cancel
Showing results for 
Search instead for 
Did you mean: 

PC Graphics

brandonm
Adept I

RX590 Wrong Memory Clock in GPUZ, AMD Radeon Software

Radeon Software and GPU-Z and Radeon Bios Editor report GDDR5 2000MHz on my RX590. However, speed is 256GB/s. According to datasheet of memory chip K4G80325FB-HC25 which is on my RX560, the memory timing is 0.25ns. 1/0.25ns = 4000MHz. At DDR, 2bit/clk * 4000MHz = 8GT/s. The datasheet specifies DDR at 8gbit/s per pin. (The Samsung chip/datasheet for my RX560 is more convenient to mention because it has that very explicit 0.25ns number. My RX590 uses Micron memory.) I downloaded several BIOS files from TechPowerUp BIOS Collection, they also report 2000MHz. Apparently all the tools/software are wrong? (QDR is a thing, but I have only seen DDR chips on my GPUs.)

Unrelated, but it would be cool if AMD published "datasheets" for their GPU chips. Maybe have a "Recommended Core Voltage vs Core Frequency" plot from like 1100MHz to 1600MHz or something like that. Would help quickly achieve a desired power rating when under-/over- volting/clocking.

P.S. I read online that BIOS mod is no longer possible on RX5700. I only use Windows to tune GPU, then RBE to flash, then move to Linux 100% (uninstall Windows, it's useless spyware other than some programs with lazy developers -- unless they got stuck with MS managed C++ or worse). Would it be possible to put unsigned tuning region in VBIOS? I know application better than GPU factory, especially considering case, airflow, performance needs... cooling solution is small part of the equation.

P.P.S.S. Any possibility of making a MEMTEST86 for GPU? Maybe using Linux kernel driver as a base? Or adding 9th chip for ECC? I really, REALLY hate fretting over "dead pixels" in memory chips........... my RX560 has some bad VRAM for sure, but has only crashed once in a few months, not too bad. Usually only see a few artifacts here and there.

0 Likes
10 Replies
brandonm
Adept I

Ignore core clock, I was running at 920mV to test efficient performance.

0 Likes
brandonm
Adept I

I am aware of GDDR "EDC" bits, but this is not ECC, and apparently it is not effective/used on my RX560.

0 Likes
brandonm
Adept I

Just remembered, my experiences with bad VRAM in the past (maybe I'm unlucky) have scared me away from HBM cards, especially used ones. Does Vega HBM have data lines for ECC codes?

0 Likes
brandonm
Adept I

I have witnessed brief "flashes" of a small, say 250-500px horizontal, 5 plus or minus 5 "line" using both my RX590 and RX560. I do not believe the RX590 has bad memory, though I have relatively recently acquired it and not exercised it very much. Plus I am skeptical both happen to have same defect in general, plus they are using Samsung vs Micron memory which we know is a little different internally

Is this a driver bug? Running 2560x1440 on Linux 5.3 with Mesa 19.2 (and 17.1 or whatever was in Kubuntu 19.10). I am using the same cables, monitors, and installation (did not move things other than needed to plug in dual-link DVI cable) as I have used with Radeon 5670 and GeForce 750 Ti.

0 Likes
brandonm
Adept I

5 line vertical, sorry for the omission. And Kubuntu 20.04/Kernel 5.4/Mesa 19.2 and Kubuntu 19.10/Kernel 5.3/Mesa 17 or whatever.

0 Likes
brandonm
Adept I

I just checked this verified AMD-brand AMD GPU using Samsung DDR and checked the datasheet to confirm this series uses 0.25ns timing:

VGA Bios Collection: AMD RX 570 4 GB | TechPowerUp 

0 Likes
brandonm
Adept I

I'm guessing locked BIOS is to prevent flashing RX5600 into RX5700 despite ability to overclock in software? It might make more sense to limit clocks like Intel mainstream CPUs. You could even do that on GPU memory controller, that way RX5600 could have used 8GB 256bit at hardware-locked clocks for full interleaving without that awkward 6GB vs 12GB divide. (RX 5500 seems to commonly sell with 8GB, so 5500/5600/5700 goes 8GB/6GB/8GB.)

0 Likes

Or even only limit clocks on memory controller.

0 Likes
brandonm
Adept I

Re: my ECC comment: never mind 9th chip, I assume you have 8 36/32bit SED channels (with SED disabled apparently per my RX560). But what if you buffer the first set of bits until the second set is sent/received to add a cycle of latency to create a virtual 72/64 channel for desktop-style SECDED?

I "fixed" my RX560 by underclocking RAM to 6700MT/s with same timings.

0 Likes

Also my family owns the Xbox One, this appears to utilize AMD APU with no memory protection. It does crash occasionally. I understand the Xbox One does not utilize GDDR so cannot use my buffer idea to convert 36/32 chips into 72/64 virtual chips, but this would benefit from the 9x8bit chips unbuffered ECC idea. (If new Xbox uses GDDR, what about using this buffer idea for "free" ECC?)

0 Likes