Always seems to be two years away, like the Aurora supercomputer at Argonne.
Early 2022 is just 4 months away actually
:O ;(
Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
31–40 of 218 posts
Re: Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
#32https://software.intel.com/content/dam/develop/external/us/e...
The above is Intel's Gen11 architecture whitepaper, describing how Gen11 iGPUs work. I'd assume that their next-generation discrete GPUs will have a similar architecture (but no longer attached to CPU L3 cache).
I haven't really looked into Intel iGPU architecture at all. I see that the whitepaper has some oddities compared to AMD / NVidia GPUs. Its definitely "more different".
The SIMD-units are apparently only 4 x 32-bit wide (compared to 32-wide NVidia / RDNA or 64-wide CDNA). But they can be reconfigured to be 8x16-bit wide instead (a feature not really available on NVidia. AMD can do SIMD-inside-of-SIMD and split up its registers once again however, but its a fundamentally different mechanism).
--------
Branch divergence is likely to be less of an issue with narrower SIMD than its competitors. Well, in theory anyway.
Re: Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
#33I really hope this breaks Nvidia's stranglehold on deep learning. Some competition would hopefully bring down prices at the compute high-end. AMD don't seem to even be trying on the software-side at the moment. ROCm is a mess.
ROCm seems to be tolerably decent, if you are willing to spend a couple hours, and, big if, if HIP supports all the various libraries you were relying on. CUDA has a huge support library, and ROCm has been missing not just the small fry stuff but a lot of the core stuff in the that library.
Long term, AI (& a lot of other interests) need to serve themselves. CUDA is excellently convenient, but long term I have a hard time imagining there being a worthwhile future for anything but Vulkan. There don't seem to be a lot of forays into writing good all-encompassing libraries in Vulkan yet, nor many more specialized AI/ML Vulkan libraries, so it feels largely like we more or less haven't started really trying.
Re: Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
#34Earlier quoted context omitted.
Intel has the manufacturing capability to really beat up Nvidia. Even if the cards don’t perform like top-tier cards they could still win bigly here. Very exciting!
it seems AMD manufactures most 7nm all at TSMC, but intel has a factory coming online next year in Arizona... https://en.wikipedia.org/wiki/List_of_Intel_manufacturing_si... I could see gov/Military investing/awarding more contracts based on these 'locally' situated plants
Nope, wiki is wrong. According to Intel, the facility in Chander, AZ will start being built next year, but won't be producing chips until 2024. See https://www.anandtech.com/show/16573/intels-new-strategy-20b...
Re: Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
#35Earlier quoted context omitted.
I wholeheartedly agree. PyTorch did recently release AMD support which I was happy to see (though I have not tested it), I’m hoping there is more to come. https://pytorch.org/blog/pytorch-for-amd-rocm-platform-now-a...
Unfortunately that support is via ROCm, which doesn't support the last three generations (!) of AMD hardware: https://github.com/ROCm/ROCm.github.io/blob/master/hardware....
ROCm supports Vega, Vega 7nm, and CDNA just fine.
The issue is that AMD has split their compute into two categories:
* RDNA -- consumer cards. A new ISA with new compilers / everything. I don't think its reasonable to expect AMD's compilers to work on RDNA, when such large changes have been made to the architecture. (32-wide instead of 64-wide. 1024 registers. Etc. etc.)
* CDNA -- based off of Vega's ISA. Despite being "legacy ISA", its pretty modern in terms of capabilities. MI100 is competitive against the A100. CDNA is likely going to run Frontier and El Capitan supercomputers.
------------
ROCm focused on CDNA. They've had compilers emit RDNA code, but its not "official" and still buggy. But if you went for CDNA, that HIP / ROCm stuff works enough for the Oak Ridge National Labs.
Yeah, CDNA is expensive ($5k for MI50 / Radeon VII, and $9k for MI100). But that's the price of full-speed scientific-oriented double-precision floating point GPUs these days.
Re: Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
#36I really hope this breaks Nvidia's stranglehold on deep learning. Some competition would hopefully bring down prices at the compute high-end. AMD don't seem to even be trying on the software-side at the moment. ROCm is a mess.
ROCm seems to be tolerably decent, if you are willing to spend a couple hours, and, big if, if HIP supports all the various libraries you were relying on. CUDA has a huge support library, and ROCm has been missing not just the small fry stuff but a lot of the core stuff in the that library. Long term, AI (& a lot of other interests) need to serve themselves. CUDA is excellently convenient, but long term I have a hard…
is there any indication that ROCm has solved its stability issues? I wasn't doing the testing myself, but the reason why we rejected ROCm a while back (2 years?) was because you could get segfaults hours into a ML training run, which is... frustrating, to say the least, and not easily identifiable in quickie test runs (or CI, if ML did more CI).
Re: Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
#37Interesting, but the add-in-card GPU market for graphics purposes is so small, it's hard to get worked up about it. The overwhelming majority of GPU units sold are IGPs. Intel owns virtually 100% of the computer (excluding mobile) IGP market and 70% of the total GPU market. You can get almost the performance of Intel's discrete GPUs with their latest IGPs in "Tiger Lake" generation parts. Intel can afford to nibble at the edges of the discrete GPU market because it costs them almost nothing to put a product out there and to a large extent they won the war already.
Re: Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
#38As TFA rightly points out, unless something drastically changes in the next ~6mo, Intel is going to launch into the most favorable market situation we've seen in our lifetimes. Previously, the expectation is that they needed to introduce something that was competitive with the top end cards from nVidia and AMD. With basically all GPU's out of stock currently they really just need to introduce something competitive wi…
Intel has the manufacturing capability to really beat up Nvidia. Even if the cards don’t perform like top-tier cards they could still win bigly here. Very exciting!
What about design capabilities? If they had it in them what were they doing all these yrs? i mean since 2000 i can remember a single GPU from intel that wasnt already behind the market.
Re: Intel’s Arc GPUs will compete with GeForce and Radeon in early 2022
#39Earlier quoted context omitted.
Intel has the manufacturing capability to really beat up Nvidia. Even if the cards don’t perform like top-tier cards they could still win bigly here. Very exciting!
Intel is not even manufacturing these, they are TSMC 7nm so they are competing for the same fab capacity that everyone else is using.
*AMD/Apple is using. Nvidia's always-sold-out Ampere-based gaming chips are made in a Samsung fab.