Originally posted by bridgman
View Post
Announcement
Collapse
No announcement yet.
AMD Sends Out Linux Graphics Driver Patches For "Arcturus" As New Vega Derived GPU
Collapse
X
-
- Likes 1
-
Putting on my speculation hat, and having glanced over available reports and through the patches, it looks very much like a widened Vega 20 -- the reported 8192 SPs and 2x cache size, the 'MI-100' callsign, wide-loads, and what seem to be indications of 2x wider SIMD registers all point that way. So how are they doing it? I don't think its a dual-package or dual-die with InfinityFabric in between, or I don't think we'd see all this 2x widening going on; and besides that Vega 20 is only ~330mm square with graphics function intact, doubling the SPs while dropping graphics functions would probably come in comfortably under 600mm square, and much smaller than nVidia's reticle-busting ~800mm square dies.
I also believe more recent versions didn't so much hit the 4096 SP limit due hard architecture limits, but moreso with the difficulty in balancing them against the practical limits of adding more graphics-related things around them, in balance. That idea is bolstered by the fact that available GCN projects have scaled far better with compute workloads than with rendering, and perhaps also that memory overclocking has been more profitable than core overclocking for rendering workloads.
The simplest idea might be to just double the SIMD, register-width, and datapath across the board, which would seem to be straight-forward once freed of concerning itself with rendering workloads.
But I also wonder if compute HPC workloads would benefit at all from more ACEs and if their might be some vestigial architecture limits related to that... So looking at how Navi is organized, I wonder if they're doing the dual shader-engine thing with 64 x 2 shader engines -- I believe the fine-grained scheduling at the shader engine level already has enough work available to keep two independent groups of SIMDs busy.
Now, what would be really cool is if this were a not just a socketed product, but if there were also CPU cores in the package or on-die along with a full compliment of DDR4 channels in addition to on-package HBM acting as cache or scratchpad. It seems to me that some number of sockets all filled with those would make a more-regular HPC architecture than one where you have the same number of sockets where one-out-of-N is the odd man out. If you look at some of the other vector architectures out there, they have fairly strong scalar CPUs alongside the vector engines -- not ~8-wide super-scalars like Ryzen or Skylake, but more than the scalar capabilities inside GPU architectures.
Leave a comment:
-
Originally posted by ihatemichaelWith all due respect: I don't like having to set up different environment variables for my desktop to work correctly. While I appreciate the fact that this helped, it still feels a lot like a workaround and not a real fix.
Depending on whether other users are seeing the same problem it may make sense to change the default in upstream code... I don't *think* I have seen this as a common problem on Raven but Marek might know better.
- Likes 2
Leave a comment:
-
Originally posted by ihatemichaelI have a Raven Ridge and the experience is still as awful as the first day I got this computer (it's been almost a year and I still see screen artifacts).
Phoronix: RadeonSI Primitive Culling Lands In Mesa 19.2 The past few months AMD's Marek Olšák has been working on primitive culling support for the RadeonSI Gallium3D driver and last week that code was merged into the Mesa 19.2 development code... http://www.phoronix.com/scan.php?page=news_item&px=Mesa-19.2-RadeonSI
Leave a comment:
-
Originally posted by bridgman View Post
Are you suggesting we stop releasing new chips, or that we release them without Linux support ?
Neither seems workable to me.
Personally have big expectations for this chip, only left wait to 2020
- Likes 1
Leave a comment:
-
Originally posted by ihatemichaelAs always, working on new stuff without finishing the old stuff, typical AMD.
Neither seems workable to me.Last edited by bridgman; 16 July 2019, 12:43 PM.
- Likes 3
Leave a comment:
-
Originally posted by L_A_G View Post
Considering how that thing is connected to the new Pro trough a dual connector setup with a standard PCIe and an additional proprietary connector, most probably so they don't have to put in any additional power connectors on the other side or back of the card, I don't think this is that exact card. Specially not when, unlike this card, that card is also used as a video card.
However you probably are right on the money about the card being a dual Vega 20 board. Most probably it's a data center variant of that design and will probably offer some fairly competitive performance like AMD cards have long since done if you know how to utilize them properly.
- Likes 1
Leave a comment:
-
Originally posted by tuke81 View PostDual Vega 20 for apple maybe?
However you probably are right on the money about the card being a dual Vega 20 board. Most probably it's a data center variant of that design and will probably offer some fairly competitive performance like AMD cards have long since done if you know how to utilize them properly.
- Likes 1
Leave a comment:
-
Something the likes of Radeon Pro V340 in "less expensive" would be really nice. ;-)
- Likes 1
Leave a comment:
Leave a comment: