enable ray tracing on its APUs, which will necessarily feature few of the corresponding hardware accelerators due to die size constraints (limiting ray tracing performance), rather than using that same space for XMX accelerators, which could help improve performance through access to XeSS. Of course, these are all bits and pieces gleaned from IGC, and Intel still has work ahead before Meteor Lake ever hits the market.
Considering Intel's recent execution woes, it remains to be seen if Meteor Lake will see the light of day in or around its planned launch. A cadre of new technologies are being integrated at the same time - such as the CPU tile, which is being manufactured in Intel's new manufacturing process, Intel 4. The GPU tile itself has been confirmed by Intel to make use of TSMC's 3 nm manufacturing process, but all may not be well over at graphics land; it seems that Intel has elected to stop GPU tile production at the 3 nm process on TSMC's foundries. Speculation places Intel as electing to move the manufacturing process towards the more performant and power-efficient TSMC N3E process, which could make sense considering power efficiency requirements of APU designs.
Sources:
Coelacanth's Dream, via Tom's Hardware
Puzzlingly, Intel's upscaling technology, Xe SuperSampling (XeSS) could be out of the picture - at least for now. It seems that IGC patches for the upcoming APU family still don't allow for DPAS (Dot Product Accumulate Systolic) instructions - instructions that rely on XMX (Intel Xe Matrix Extensions), the AI engines responsible for executing 128 FP16/BF16, 256 INT8, or 512 INT4/INT2 operations per clock. These low-precision operations are the soul of algorithmic supersampling technologies such as XeSS.
Considering Intel's recent execution woes, it remains to be seen if Meteor Lake will see the light of day in or around its planned launch. A cadre of new technologies are being integrated at the same time - such as the CPU tile, which is being manufactured in Intel's new manufacturing process, Intel 4. The GPU tile itself has been confirmed by Intel to make use of TSMC's 3 nm manufacturing process, but all may not be well over at graphics land; it seems that Intel has elected to stop GPU tile production at the 3 nm process on TSMC's foundries. Speculation places Intel as electing to move the manufacturing process towards the more performant and power-efficient TSMC N3E process, which could make sense considering power efficiency requirements of APU designs.
25 Comments on Intel 14th Gen "Meteor Lake" APUs Reportedly Feature Ray Tracing, May Lack XeSS
Typical Intel.
AMD's integrated graphics suck. There I said it.
Look, I know the new RDNA2-infused 6000-series APUs have the best integrated graphics ever, but AMD still don't dedicate much silicon area to the IGP. 12CUs on the fully-enabled silicon is still pretty anemic. You can't run AAA games at 1080p still, you still can't really run some AAA games at all, and you still get terrible IGPs with the Ryzen 5 models that lose HALF of the CUs and run about 25% slower as well.
Honestly, for a general-purpose machine without a dGPU, the IGP should use more die area than the CPU cores. AMD currently allocate almost 2:1 ratio of die area between CPU and IGP. Intel's GT3 models are the opposite - the IGP dwarfs the CPU cores.
If Intel's IGPs get better, AMD will have to redress the balance between their CPUs and IGPs. 8C/16T are rarely a necessity in an APU, especially when most of them aren't given enough TDP to run all 8 cores at decent clockspeeds simultaneously. 6C/12T and giving over the extra die area to a 20CU or 24CU IGP would be a huge improvement. It would likely unlock 1080p60 gaming for most titles and enable some of the heavier AAA titles to run at playable framerates with reduced settings.
idk man, I think AMD is pretty impressive considering that thing is just build into the cpu.
But sure, moar is better, tech progress needs to keep going and competition there is good.
In fact there's strong evidence to suggest that for the sort of resolutions that old (2017) Vega8 was capable of, even DDR4-2400 was more than enough. Adding in Infinity Cache to the IGP removes that dependence on bandwidth, and that bandwidth has more than doubled despite being plenty five years ago. At the very least, whilst it's possible to improve performance by adding bandwidth, there's still a lot of performance to be gained by adding compute units. Going all the way back to the 2700U in 2017, there has never been a bandwidth bottleneck, always a shader bottleneck caused by AMD being too stingy with silicon area dedicated to the IGP. I'm 100% sure there has always been a bandwidth bottleneck at some size of IGP, but AMD have never come close to it. Desktop APUs proved that because you could add stupidly-expensive, silly-fast RAM with 80% more bandwidth and far lower latency to them and it would add maybe 5-10% more performance over 2133 dual-channel, by 2400 or 2666 you were already looking at diminishing returns.
Just so we're lear here, a threadripper with 4 3200 dimms hits 95 GB/s of memory bandwidth, and that has to feed both a hypothetial iGPU and the system itself. a 5500xt alone has 224GB/s. The old 1650 non super has 128GB/s.
Your performance would be total dog feces.
If I am not wrong AMD Zen4 laptop SKUs will be featuring ~1500 cores in their APU as well and if u go by the leaks the performance was close to 3050M in paper.
Even with one of the older 2700U laptops we still have, the 2.2GHz Zen1 quad-core is more than enough CPU for general use but the IGP is woefully underequipped for almost any AAA title from the last decade. Tomb Raider 2013 on minimum settings at 720p? 18fps. CPU cores almost entirely idle as the IGP chokes on a lack of shader/compute power.
Take the two new 6800U laptops I just bought. With dual-channel DDR5, ~2GHz of RDNA2 cores across 12CUs, it's immensely impressive for a laptop IGP, and yet it's still barely enough to manage 1080p in older/simpler titles like 2015's GTA5 or F12020 which even low-end hardware like the insultingly-bad GTX 1630/RX6400 can manage at 100fps with ease. Outside of old games and extremely light e-sports titles, it's simply not enough for modern AAA gaming, even at reduced settings. The XBox Series S is a 75W device (from the wall) which is about what Zen3 ultrabooks pull from the wall under full boost - and the XBSS is on an older process node with less efficient CPU cores and 20CU of RDNA2. AMD clearly *can* make a 20CU IGP, they just don't have any incentive to due to a lack of competition, and that's what I want to see change.
I played CS:GO and Strange Brigade. With the last one 720p ran OK. I was impressed. Strange Brigade is not a wimp of a game though and it still ran in a manner, I could literally play and enjoy it.
I think it is enough for the time being and AMD has been increasing the iGPU performance ever since. Now you get a totally different iGPU in current AMD APUs. I hope these improvements wont stop but there should be no sacrifices in the CPU performance. I think AMD could make iGPU with higher CU's without impacting CPU performance but it also would increase the price for the product in few areas and I don;t think they want to do that. It is still an APU and maybe the gains from that move are not that impressive to what it would cost.
That means that If you specifically buy an APU with the stronger IGP, it's because the basic IGP of a regular Zen4 CPU isn't enough and you absolutely NEED more GPU performance; With the current Zen3/RDNA2 APUs, the extra CPU cores are wasted when the anemic IGP is bottlenecking everything. Hell, 6C/12T might still be overkill for a 12CU IGP; If you watch modern games running on a 5700G, for example, the CPU is sitting at 20% (effectively less than a dual-core) whilst the IGP is always running flat-out and still struggling to provide an enjoyable experience compared to even some of the more dated low-end dGPUs.
If I had the choice of a 4C/8T 24CU part or an 8C/16T 12CU part, I'd pick the quad-core every time. You can count the number of things that benefit from more than 8 threads on the fingers of one hand, whilst almost every AAA game from the last decade will be instantly and dramatically improved with twice the GPU power. The people who are software-encoding and software-rendering in Blender/Cinema4D etc are not the target audience for an APU, I don't think.
Desktop APUs never made any sense to me; they cost like $100 more than the regular CPU variant, but they perform worse because they only have half the L3 cache. For that $100 you could always just get a much better dGPU and run absolute circles around the IGP. Even in the darkest depths of the 2020 shortages you could still find old 2GB GTX 960 cards for $100 that were better than the best Vega8 IGP AMD were offering at the time.