Jump to content

I currently have a Radeon 6650xt which I have enjoyed quite a bit but I've noticed that it's a real pain to use for machine learning, at least relative to Nvidia's cuda. I want to upgrade and have been looking at the RTX 4070 but.... I find it to be a bit of a hard buy especially since AMD cards tend to offer more vram and more performance in games (I only care about rastorization). I've heard that higher end AMD cards have access to ROCm so... My question is, do any of you know which AMD cards support ROCm? If there is a relatively priced AMD card to the 4070 (around $550), is ROCm as good as cuda and if not, would the performance and price difference make up the difference?

 

Thanks is advanced!

Link to comment
https://linustechtips.com/topic/1541642-cuda-vs-rocm/
Share on other sites

Link to post
Share on other sites

ROCm is supported on Radeon RX 400 and newer AMD GPUs.

 

I’ve never personally tried to use it although I did investigate using it awhile back. From looking around, it appears that not much has changed. It’s main problem was that it wasn’t not supported by the same wide range of packages and applications as CUDA.

 

As for its performance, no idea. But if I were a betting man, I’d guess that its performance is weaker and part of the reason the re is less support for it. (Tbf to AMD, It could also be that AMD have a far smaller market share)

I might be experienced, but I'm human and I do make mistakes. Trust but Verify! I edit my messages after sending them alot, please refresh before posting your reply. Please try to be clear and specific, you'll get a better answer. Please remember to mark solutions once you have the information you need. Expand this signature for common PC building advice, a short bio and a list of my components.

 

Common build advice:

1) Buy the cheapest (well reviewed) motherboard that has the features you need. Paying more typically only gets you features you won’t use. 2) only get as much RAM as you need, getting more won’t (typically) make your PC faster. 3) While I recommend getting an NVMe drive, you don’t need to splurge for an expensive drive with DRam cache, DRamless drives are fine for gamers. 4) paying for looks is fine, just don’t break the bank. 5) Tower coolers are usually good enough, unless you go top tier Intel or plan on OCing. 6) OCing is a dead meme, you probably shouldn’t bother. 7) "Bottlenecks" rarely matter and "Future-proofing" is a myth. 8) AIOs don't noticeably improve performance past 240mm and don't improve at all past 360mm. 9) RTFM.

 

Useful Websites:

https://www.productchart.com - helps compare monitors, https://uk.pcpartpicker.com - makes designing a PC easier.

 

Bio:

He/Him - I'm a PhD student working in the fields of reinforcement learning and traffic control. PCs are one of my hobbies and I've built many PCs and performed upgrades on a few laptops (for myself, friends and family). My personal computers include 4 windows (10/11) machines and a TrueNAS server (and I'm looking to move to dual booting Linux Mint on my main machine in future). Aside from computers, I also dabble in modding/homebrew retro consoles, support Southampton FC, and enjoy Scuba Diving and Skiing.

Fun Facts

1) When I was 3 years old my favourite toy was a scientific calculator. 2) My father is a British Champion ploughman in the Vintage Hydraulic Class. 3) On Speedrun.com, I'm the world record holder for the Dream Bobsleigh event on Mario & Sonic at the Olympic Winter Games 2010.

 

My Favourite Games: World of Tanks, Runescape, Subnautica, Metroid (Fusion and Dread), Spyro: Year of the Dragon (Original and Reignited Trilogy), Crash Bash, Mario Kart Wii, Balatro

 

My Computers: Primary: My main gaming rig - https://uk.pcpartpicker.com/user/will0hlep/saved/NByp3C Second: Hosts Discord bots as well as a Minecraft and Ark server, and also serves as a reinforcement learning sand box - https://uk.pcpartpicker.com/user/will0hlep/saved/cc9K7P NAS: TrueNAS Scale NAS hosting SMB shares, DDNS updater, pi-hole, and a Jellyfin server - https://uk.pcpartpicker.com/user/will0hlep/saved/m37w3C Foldatron: My folding@home and BOINC rig (partially donated to me by Folding Team Leader GOTSpectrum) - Mobile: Mini-ITX gaming rig for when I'm away from home -

Link to comment
https://linustechtips.com/topic/1541642-cuda-vs-rocm/#findComment-16211431
Share on other sites

Link to post
Share on other sites

49 minutes ago, eric_s said:

My question is, do any of you know which AMD cards support ROCm?

Take a look at: https://rocm.docs.amd.com/en/latest/release/windows_support.html#windows-supported-gpus

 

Make sure to switch to the "Radeon" tab, by default it'll only show supported Radeon Pro cards. The 6650 XT seems to support the runtime, but not the HIP SDK. You need at least an RX 6800 for that.

 

Unless you know that the software you use also supports ROCm, it's not going to matter much though. A lot of software requires CUDA.

Remember to either quote or @mention others, so they are notified of your reply

Link to comment
https://linustechtips.com/topic/1541642-cuda-vs-rocm/#findComment-16211443
Share on other sites

Link to post
Share on other sites

21 minutes ago, will0hlep said:

ROCm is supported on Radeon RX 400 and newer AMD GPUs.

 

I’ve never personally tried to use it although I did investigate using it awhile back. From looking around, it appears that not much has changed. It’s main problem was that it wasn’t not supported by the same wide range of packages and applications as CUDA.

 

As for its performance, no idea. But if I were a betting man, I’d guess that its performance is weaker and part of the reason the re is less support for it. (Tbf to AMD, It could also be that AMD have a far smaller market share)

Wait what?

RocM support on consumer class cards only launched this year, I know one guy was pushing for it hard and was able to do some of it, but he wasn't AMD
https://rocm.docs.amd.com/en/latest/release/windows_support.html

 

Vega 7 was supported(cause it was like a titan to AMD), but the support is supposed to drop next year. 

Link to comment
https://linustechtips.com/topic/1541642-cuda-vs-rocm/#findComment-16211445
Share on other sites

Link to post
Share on other sites

2 hours ago, starsmine said:

RocM support on consumer class cards only launched this year, I know one guy was pushing for it hard and was able to do some of it, but he wasn't AMD
https://rocm.docs.amd.com/en/latest/release/windows_support.html

https://en.wikipedia.org/wiki/ROCm

 

Your right, official support is more limited. But unoffical support is available on cards going back to the Radeon RX 400. No idea what it takes to get them running though (and it probably isn't on windows).

I might be experienced, but I'm human and I do make mistakes. Trust but Verify! I edit my messages after sending them alot, please refresh before posting your reply. Please try to be clear and specific, you'll get a better answer. Please remember to mark solutions once you have the information you need. Expand this signature for common PC building advice, a short bio and a list of my components.

 

Common build advice:

1) Buy the cheapest (well reviewed) motherboard that has the features you need. Paying more typically only gets you features you won’t use. 2) only get as much RAM as you need, getting more won’t (typically) make your PC faster. 3) While I recommend getting an NVMe drive, you don’t need to splurge for an expensive drive with DRam cache, DRamless drives are fine for gamers. 4) paying for looks is fine, just don’t break the bank. 5) Tower coolers are usually good enough, unless you go top tier Intel or plan on OCing. 6) OCing is a dead meme, you probably shouldn’t bother. 7) "Bottlenecks" rarely matter and "Future-proofing" is a myth. 8) AIOs don't noticeably improve performance past 240mm and don't improve at all past 360mm. 9) RTFM.

 

Useful Websites:

https://www.productchart.com - helps compare monitors, https://uk.pcpartpicker.com - makes designing a PC easier.

 

Bio:

He/Him - I'm a PhD student working in the fields of reinforcement learning and traffic control. PCs are one of my hobbies and I've built many PCs and performed upgrades on a few laptops (for myself, friends and family). My personal computers include 4 windows (10/11) machines and a TrueNAS server (and I'm looking to move to dual booting Linux Mint on my main machine in future). Aside from computers, I also dabble in modding/homebrew retro consoles, support Southampton FC, and enjoy Scuba Diving and Skiing.

Fun Facts

1) When I was 3 years old my favourite toy was a scientific calculator. 2) My father is a British Champion ploughman in the Vintage Hydraulic Class. 3) On Speedrun.com, I'm the world record holder for the Dream Bobsleigh event on Mario & Sonic at the Olympic Winter Games 2010.

 

My Favourite Games: World of Tanks, Runescape, Subnautica, Metroid (Fusion and Dread), Spyro: Year of the Dragon (Original and Reignited Trilogy), Crash Bash, Mario Kart Wii, Balatro

 

My Computers: Primary: My main gaming rig - https://uk.pcpartpicker.com/user/will0hlep/saved/NByp3C Second: Hosts Discord bots as well as a Minecraft and Ark server, and also serves as a reinforcement learning sand box - https://uk.pcpartpicker.com/user/will0hlep/saved/cc9K7P NAS: TrueNAS Scale NAS hosting SMB shares, DDNS updater, pi-hole, and a Jellyfin server - https://uk.pcpartpicker.com/user/will0hlep/saved/m37w3C Foldatron: My folding@home and BOINC rig (partially donated to me by Folding Team Leader GOTSpectrum) - Mobile: Mini-ITX gaming rig for when I'm away from home -

Link to comment
https://linustechtips.com/topic/1541642-cuda-vs-rocm/#findComment-16211568
Share on other sites

Link to post
Share on other sites

9 hours ago, eric_s said:

is ROCm as good as cuda

It's not, far from it.

9 hours ago, eric_s said:

and if not, would the performance and price difference make up the difference?

IMO, no.

 

You can get it to work with most AMD GPUs after 6000 series, but it's still going to be a pain no matter if you go for a 6500xt or 7900xtx, and performance will also be subpar. If your focus is ML, steer from AMD at this moment, maybe look back at them in a couple years.

FX6300 @ 4.2GHz | Gigabyte GA-78LMT-USB3 R2 | Hyper 212x | 3x 8GB + 1x 4GB @ 1600MHz | Gigabyte 2060 Super | Corsair CX650M | LG 43UK6520PSA
ASUS X550LN | i5 4210u | 12GB
Lenovo N23 Yoga

Link to comment
https://linustechtips.com/topic/1541642-cuda-vs-rocm/#findComment-16212152
Share on other sites

Link to post
Share on other sites

So if I'm relatively new to AI and ML and don't really want to tinker much to get it going, then I should probably choose an Nvidia GPU? With that, I'm seeing a lot of people saying to wait until the 4th of January and to ignore any black Friday deals since Nvidia is about to launch a new revision. Do you agree?

Link to comment
https://linustechtips.com/topic/1541642-cuda-vs-rocm/#findComment-16212904
Share on other sites

Link to post
Share on other sites

14 minutes ago, eric_s said:

So if I'm relatively new to AI and ML and don't really want to tinker much to get it going, then I should probably choose an Nvidia GPU? With that, I'm seeing a lot of people saying to wait until the 4th of January and to ignore any black Friday deals since Nvidia is about to launch a new revision. Do you agree?

If you're getting started you should just be using colab to be honest, until you can properly understand what are your requirements. 

Other than that, yeah, wait for nvidia's new products so you can see if a price drop on current products happen. 

FX6300 @ 4.2GHz | Gigabyte GA-78LMT-USB3 R2 | Hyper 212x | 3x 8GB + 1x 4GB @ 1600MHz | Gigabyte 2060 Super | Corsair CX650M | LG 43UK6520PSA
ASUS X550LN | i5 4210u | 12GB
Lenovo N23 Yoga

Link to comment
https://linustechtips.com/topic/1541642-cuda-vs-rocm/#findComment-16212914
Share on other sites

Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

×