Likely due to the shortage of HBM2 and cryptocurrency miners who should all be crushed by their own machines, it's not on the market yet. Not even the vaunted AMAX Project 47 1 Petaflop EPYC/Radeon Instinct Supercomputer is available, and it was supposed to be for sale starting Q4 2017.
You won't unless you're a Business ktorn, because there's a minimum Purchase Number., although I'd have to talk to my Supplier about what that is exactly.
Now note, while I have Instinct (MI25) that power our Render and Codebase Server., we didn't purchase these "OEM" but rather as part of Pre-Built Server Blades.
Also as another key point the Radeon Instinct Series are ACCELERATORS not Graphics Cards.
As such they're only accessible as Compute Devices / Processors, and have no Display Outputs... these are NOT Consumer Products., as for all intended purposes you have to create Custom Software Chains specifically to utilise them for Purpose., hack I haven't even tried to see if DirectX 12 or Vulkan recognise them as Graphics Processors for Multi GPU or Compute purposes; instead we use them exclusively as part of our ROCM and OpenCL Pipelines to Accelerate Compute Workloads (i.e. Radeon Pro Render and Visual Studio Build) as well as Compute Simulations.
If you are still interesting in acquiring one, I'd recommend one of AMDs Supply Partners; such-as HP.
The only reason that the NVIDIA Equivalents appear to have a Price Tag (and are sold via NVIDIA.com)., is because they also provide Stand-Alone Versions along side their Server MCM Versions; and believe me the MCM Versions are quite a fair bit more expensive than the Stand-Alone Workstation Cards; with simile minimum purchase requirements.
Remember that Accelerators are essentially produced on an Order Basis., as opposed to Mass Production that Retail or even Workstation Hardware is produced on.
With this said., I'm almost certain that AMD is planning to release their 7nm Retail and Workstation Series from Q2 2019 on-ward... in-fact I'd put good money on it being Mid-April 2019 that we'll see them "Launch" the RX 600-Series.
Now if they call it Vega-Based or Navi-Based., well rumours are going somewhat nuts over this... but I can guarantee you that they are due in the very near future.
In fact we only have to look at the Athlon 200GE as proof of this., because the performance of the Vega-Based 3CU at 1000MHz, is frankly far beyond the performance difference between Fiji, Polaris and Vega as they're on the market right now.
Keep in mind that the ONLY reason that the Fiji-Based (3rd Gen) GCN in the A-Series 7th Gen APU is faster than the 5th Gen APU is because they're running at 1,100MHz as opposed to 750MHz … the only reason that the Vega (Polaris) Based (4th Gen) GCN in the Ryzen Mobile APU is because they're running at 1,400MHz as opposed to 1,100MHz.
For all intended purposes 3rd, 4th and 5th Gen Architecture has been almost identical in terms of "Graphics IPC" Performance., but curiously these new Athlon 200GE Series Processors see a DRAMATIC increase in performance that has essentially flown under-the-radar because … well let's be honest, who care about the Absolute Low-End Budget Hardware for Gaming or Graphics?
Very informative response, thanks leyvin.
I have contacted them through my company, even though I do want them for personal use. I'm building an EPYC-based home server for machine learning using ROCm and hence my interest in those cards, however...
I have just found an official distributor that lists/sells the MI25 online but the price ($9,959) is more than I budgeted for (that's about my total budget for the whole system, and I was looking to buy 2 cards), so...
I will probably wait for the 7nm products as you suggest. That's also the reason why I chose a relatively "low-end" EPYC 7351P, as I plan to upgrade to a 7nm EPYC when AMD releases them next year.
I might stick a couple of consumer grade cards in the mean time, just to start preparing my ROCm setup. Perhaps a couple of refurbished R9 Nanos, since they'll be just temporary.
For a Home Server (without external Refrigeration / Cooling) then I'd actually recommend the Workstation Cards., either the WX9100 (16GB / Vega 64 on 12nm / £1600) or WX8200 (8GB / Vega 56 / 12nm) … I don't know how much the WX8200 as they only announced it recently, but it's likely ~£1200 given it essentially replaces the Polaris-Based Pro Duo.
Aside from not requiring an external Cooling Solution, getting access to the Professional Drivers., ROCm Compliancy., as well as also being able to later repurpose them for Individual Workstations as opposed to Reseller dumping. It ends up being better value and more flexible for a small solution.
In terms of specs the WX9100 seems suitable, and the price is definitely within budget, but the blower style cooling will surely be too loud for me (especially since I plan to run multiple cards).
I was already planning a custom cooling solution for the Instincts, including dedicated ducts for intake fans from bottom of the case.
If I get the WX9100 I will likely have to mod it to make it more like an Instinct in terms of cooling, with airflow injected from the back. That'll mean a custom 3D printed shroud, and a bit of dremeling to remove the round metal frame around the back of the fan, to allow airflow from the back of the card. Not crazy about subjecting a brand new $1,500 card to a Dremel though.
The "Blower" Coolers are actually better suited to a Front-to-Rear Airflow Case., but more than that they're designed to be Semi-Passive Cooling … so while the Cooler might look identical to the Reference RX Vega... it's a somewhat higher capacity Heatsink, which with the lower Clocks keeps them relatively Low-Noise; especially as noted if you have a decent airflow case.
As another point, Blower Coolers are better for drawing in Air from Confined Spaced (i.e. with Multiple Closely Mounted together),. where-as Traditional Axial Fans require a reasonable (about 1-3") clearance to be effective.
This said, they have identical Mounting Points to the Retail RX Vega / Polaris Hardware... so I mean there are plenty of options available.
I suppose there's nothing like actually getting one and testing it. Perhaps inside the Define R6 it won't be too loud.
leyvin, thanks a lot for all the great feedback.