AMD RADEON PRO V620 32GB GDDR6 GPUs - $565

[CyklonDX]

maybe a good news or final nails into the topic - but appears amd changed their approach again - and started updating gim on github (6-7h ago) currently supporting mi200(rev0-3), mi300, mi325x, mi308x, mi300x_hc, and v710(navi32). *even tho they aren't listed, so there's hope for streamlined gim again.

GitHub - amd/MxGPU-Virtualization at staging/8.2.0.K

Contribute to amd/MxGPU-Virtualization development by creating an account on GitHub.

github.com

The staging branch for 8.2.0K tag outlines

AMDGV_CMD_CHIP_MI200 = 2,
AMDGV_CMD_CHIP_NAVI32 = 8,
AMDGV_CMD_CHIP_MI300X = 9,
AMDGV_CMD_CHIP_MI308X = 11,
AMDGV_CMD_CHIP_UNKNOWN,
AMDGV_CMD_CHIP_LAST,

potentially v620 is under older common/legacy devices.


i've added an issue to github, i would appreciate if you guys could also spam there - if there's a lot of activity maybe they'll do us something nice.

KVM Support for Navi21 (v620) · Issue #6 · amd/MxGPU-Virtualization

Please add support for v620 and bring new life into this e-waste. Thank You! Regards

github.com

 

[CyklonDX]

Here's official response from amd - more than at any time anyone got so far.

 

[tjk]

And someone after AMD post linked this:

AMD RDNA2 system optimization — ROCm Documentation

Learn about system settings and performance tuning for RDNA2-based GPUs.

rocm.docs.amd.com

So, is it supported or not?

 

[CyklonDX]

And someone after AMD post linked this:

AMD RDNA2 system optimization — ROCm Documentation

Learn about system settings and performance tuning for RDNA2-based GPUs.

rocm.docs.amd.com

So, is it supported or not?

The amd response should be explanatory.

(This url is a left-over from another time.)

 

[tjk]

Thanks, but they mention the V620 in the latest ROCm release - not doubting you just trying to understand why they call out the V620 in the current release - if it isn't supported?

Educate me pls.

 

[CyklonDX]

v320/340 which are vega which supposedly do support sr-iov; [No drivers]
v420 vega2 unreleased which supposedly do support sr-iov; [No drivers, and only 1-2 cards on the market ever spotted]
v520/540 rdna1 which supposedly do support sr-iov [No sr-iov drivers anymore they used to be on amd site for brief while - so no drivers]
v620 rdna2 which supports sr-iov but drivers behind partnership program for now. [No Drivers]
v720 rdna3 navi32 rdna3 which supports sr-iov and drivers are available in the github code - staging repo(beta) [No official driver/s, but staging branch does have them listed in code]

As per amd official response, they might release all rdna(1-4) sr-iov open drivers publicly at some point - which will be for v620, and v520/540 gpu's.

*1 i don't expect v520/540 since its not listed here at all anymore.
*2 kvm drivers we all want for v620 do exist, but are not being shared publicly - they likely are in some kind of agreement with microsoft blocking them from sharing it unless its with their partners.
*3 the link you posted, is a left-over from when amd rep was guiding me to use rocm to get the driver *(i couldn't test it then since i didn't have the card), that isn't there anymore, and i don't think rocm can download any driver packages for you anymore anyway.

The support virtualization matrix in rocm states v620 is supported in hyper-v on Azure Host 2024 *and that's only available through/for microsoft. *This is not ms server 2024 whatever edition.

 

[custom90gt]

Part of me is surprised these paywall drivers were never leaked anywhere. Without knowing that these will be released in a reasonable time frame, I think these are probably priced too high.

 

[jbrukardt]

yeah, i think ill wait for the Arc B60 cards to hit consumer markets

 

[pcmoore]

Aside from the published 300W TDP, does anyone have any information on idle power consumption as well as any info on observed consumption when running various LLM tasks? Also, I'm reading that AMD doesn't appear to support power limits on modern Linux systems, but people have had luck limiting power by indirectly limiting the clock rate; anyone have any experience with that?

 

[Cruzader]

Part of me is surprised these paywall drivers were never leaked anywhere. Without knowing that these will be released in a reasonable time frame, I think these are probably priced too high.

Atleast for the vendor id be able to get them from the downloads are uniqe/fingerprinted.

That pack tracing back is a solid incentive not to leak it public.

 

[juddle14]

Aside from the published 300W TDP, does anyone have any information on idle power consumption as well as any info on observed consumption when running various LLM tasks? Also, I'm reading that AMD doesn't appear to support power limits on modern Linux systems, but people have had luck limiting power by indirectly limiting the clock rate; anyone have any experience with that?

We did some *very* basic LLM testing with these (using publicly available drivers - ROCm driver 6.0 with Ubuntu and Ollama).

* Rdeepseek-r1:70b using 4x V620s - 7 tokens / Sec
* mistal7b using 1x V620 - 54 tokens / Sec

That is what we were seeing with just a few tests and no finetuning
Idle power draw was 6 watts (which is 1/3 draw of 3090s)

We primarily sell hardware, so I'm sure others with more LLM experience can get more out of these cards.

 

[pcmoore]

@juddle14 just sent you a message

 

[pcmoore]

Just a FYI on the V620 for those of you who are considering putting these into a GPU server with dedicated GPU support brackets: the support brackets might block one of the two power connectors on the V620, and the V620 doesn't appear to come online if only one power connector is attached. Of course that doesn't mean you can't use these cards, just be prepared to jimmy rig some supports if needed.

 

[Gunnyp]

It's the KVM driver I am talking about.

You have a machine running Ubuntu, you can get a driver so that your Ubuntu machine recognizes the GPU. That's great.

What I want is to install KVM (Kernel-based Virtual Machine) to turn the Ubuntu machine into a hypervisor host. That way, I can run a bunch of virtual machines on that host. The driver that I will need to install would allow the host (Ubuntu with KVM installed in this case) to share the GPU with multiple virtual machines. That driver is the file: gim-dkms_1.0.0.1234577_all

The driver that allows your Ubuntu machine to share its GPU with multiple virtual machines is very different from the driver that allows your Ubuntu machine to use the GPU itself. The former driver is what people can't find.

Nvidia is doing the same thing. Its host driver (KVM driver) is locked behind a paywall while its guest drivers are publicly available.

We purchased 4 V620s at the end of 2023, before GPU prices went through the roof, intending to use their one notable feature of partitioning via SR-IOV to share over multiple virtual machines or allocate multiple 32Gb V620s memory to a single virtual machine. Unfortunately, as you've so kindly warned others, these attributes are not accessible as AMD does not yield access to the required drivers outside of Microsoft and their NGad Azure servers via Windows.

This article from late Dec 2024 would seemingly make clear that access to support for virtual machines is highly unlikely to ever occur in the MS Windows arena anytime soon as this partitioning is Microsoft's main selling point on their highly promoted NGAD cloud vs Amazon AWS.

Exploring Azure NGads V620 for GPU-Accelerated Graphics

Explore Azure NGads V620 series for GPU-accelerated apps like cloud gaming, 3D rendering, & virtual desktops. Compare pricing, configs, and AWS GPU options.

cloudoptimo.com

Thankfully, this became clear to us before the 30-day window for eBay returns and fortunately eBay honors their 30-day policy no matter what "no return" policy a seller seeks to add.

We purchased and returned the V620s for under $400 in the days before these tariff wars shot prices ever skywards. With less than half the throughput of a RTX 3090 and lacking crucial virtual machine support in the manner of Microsoft NGAD these cards are utterly neutered, but would be worth a look if given access to the required VM support software.

 

[jbrukardt]

oof, unfortunate that these are basically ewaste then

 

[iraqigeek]

Just to add a data point: there seems to be a gagillion Radeon Instinct Mi50 being sold out of China. They come with 32GB of HBM2 with 1TB/s bandwidth. The Mi50 has about 2/3rds the compute of the V620 and there are no Windows drivers for the Mi50, but they're selling for ~150-160 a pop including shipping if you order a few. You can get three Mi50 cards for the price of one V620. Four Mi50 cards would cost a tad more than a single V620.

 

[EasyRhino]

Just to add a data point: there seems to be a gagillion Radeon Instinct Mi50 being sold out of China. They come with 32GB of HBM2 with 1TB/s bandwidth. The Mi50 has about 2/3rds the compute of the V620 and there are no Windows drivers for the Mi50, but they're selling for ~150-160 a pop including shipping if you order a few. You can get three Mi50 cards for the price of one V620. Four Mi50 cards would cost a tad more than a single V620.

sounds fun but are those based on vega architecture from 2018 right?

 

[jbrukardt]

sounds fun but are those based on vega architecture from 2018 right?

yeah, theyre vega 20, but theyre also HBM which the v620s arent. So older arch, meaning poorer software support, but way faster mem

mi50 has ridiculous FP64 performance, but thats unfortunately not that useful for many things besides initial training, where you need TBs worth of VRAM. You're really trading FP16 flops against memory BW.

Memory Type HBM2 Memory Bus 4096 bit Bandwidth 1.02 TB/s

P16 (half) 26.82 TFLOPS (2:1)
FP32 (float)13.41 TFLOPS
FP64 (double)6.705 TFLOPS (1:2)

V620

Memory Size 32 GB Memory Type GDDR6 Memory Bus 256 bit Bandwidth512.0 GB/s

FP16 (half) 40.55 TFLOPS (2:1)
FP32 (float)20.28 TFLOPS
FP64 (double)1,267.2 GFLOPS (1:16)

 

[iraqigeek]

So older arch, meaning poorer software support

They're marked as deprecated in ROCm 6 and will be unsupported in ROCm7. Got 6.4.1 working, and I'm fairly certain I can get 6.4.3 working without any additional hassle.
The issue with AMD, whether you're on their latest hardware or older architectures, is QA and documentation. Installing ROCm 6.3.x and 6.4.x is an exercise in figuring what steps are missing from the installation instructions in AMD's documentation.... Let's just say that ChatGPT was immensely helpful today.

sounds fun but are those based on vega architecture from 2018 right?

Yes, in absolute terms they're not that powerful anymore. Their FP32 performance is a bit higher than the Tesla P40, but whereas the P40 falls apart in FP16, the Mi50 has 2:1 FP16 making it much faster than the P40. But for 150-160 a pop for 32GB HBM VRAM they're a steal, IMO.

On a Broadwell Xeon (LGA2011) or Skylake-SP (LGA2066 or LGA3647), you can have a full rig built with four Mi50s for ~1k. I just received five of them today, and I'm very tempted to get 6-7 more.

 

[larkinwc]

v320/340 which are vega which supposedly do support sr-iov; [No drivers]
v420 vega2 unreleased which supposedly do support sr-iov; [No drivers, and only 1-2 cards on the market ever spotted]
v520/540 rdna1 which supposedly do support sr-iov [No sr-iov drivers anymore they used to be on amd site for brief while - so no drivers]
v620 rdna2 which supports sr-iov but drivers behind partnership program for now. [No Drivers]
v720 rdna3 navi32 rdna3 which supports sr-iov and drivers are available in the github code - staging repo(beta) [No official driver/s, but staging branch does have them listed in code]

As per amd official response, they might release all rdna(1-4) sr-iov open drivers publicly at some point - which will be for v620, and v520/540 gpu's.

*1 i don't expect v520/540 since its not listed here at all anymore.
*2 kvm drivers we all want for v620 do exist, but are not being shared publicly - they likely are in some kind of agreement with microsoft blocking them from sharing it unless its with their partners.
*3 the link you posted, is a left-over from when amd rep was guiding me to use rocm to get the driver *(i couldn't test it then since i didn't have the card), that isn't there anymore, and i don't think rocm can download any driver packages for you anymore anyway.

The support virtualization matrix in rocm states v620 is supported in hyper-v on Azure Host 2024 *and that's only available through/for microsoft. *This is not ms server 2024 whatever edition.
View attachment 43885

v340 drivers are available from OEMs such as lenovo, now indexed by google last time I checked. Just only applicable on ESXi 6.5/6.7.

EDIT:

Looks like AMD has made v340 host drivers available (again?) from their support/drivers page, I don't believe it was in the past when I had looked.