4 ms·

Consumer gpus are totally different products from the high end gpus now. Intel has failed on the gpu market and has effectively zero market share, so it is not

Consumer gpus are totally different products from the high end gpus now. Intel has failed on the gpu market and has effectively zero market share, so it is not actually clear there is an antitrust issue in that market. It would be nice if there was more competition but there are other players like AMD and a long tail of smaller ones

tw04 1y ago
>Consumer gpus are totally different products from the high end gpus now. Intel has failed on the gpu market and has effectively zero market share, so it is not actually clear there is an antitrust issue in that market. It would be nice if there was more competition but there are other players like AMD and a long tail of smaller ones I'm sorry that's just not correct. Intel is literally just getting started in the GPU market, and their last several releases have been nearly exactly what people are asking for. Saying "they've lost" when the newest cards have been on the market for less than a month is ridiculous. If they are even mediocre at marketing, the Arc Pro B50 has a chance to be an absolute game changer for devs who don't have a large budget: https://www.servethehome.com/intel-arc-pro-b50-review-a-16gb-sff-mini-gpu/ https://www.servethehome.com/intel-arc-pro-b50-review-a-16gb... I have absolutely no doubt Nvidia sees that list of "coming features" and will do everything they can to kill that roadmap.
- bigyabai 1y ago
  224 GB/s 128 bit The monkey's paw curls... I love GPU differentiation, but this is one of those areas where Nvidia is justified shipping less VRAM. With less VRAM, you can use fewer memory controllers to push higher speeds on the same memory! For instance, both the B50 and the RTX 2060 use GDDR6 memory. But the 2060 has a 192-bit memory bus, and enjoys ~336 GB/s bandwidth because of it.
  - privatelypublic 1y ago
    Tell me again, how fast can you move data from system ram to vram?
    - bigyabai 1y ago
      Over a PCIe5 x8, ~31.5gb/s.
- bpt3 1y ago
  Intel has been making GPUs for over 25 years. Claiming they are just getting started is absurd. To that point, they've been "just getting started" in practically every chip market other than x86/x64 CPUs for over 20 years now, and have failed miserably every time. If you think Nvidia is doing this because they're afraid of losing market share, you're way off base.
  - cptskippy 1y ago
    There's a very big difference between the MVP graphics chips they've included in CPUs and the Arc discrete GPU.
    - bpt3 1y ago
      Sure, but claiming they have literally just started is completely inaccurate. They've been making discrete GPUs on and off since the 80s, and this is at least their 3rd major attempt at it as a company, depending on how you define "major". They haven't even just started on this iteration, as the Arc line has been out since 2022. The main thing I learned from this submission is how much people hate Nvidia.
      - cptskippy 1y ago
        > The main thing I learned from this submission is how much people hate Nvidia. I think there's a lot of frustration with Nvidia as of late. Their monopoly was mostly won on the merits of their technology but now that they are a monopoly they have shifted focus from building the best technology to building the most lucrative technology. They've demonstrated that they no longer have interested in producing the best gaming GPUs because those might cannibalize their server technology. Instead they seem to focus on crypto and AI while shipping over priced knee capped cards at outrageous prices. People are upset because they fear this deal will somehow influence Intel's GPU ambitions. Unfortunately I'm not sure these folks want to buy Intel GPUs, they just want Nvidia to be scared into competing again so they can buy a good Nvidia card. People just need to draw a line in the sand and stop supporting Nvidia.
- raincole 1y ago
  "Intel getting started in GPU market" is like a chain smoker quitting smoking. It's so easy that they have done it 20 times!
- tapland 1y ago
  The lastest Arc GPUs were doing good, and were absolutely an option for entry/mid level gamers. I think lack of maturity was one of the main things keeping sales down.
  - Seattle3503 1y ago
    I've been seeing a lot of homelab types recommending their video cards for affordable Plex transcoding as well.
- Sohcahtoa82 1y ago
  I don't know what anybody would do with such a weak card. My RTX 5090 is about 10x faster (measured by FP32 TFLOPS) and I still don't find it to be fast enough. I can't imagine using something so slow for AI/ML. Only 2.2 tokens/sec on an 8B parameter Llama model? That's slower than someone typing. I get that it's a budget card, but budget cards are supposed to at least win on a pure price/performance ratio, even with a lower baseline performance. The 5090 is 10x faster but only 6-8x the price, depending on where in the $2-3,000 price range you can find one at.
  - jpalawaga 1y ago
    you have outlier needs if an rtx, the fastest consumer grade card, is not good enough for you. the intel card is great for 1080p gaming. especially if you're just playing counterstrike, indie games, etc, you don't need a beast. very few people are trying to play 4k tombraider on ultra with high refresh rate.
    - Sohcahtoa82 1y ago
      FWIW, my slowness is because of quantizing. I've been using Mistral 7B, and I can get 45 tokens/sec, which is PLENTY fast, but to save VRAM so I can game while doing inference (I run an IRC bot that allows people to talk to Mistral), I quantize to 8 bits, which then brings my inference speed down to ~8 tokens/sec. For gaming, I absolutely love this card. I can play Cyberpunk 2077 with all the graphics settings set to the maximum and get 120+ fps. Though when playing a much more graphically intense game like that, I certainly need to kill the bot to free up the VRAM. But I can play something simpler like League of Legends and have inference happening while I play with zero impact on game performance. I also have 128 GB of system RAM. I've thought about loading the model in both 8-bit and 16-bit into system RAM and just swap which one is in VRAM based on if I'm playing a game so that if I'm not playing something, the bot runs significantly faster.
      - mysteria 1y ago
        Hold on, you're only getting 45 tokens/sec with Mistral 7B on a 5090 of all things? That gets ~240 tokens/sec with Llama 7B quantized to 4 bits on llama.cpp [1] and those models should be pretty similar architecturally. I don't know exactly how the scaling works here but considering how LLM inference is memory bandwidth limited you should go beyond 100 tokens/sec with the same model and a 8 bit quantization. 1. https://github.com/ggml-org/llama.cpp/discussions/15013 https://github.com/ggml-org/llama.cpp/discussions/15013
realityking 1y ago
> it is not actually clear there is an antitrust issue in that market Preempting a (potential) future competitor from entering a market is also an antitrust issue.
Dylan16807 1y ago
Other than the market segmentation over RAM amounts, I don't see very much difference. There's some but there's been some for a long time. Isn't AMD re-unifying their architectures?
- 0x457 1y ago
  > There's some but there's been some for a long time. Isn't AMD re-unifying their architectures? Yes. > Other than the market segmentation over RAM amounts, I don't see very much difference. The difference between CDNA and RDNA is pretty much how fast it can crunch FP64 and SR-IOV. Prior to RDNA, AMD GPUs were jacks of all trades with compute bias. Which made them bad for gaming unless the game is specifically written around async compute. Vega64 has more FP64 compute than the 4080 for context. I think if AMD was able to get a solid market share of datacenter GPUs, they wouldn't have unified. This feels like CDNA team couldn't justify its existence.

tw04 1y ago

>Consumer gpus are totally different products from the high end gpus now. Intel has failed on the gpu market and has effectively zero market share, so it is not actually clear there is an antitrust issue in that market. It would be nice if there was more competition but there are other players like AMD and a long tail of smaller ones I'm sorry that's just not correct. Intel is literally just getting started in the GPU market, and their last several releases have been nearly exactly what people are asking for. Saying "they've lost" when the newest cards have been on the market for less than a month is ridiculous. If they are even mediocre at marketing, the Arc Pro B50 has a chance to be an absolute game changer for devs who don't have a large budget: https://www.servethehome.com/intel-arc-pro-b50-review-a-16gb-sff-mini-gpu/ https://www.servethehome.com/intel-arc-pro-b50-review-a-16gb... I have absolutely no doubt Nvidia sees that list of "coming features" and will do everything they can to kill that roadmap.

bigyabai 1y ago
224 GB/s 128 bit The monkey's paw curls... I love GPU differentiation, but this is one of those areas where Nvidia is justified shipping less VRAM. With less VRAM, you can use fewer memory controllers to push higher speeds on the same memory! For instance, both the B50 and the RTX 2060 use GDDR6 memory. But the 2060 has a 192-bit memory bus, and enjoys ~336 GB/s bandwidth because of it.
- privatelypublic 1y ago
  Tell me again, how fast can you move data from system ram to vram?
  - bigyabai 1y ago
    Over a PCIe5 x8, ~31.5gb/s.
bpt3 1y ago
Intel has been making GPUs for over 25 years. Claiming they are just getting started is absurd. To that point, they've been "just getting started" in practically every chip market other than x86/x64 CPUs for over 20 years now, and have failed miserably every time. If you think Nvidia is doing this because they're afraid of losing market share, you're way off base.
- cptskippy 1y ago
  There's a very big difference between the MVP graphics chips they've included in CPUs and the Arc discrete GPU.
  - bpt3 1y ago
    Sure, but claiming they have literally just started is completely inaccurate. They've been making discrete GPUs on and off since the 80s, and this is at least their 3rd major attempt at it as a company, depending on how you define "major". They haven't even just started on this iteration, as the Arc line has been out since 2022. The main thing I learned from this submission is how much people hate Nvidia.
    - cptskippy 1y ago
      > The main thing I learned from this submission is how much people hate Nvidia. I think there's a lot of frustration with Nvidia as of late. Their monopoly was mostly won on the merits of their technology but now that they are a monopoly they have shifted focus from building the best technology to building the most lucrative technology. They've demonstrated that they no longer have interested in producing the best gaming GPUs because those might cannibalize their server technology. Instead they seem to focus on crypto and AI while shipping over priced knee capped cards at outrageous prices. People are upset because they fear this deal will somehow influence Intel's GPU ambitions. Unfortunately I'm not sure these folks want to buy Intel GPUs, they just want Nvidia to be scared into competing again so they can buy a good Nvidia card. People just need to draw a line in the sand and stop supporting Nvidia.
raincole 1y ago
"Intel getting started in GPU market" is like a chain smoker quitting smoking. It's so easy that they have done it 20 times!
tapland 1y ago
The lastest Arc GPUs were doing good, and were absolutely an option for entry/mid level gamers. I think lack of maturity was one of the main things keeping sales down.
- Seattle3503 1y ago
  I've been seeing a lot of homelab types recommending their video cards for affordable Plex transcoding as well.
Sohcahtoa82 1y ago
I don't know what anybody would do with such a weak card. My RTX 5090 is about 10x faster (measured by FP32 TFLOPS) and I still don't find it to be fast enough. I can't imagine using something so slow for AI/ML. Only 2.2 tokens/sec on an 8B parameter Llama model? That's slower than someone typing. I get that it's a budget card, but budget cards are supposed to at least win on a pure price/performance ratio, even with a lower baseline performance. The 5090 is 10x faster but only 6-8x the price, depending on where in the $2-3,000 price range you can find one at.
- jpalawaga 1y ago
  you have outlier needs if an rtx, the fastest consumer grade card, is not good enough for you. the intel card is great for 1080p gaming. especially if you're just playing counterstrike, indie games, etc, you don't need a beast. very few people are trying to play 4k tombraider on ultra with high refresh rate.
  - Sohcahtoa82 1y ago
    FWIW, my slowness is because of quantizing. I've been using Mistral 7B, and I can get 45 tokens/sec, which is PLENTY fast, but to save VRAM so I can game while doing inference (I run an IRC bot that allows people to talk to Mistral), I quantize to 8 bits, which then brings my inference speed down to ~8 tokens/sec. For gaming, I absolutely love this card. I can play Cyberpunk 2077 with all the graphics settings set to the maximum and get 120+ fps. Though when playing a much more graphically intense game like that, I certainly need to kill the bot to free up the VRAM. But I can play something simpler like League of Legends and have inference happening while I play with zero impact on game performance. I also have 128 GB of system RAM. I've thought about loading the model in both 8-bit and 16-bit into system RAM and just swap which one is in VRAM based on if I'm playing a game so that if I'm not playing something, the bot runs significantly faster.
    - mysteria 1y ago
      Hold on, you're only getting 45 tokens/sec with Mistral 7B on a 5090 of all things? That gets ~240 tokens/sec with Llama 7B quantized to 4 bits on llama.cpp [1] and those models should be pretty similar architecturally. I don't know exactly how the scaling works here but considering how LLM inference is memory bandwidth limited you should go beyond 100 tokens/sec with the same model and a 8 bit quantization. 1. https://github.com/ggml-org/llama.cpp/discussions/15013 https://github.com/ggml-org/llama.cpp/discussions/15013

realityking 1y ago

> it is not actually clear there is an antitrust issue in that market Preempting a (potential) future competitor from entering a market is also an antitrust issue.

Dylan16807 1y ago

Other than the market segmentation over RAM amounts, I don't see very much difference. There's some but there's been some for a long time. Isn't AMD re-unifying their architectures?

0x457 1y ago
> There's some but there's been some for a long time. Isn't AMD re-unifying their architectures? Yes. > Other than the market segmentation over RAM amounts, I don't see very much difference. The difference between CDNA and RDNA is pretty much how fast it can crunch FP64 and SR-IOV. Prior to RDNA, AMD GPUs were jacks of all trades with compute bias. Which made them bad for gaming unless the game is specifically written around async compute. Vega64 has more FP64 compute than the 4080 for context. I think if AMD was able to get a solid market share of datacenter GPUs, they wouldn't have unified. This feels like CDNA team couldn't justify its existence.