LabRat

I suppose I'm gonna have to dual-box the machine it's in @ my desk... It wouldn't be fair to pass it thru to a Win9X/XP VM via PCI-PCIe bridge in my Daily Driver R5 5600 build. I literally forgot that I'd already built the thing; I have a different case for it, but this one appears to 'work'.

Dangit, @pioneerisloud! I don't even know how I got here... oh wait... yes, I do.

All this does is move where you plug in the power, and remove (plug-n-play) backwards compatibility. Even as a standard, this would effectively separate "HEDT" and "Desktop" GPUs into entirely separate branches. Where, a 'HEDT card' isn't compatible with any 'value-oriented' boards. That sounds like a wetdream for AMD/Intel/nVidia and their partners but, a nightmare for Small-SIs and DIYers. No Thank You. OTOH, I'd be more than on board for -12V_N_+12V "Split Rail" 24VDC becoming 'a thing' for PSUs and GPUs. https://kbreee.blogspot.com/2013/11/dc-transmission-and-distribution.html Edison 3-Wire Power Distribution - https://pubmed.ncbi.nlm.nih.gov/17757744/ https://www.eeeguide.com/comparison-of-conductor-material-in-overhead-system/ Basically, (If you've done any home electrical work in US/CAN) just think of the GPU as an Electric Range, Dryer, or Water Heater. Split-phase electric power - Wikipedia EN.WIKIPEDIA.ORG

No. Been there Done that This is just Proprietary PCIe Pro

As you've found out w/ Vega's predecessor (Fiji), they're compute monsters. Which, is also why we only see HBM on 'CDNA' products. (-and are direct-descendants of Fiji - Vega10/20 - Navi12) For rapid processing of memory transactions, HBM is fantastic. AI/MI and 'compute' seem to benefit much more than 3D Game Engines. As @Lifted points out, for strictly (max price/perf) gaming, Vega(and Fiji) isn't(aren't) really a 'value'.

I haven't really tried much (due to aforementioned cooling limits), but I HAD gotten even more than +50% working by loading VegaFELC PPT. Haven't tried it again since Zone drivers but, I have seen more than 300W out of my MI25; it will overheat in seconds (under load*) *:

Yes. The 'trick' is using either OverdriveNtool (or Wattman) to 'emulate' the MI25's stock P-states up until the last few. Then, use wattman for the rest of your tweaking. (Edit/note: P-State mods are often ignored, especially 'oddball' clock and mV settings *until* you load the Zone drivers) Not sure how to easily prove by screen shot, but (I'll try) @ my 'gimped' 1400/1420 / 1100HBM, I can exceed the 'stock' WX 9100 limit w/ +50% power set in wattman. I've pushed 275w+ before, but hotspot and HBM temps kill my stability, even w/ maxfan (my ghettomod'd server blower cooling is attached to the card's fan header)

What forces dare defend this world from the steel talons of the Jade Falcon?! Sorry... memetic trigger...

Started with Navi 12 / CDNA; been going that way since. IMO, AI/MI has drastically changed the prosumer/gamer/enthusiast market (relative) profits. It's just not 'attractive' to make 'halo tier' GPUs and sell them for $1500 or less. Example: My (Vega 10) MI25 used to be a $6k+ card, (x2 Navi 12) V540s were multiple-$10Ks, IIRC.

I wonder if AMD would ever consider MI250(X)/MI300-like "Big APUs" for HEDT? Nonetheless, I'm pleased to see there's even plans for Zen4 HEDT. I'd assumed HEDT'd been (basically) abandoned.

A Melty 8500GT 256MB DDR2 Half-Height. ~$20 shipped I haven't tested it yet, but it was sold as working. I figure the cooler needs replacing/repair. (eventually) to go in a PCI-PCIe bridge card, in a VIA C7 ITX build w/ Mod Win9x Drivers. (8500GT vs. PCI FX5500 256MB 128-bit vs. PCI MX4000 128MB 128-bit)

ATI had, arguably twice (or more) 9600/9800Pro and 9600/9800XT wiped the floor of all GF4 and FX 5000 cards. x850 XT PE kicked butt, but lacked SM3.0 v. GF6800s X1800XT(X) v. 7800GTX 256MB, X1900XT(X) vs. 7800GTX 512MB, and X1950XTX vs. 7900 GTX were all tight competition, as well. (With the short-ran soon-superseded-by-G80 X1950XTX GDDR4 reigning undisputed king) AMD... I'd say the HD 7970 (Ghz Ed. / Rev B. silicon) qualifies. The R9 290X 8GB w/ Water Cooling or excessive Air Cooling and a heavy OC would topple the competition. But... that's not out-of-box/apples-apples.

Can I just wear my card like a breastplate? Find me some better (air) cooling for my card, and you got a duel. (@TM my MI25-convert is 'gimped' to 1400/1420 Core and 1100 HBM)

A Vega 64 is about the best one can commonly acquire for such uses, yes. Radeon Instinct MI25 cards are 16GB Vega GPUs of the same rev. as Vega 56/64, "Vega 10". Using MI25s as a 'gaming GPU' is anything but a plug-and-play experience, however. You're asking the wrong person about a 5700XT being faster than a Vega 10/20 HBM-equipped card. (Vega 20 is Radeon VII, and we shall not speak of this card, just the GPU die. ) In all seriousness: A 'top-tier' board partner and Vega 64 Liquid Cooled Frontier Edition is meeting-beating a 5700XT and a 'boring' Vega 64 is w/in spitting distance of a 5700XT. With undervolt, overclock, and p-state tweaking; Vega 10 and its HBM2 will beat any (out of box) 5700XT. (perhaps, excepting top-shelf OC cards like the Red Devil). The Fiji (Fury), Vega 10, and Vega 20 GPUs were really something else. Even when benchmarks may say they're obsolete, there's just something about 'em Atop their impressive qualities they're a curio/relic already to collectors; we're extremely unlikely to ever see HBM GPUs in the consumer-facing market, ever again. For gaming and many other User-level multi-tasking(s), yes; 64GB is more than sufficient, and faster is better (generally). I doubt there's any 'tweaking' or 'OCing' options in the SuperMicro BIOS, and I don't even want to start researching BIOS mods/modding... So, fastest supported-JEDEC in just 2 slots would've been my choice too. Similarly, I've accepted that 24GB RAM just isn't gonna happen on my K8N-DL. 2x4GB on ea. CPU is much easier to get stable. (BTW, It took 4 separate purchases of literal-lots of RAM to find 4GB ECC DDR1 that worked)

Have you tried your Navi cards in a ("Legacy") BIOS-only (no UEFI) board? IIRC, that can be a problem w/ super-new cards; the vBIOS is UEFI-only.

Ya looking for the fastest thing that's not a crazy bottleneck? Or, maybe something a lil more era-appropriate? There is a semi-true 'trope' that AMD drivers can be 'lighter weight' than nVidias (and playing w/ a GTX 970 on a dual-940 did reinforce that) Also, 3rd Party AMD drivers give very long legs to older cards. I'm about 'set' on putting my spare Vega 64 8GB ref card in my K8N-DL w/ 2 885s. Depending on what CPU you're rocking in that C32 server board, that's 'not far off' from what I decided on Edit: If you are indeed looking to shove something quite 'new' in there, BIOS-UEFI compatibility *may* become an issue. I've actually had better luck in that department than most folks. Ex: some people said a WX 9100/MI25 wouldn't work in an nForce 4 Pro board (like my K8N-DL). They were wrong. Works great! Looks like C32 was largely a 2011-2013 platform. (If I'm going for era-appropriate, I like to RP that the build got some pre-end-of-life upgrades after Sales-EoL. Sometimes 'workstation' cards are neat too, in-theme.) So, lets see what's good about 2014-2015. 2015 AMD: R9 390X 8GB 512-bit GDDR5 Fury X 4GB 4096-bit HBM FirePro S9170 32GB 512-bit GDDR5 2014 AMD: R9 290X 4GB 512-bit GDDR5 FirePro W9100 16GB 512-bit GDDR5 FirePro S9150 16GB 512-bit GDDR5 2014 nVidia: GTX 980 GTX Titan Black/Z Quadro K5200 (Kepler) 2015 nVidia: GTX 980 Ti GTX Titan X Quadro M6000/M40

In investigating trying to get 'AMD PRO' features all-around, I stumbled into an awesome fix for that problem Amernime Zone - Browse /Release Polaris-Vega-Navi at SourceForge.net SOURCEFORGE.NET Unique AMD WHQL Driver Tweaking Project That's a bummer, but that was actually (specifically) a question I'd been looking for answer on: What's the largest mount on the factory Accelero brackets? But, that's not the end for trying to mount that behemoth to Vega. the GPU mount-bracket is mounted to the Accelero itself w/ 2 screws, and the receiving threads are in the Accelero. Meaning, one could (fairly-easily) DIY-fabricate a mount w/ a Rotary Tool (Dremel) and other common power tools. Just need to find some appropriate thickness scrap metal. Get creative and you (or a neighbor) may already have something that the material can be salvaged from. My uhmmm 'homelab' is currently in a state of disabling disarray, otherwise: yes, I do have an 8GB V64 not currently in a machine (and calipers). I don't want to mod or disassemble that reference card; it's fated for my Win7x64+Linux, K8N-DL, Dual 885s, 2x2x4GB ECC DDR, Optane NVMe build that was last 3/4 working (with all said parts) on my bench. @Lifted it's all yours if you want it. I want to aircool mine; specifically, for the indefinite longevity. My 16GB HBM2 Vega is a DD'er but, working collectable. I really, probably, have an unhealthy fascination with HBM and other old high-bit-width cards. (Also, I already have mod-able kit for an H115i AIO that I bought back w/ my last i7-3770k's rebuild. JIC)

I was wondering if anyone else would 'think this direction'. However in (my) doing so, I stumbled upon something that really deserves its own thread... Between model-series and brand (on a given 'GPU SKU'), cooler mounts differ. Go onto Aliexpress, and start looking up GPU coolers for those cards; you'll find multiple different mounting patterns. At least 1 'type' is 65x65mm. "close", but not close enough. If it wasn't for the fact that the 'threaded part' is in the cooler, you could just take a drillbit and 'worble-out' the hole a mm. I know it's a bit 'tinfoil hat', but I can't help but feel that's related to Board Partners classifying such specs as 'confidential'. I'd say I've observed purposeful non-intercompatibility

I thought so too but... Amazon has it *from* Raijintek. https://www.amazon.com/RAIJINTEK-Morpheus-Heat-Pipe-Heat-Sink-Options/dp/B00R6YQYQ2 Raijintek also still lists the cooler. RAIJINTEK WWW.RAIJINTEK.COM Raijintek provides the products - high compatibility, extreme engineering, remarkable performance, amazing design, necessity of demand and multifunctional usage – for high recommendation...

Good idea. I did manage to non-physically verify that is the mount via both Aliexpress listings for coolers, and old Reddit Posts. I too, have been on a quest to try and find an affordable air cooler that'll fit my MI25 -> WX9100. I've almost bought 3 'for parts' Vega/Fury 'partner' cards that came w/ better coolers. Sadly, many completely delete the frontplate and have too big of a coldplate to just mount-on unmodified. I want to keep the frontplate for VRM and board-component cooling, and I don't want to have to dismount it (potentially, ruining thermal pads). So far, the 'best' option is just 'eating it' and buying a Morpheus II Core off Amazon. Unfortunately, that costs more than the card did... Note: The Morpheus series will not mount onto Vega 56/64 cards while the frontplate is still installed (at least, not without A LOT of cutting). However! MI25's frontplate is MUCH SHORTER than Vega56/64's. At most, I see a teensy bit of material towards the Bracket needing filed-away. (which, could be CAREFULLY done w/o dismounting the plate)

Fury, Vega, and VII's GPU+HBM Heatsink mounting pitch is 64mm x 64mm. s far as I can research, that 64x64mm square mount is 'unique' to AMD's HBM-GCN/NCU cards (not sure about BC-160/V520/V540 HBM-Navi, though).