Power pins on a MI50
- Thread starter fp64
- Start date
-
- Tags
- power pins
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
if all manufacturers assign the same function/voltage to each pin shape, then one can determine with confidence what type of power plug is used on a gpu. the question is, do all use the same standard?
i am using gpus for linear algebra computations. virtualization is of no interest. double precision compute speed is.
i am using gpus for linear algebra computations. virtualization is of no interest. double precision compute speed is.
They all use standards pin-outs - There are 2 standards PCIe and EPS-12V connectors.
(The standard specifies how much voltage, amp, wattage can be passed through each pin on each standard - No1 is violating that, it would be big mess of dead gpu's if that was the case.)
The first 2 pins can be used to identify which one it is.
If first one is hot, and 2nd is square-ish then its EPS-12V.
If first one is ground, and 2nd has same shape as first, is a sense cable then it is PCIe.
As mentioned NV Tesla series cards use single 8pin EPS-12V.
MI60 uses 2x pcie 8pin's, (so MI50 100% uses same connectors - they are same gpu just cut down a bit.)
You won't find any GPU not using those
(except newest from NV that do 12pin but you won't find it anywhere else)
(and some cards in past could use molex)
// Note:
In terms of power typically 8pin EPS-12V = 2x 8pin PCIe power connectors. (and vice versa)
(The standard specifies how much voltage, amp, wattage can be passed through each pin on each standard - No1 is violating that, it would be big mess of dead gpu's if that was the case.)
The first 2 pins can be used to identify which one it is.
If first one is hot, and 2nd is square-ish then its EPS-12V.
If first one is ground, and 2nd has same shape as first, is a sense cable then it is PCIe.
As mentioned NV Tesla series cards use single 8pin EPS-12V.
MI60 uses 2x pcie 8pin's, (so MI50 100% uses same connectors - they are same gpu just cut down a bit.)
You won't find any GPU not using those
(except newest from NV that do 12pin but you won't find it anywhere else)
(and some cards in past could use molex)
// Note:
In terms of power typically 8pin EPS-12V = 2x 8pin PCIe power connectors. (and vice versa)
Last edited:
Thank you.
found pin close-up for mi60 (scroll down a bit)
i checked it against my radeon vii and are identical i.e. pcie. differ from your earlier pcie sketch tho;
the two leftmost pins in the top row are square.
found pin close-up for mi60 (scroll down a bit)
AMD Radeon Instinct MI60 GPU Accelerator - 32GB HBM2-102D1631200 | eBay
Find many great new & used options and get the best deals for AMD Radeon Instinct MI60 GPU Accelerator - 32GB HBM2-102D1631200 at the best online prices at eBay! Free shipping for many products!
www.ebay.com
i checked it against my radeon vii and are identical i.e. pcie. differ from your earlier pcie sketch tho;
the two leftmost pins in the top row are square.
Just curious is the mi25 the same PCIe 8 pin like the Mi50 and Mi60? I would assume it would be, but just want to confirm it's not EPS-12v.
Correct, not EPS 12
AMD uses PCIE 12v up through the MI100, they switch to EPS with the MI210.
@fp64 What price were they offering, the mi50s are getting long in the tooth, they lack the Matrix Math the MI100/200 have.
MI50: 250 euros each + 20 euro shipping from germany. bought one. dead loss. it is slightly slower than my surviving Radeon vii. 6.6 tflops fp64 are unattainable due to bandwidth limitation. similar story with MI60, bought one from the US supposed new. it is slightly faster than the said radeon vii - 7.33tflops ffts and linear algebra unattainable as well. add to these my two dead radeon vii which i bought from new, and i am tired of being lied to by amd.
btw, MI50 2x8 pins pcie; MI60 8+6 pcie pins; radeon vii 2x8 pcie pins.
btw, MI50 2x8 pins pcie; MI60 8+6 pcie pins; radeon vii 2x8 pcie pins.
How are you cooling them, they throttle pretty quick.
What workload are you running? You may have to tune your code to get full performance, or might be hitting other system limits.
Cards mounted on a HP ML350 G9 - all (8) fans operational. during tests the temp never exceeds 54C (monitored thru rocm-smi).
since radeon vii is intentionally crippled to half the fp64 rate, there should a doubling of the speed when this limitation is removed unless computations are bandwidth-limited. i am running batched matrix-matrix multiplications of my own writing (tiled algorithm) and some simple matrix additions. i have moved to hip now - originally they were in opencl.
i am also running batched ffts (r2c and c2r) using rocfft which is supposed to be expertly written. the same problem here.
since radeon vii is intentionally crippled to half the fp64 rate, there should a doubling of the speed when this limitation is removed unless computations are bandwidth-limited. i am running batched matrix-matrix multiplications of my own writing (tiled algorithm) and some simple matrix additions. i have moved to hip now - originally they were in opencl.
i am also running batched ffts (r2c and c2r) using rocfft which is supposed to be expertly written. the same problem here.
Hmm. i thought the MI100 was dual 8pin EPS also.
I've was running an MI25 as a VM card for a while, and moved to an MI60 to dabble with some AI models.
For cooling I found some 3D printed 25/80/120mm fan mounts which work.. okay. I'd still rather be able to slap a waterblock on it though.
I've was running an MI25 as a VM card for a while, and moved to an MI60 to dabble with some AI models.
For cooling I found some 3D printed 25/80/120mm fan mounts which work.. okay. I'd still rather be able to slap a waterblock on it though.