AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

Wait what?

DDR4 die spot price were at $10/GB in february this year, they are currently around $5/GB. Flash is collapsing even faster and the industry expect NAND producers to shift production capacity from NAND to DRAM, putting more downward pressure on DRAM going into next year.

Cheers
 
With very high DRAM cost, and HBM2 only adding a premium, AMD could not lower sales prices and make a buck. In order to command a highend sticker price, they had to be somewhat competitive with NVidias offerings. So instead of accepting a 20% lower sales price for a 10-15% less performing part, AMD pushed VEGA chips all the way up the schmoo-plot, with very high power consumption.
While this is somewhat plausible, it's hard to believe that AMD would produce a HBM2 based top-SKU with a 484/510mm2 die which would be 15% slower than the current Vega 64. That would be truly horrible compared to GP104 with GDDR and a 314mm2 die.
 
While this is somewhat plausible, it's hard to believe that AMD would produce a HBM2 based top-SKU with a 484/510mm2 die which would be 15% slower than the current Vega 64. That would be truly horrible compared to GP104 with GDDR and a 314mm2 die.

AMD almost certainly fell short of their own expectations, some of that might be GF's fault, some it was probably AMD's fault; ie. pushing higher frequency over lower power consumption is almost certain to backfire on a massively parallel SOC like a GPU (</armchair architect>)

I'm just suggesting AMD had to push Vega up the schmoo plot curve in order to not sell Vega at a loss. If the HBM2 had been half the price, they could have had a lower performing/cheaper SKU where all the high power/faulty CU dies could be sunk.

Cheers
 
Last edited:
AMD almost certainly fell short of their own expectations, some of the might be GF's fault, some it was probably AMD's fault; ie. pushing higher frequency over lower power consumption is almost certain to backfire on a massively parallel SOC like a GPU (</armchair architect>)

I'm just suggesting AMD had to push Vega up the schmoo plot curve in order to not sell Vega at a loss. If the HBM2 had been half the price, they could have had a lower performing/cheaper SKU where all the high power/faulty CU dies could be sunk.

Cheers

This had been conventional wisdom for a while, but Pascal showed that you can reach pretty high clocks without sacrificing (much) power efficiency. It runs at higher frequencies than Vega in most cases, while Turing runs higher still—albeit on 12nm, granted.
 
Last edited:
This had been conventional wisdom for a while, but Pascal showed that you can reach pretty high clocks without sacrificing (much) power efficiency. It runs at higher frequencies than Vega in most cases, while Turing runs higher still—albeit on 12nm, granted.
Using a different Fab though. Performance from a reasonably undervolted Vega isn't half bad and less leakage could be significant without knowing what the power curves of each process look like. Ignoring the possibility of a hardware defect and/or automatically compiled primitive shaders not currently working.
 
Not to mention you should never compare clocks between different architectures even if they have similarities, one could be optimized to do less work per clock but achieve higher clocks and vice versa
 
Vega Mobility is here, announced during an apple event:


https://www.amd.com/en/graphics/radeon-pro-vega-20-pro-vega-16

- 16 to 20 CUs
- 14nm (not 12nm)
- Single HBM2
- Premium ultra-portable notebooks
- Rapid Packed Math (Not Kaby Lake G's Polarega confirmed)



I wonder why this is coming so late. Also, the CU count puts it a small step above Polaris 11.

EDIT: It's available for the 15-inch Macbook Pro, so it's an incremental upgrade to the Polaris 11 560X
https://www.apple.com/macbook-pro/specs/
 
Last edited by a moderator:
Vega Mobility is here, announced during an apple event:


https://www.amd.com/en/graphics/radeon-pro-vega-20-pro-vega-16

- 16 to 20 CUs
- 14nm (not 12nm)
- Single HBM2
- Premium ultra-portable notebooks
- Rapid Packed Math (Not Kaby Lake G's Polarega confirmed)



I wonder why this is coming so late. Also, the CU count puts it a small step above Polaris 11.

EDIT: It's available for the 15-inch Macbook Pro, so it's an incremental upgrade to the Polaris 11 560X
https://www.apple.com/macbook-pro/specs/

I did not see this coming. My guess is that due to the use of HBM2, it's too expensive for anything other than MacBooks, so AMD just followed Apple's schedule, which is what it is.
 
IIRC, just before Polaris has been launched, it was stated that all future AMD chips are able to work with both GDDR and HBM. So I don't see why couldn't they make Vega 16/20 cards with 4GB GDDR5 to replace RX 560. Vega 20 CU @ 1600-1700 MHz would be a great card. Probably faster than GTX 1050 Ti. But also a decent mining card (undervolted), so that might be the answer

At least I hope we will soon see Vega 20 CU in Radeon Pro , maybe WX 4200. So if it's ~$300 or less, that might be my next GPU, with or without HBM, I don't care :)
 
I did not see this coming. My guess is that due to the use of HBM2, it's too expensive for anything other than MacBooks, so AMD just followed Apple's schedule, which is what it is.

And/or apple demanded they'd get to announce the GPU, which they clearly are in a position to do so.

This also points to Kaby Lake G not being a custom order by apple after all.
Odd stuff. Intel can't be extremely happy that apple is paying AMD for Vega Mobility when they're offering the 65W Core i7 8705G.
Kaby Lake G just seems misplaced in time, though. Intel positions it as a high-end laptop solution, but it came out at around the same time they launched the 6-core 45W solutions. I guess it was inevitably cannibalized by Intel's own Kaby Lake R.
 
IIRC, just before Polaris has been launched, it was stated that all future AMD chips are able to work with both GDDR and HBM. So I don't see why couldn't they make Vega 16/20 cards with 4GB GDDR5 to replace RX 560.
It wasn't just future chips, but several previous gens too. GCN has always been "modular" in the sense that you could easily "mix'n'match" different IP-blocks together, including memory controllers. The thing is, you still need to design the whole chip layout from the ground up, it's not that AMD's memory controllers would support which ever memory standard you want, but that you can use which ever memory standard supporting memory controller you want.
 
Well, between this and Sony/Navi rumours, it seems the story about AMD GPU Roadmap being tightly knitted with semi-custom work has some merit. No wonder Raja was not happy at AMD.

I'd buy it. Taping out chips today has gotten massively expensive, and gets even moreso the newer the node you get to. EUV might help for a bit, but that's still in the beginning stages of being deployed. "Semi-custom" would definitely be the operating word here, taping out a completely custom chip for anyone would get very expensive.

But this sudden announcement makes me wonder what other seemingly completely unleaked AMD products will just turn up. The RX 590 is a stupid idea, Polaris's architecture isn't designed to have clockspeeds even as high as it does with the 580, a jump to 12nm and pushing up clockspeeds even more isn't going to help performance much.

Vega on the other hand was designed with much higher clockspeeds in mind. With GF out of the game more or less AMD might be able to get 12nm Vegas to work on Samsung's process, since the base 14nm process GF used was licensed from Samsung anyway. With Samsung already (reportedly) being able to produce better 14nm chips than GF, a 15-25% higher clockspeed Vega might be possible on Samsung 12nm (ballpark guess). That'd put the card comfortably above a 2070 without necessarily having to spend a ton of money. But who the hell knows.
 
Didn't they said that Vega has around 40mb of cache ? So i agree, no CU or Rop, but they added a lot of cache. But maybe it was just to compensate some higher latency or something, linked to the higher frequency ? Anyway, the chip is a mess...
I believe the claim was 45MiB of SRAM across the entire chip, which would be caches, registers, buffers, and various scratch pads in everything from the graphics units to the miscellaneous cores and interface blocks.
Vega's amount is presumed to be higher than prior GPUs, but no breakdown was given and no number given for any prior GPUs to know how significant the increase was.

But this sudden announcement makes me wonder what other seemingly completely unleaked AMD products will just turn up. The RX 590 is a stupid idea, Polaris's architecture isn't designed to have clockspeeds even as high as it does with the 580, a jump to 12nm and pushing up clockspeeds even more isn't going to help performance much.

Vega on the other hand was designed with much higher clockspeeds in mind. With GF out of the game more or less AMD might be able to get 12nm Vegas to work on Samsung's process, since the base 14nm process GF used was licensed from Samsung anyway. With Samsung already (reportedly) being able to produce better 14nm chips than GF, a 15-25% higher clockspeed Vega might be possible on Samsung 12nm (ballpark guess). That'd put the card comfortably above a 2070 without necessarily having to spend a ton of money. But who the hell knows.
Perhaps it gives AMD an inexpensive option for filling out its GF wafer allocation?
Vega was noted to devote most of its additional transistor count over Fiji to reaching higher clock speeds, and its marketing mentioned getting design tweaks from the Zen team for the register files. Perhaps what Vega did for its optimization made it a little less trivial to port to an implementation on GF's 12nm (maybe the Zen team's time was taken up by either porting Pinnacle Ridge or the shuffling of resources for future architectures rather than Polaris).

Some of Vega's features would have been nice to carry into the 590's SKU, though perhaps some of them cost too much area for what revenue AMD expects of the product. The command processors and geometry front end hardware in the middle of Vega appears to be significantly larger than in prior GPUs, and since Polaris has 4 front-ends it would be inviting much of that penalty with some of the more significant motivations for the area cost being non-starters. If AMD's expected clock gains from Vega's optimizations weren't realized, perhaps the area cost for them wouldn't prove sufficient for 12nm, which gave Ryzen a bump in the upper reaches of the clock range at the cost of somewhat higher TDP.
 
Some of Vega's features would have been nice to carry into the 590's SKU, though perhaps some of them cost too much area for what revenue AMD expects of the product. The command processors and geometry front end hardware in the middle of Vega appears to be significantly larger than in prior GPUs, and since Polaris has 4 front-ends it would be inviting much of that penalty with some of the more significant motivations for the area cost being non-starters. If AMD's expected clock gains from Vega's optimizations weren't realized, perhaps the area cost for them wouldn't prove sufficient for 12nm, which gave Ryzen a bump in the upper reaches of the clock range at the cost of somewhat higher TDP.

even older Polaris (RX480) was somehow bottlenecked by slow memory so this much higher clocked version will suffer even more. Instead of tricky "cooking"Vega features into GCN4 they should rather invest in better memory controller (GDDR6/GDDR5X/wider memory bus). That would hepl a lot
 
even older Polaris (RX480) was somehow bottlenecked by slow memory so this much higher clocked version will suffer even more. Instead of tricky "cooking"Vega features into GCN4 they should rather invest in better memory controller (GDDR6/GDDR5X/wider memory bus). That would hepl a lot
Memory controllers aren't the issue, it's the caching and culling/sorting/prediction algorithms. Essentially a cache miss issue where memory latency is a factor in addition to having the miss. Hard to know precisely, but it ties into the geometry bottleneck. Somewhere there is a critical path that was likely missed and accounts for the higher voltages and power usage. Only takes one weak link to slow everything down. In theory that could be somewhat easily fixed in newer Vega variants, but no indication of that so far. Not unless Vega20 or the new mobile version is uncharacteristically fast.

There was some discussion of prefetching optimizations in patents a while back with Polaris.
 
Blackmagic eGPU Pro with Vega 56 video card
The Blackmagic eGPU Pro features a built-in AMD Radeon RX Vega 56, two Thunderbolt 3 ports, a new DisplayPort for connecting 5K displays, HDMI 2.0, 85W of charging power and four USB 3.1 connections. Designed in collaboration with Apple, the integrated design brings cutting-edge workstation-class graphics processing and computational acceleration to customers working in professional video, playing 3D games or using the latest virtual reality software.

The Blackmagic eGPU Pro features 8GB of HBM2 RAM, a massive 2048-bit memory interface with 410 gigabytes per second bandwidth, and 56 discreet compute units for up to 10.5 teraflops of processing power.
The Blackmagic eGPU Pro will be available in November for only US$1,199 from apple.com.
https://www.guru3d.com/news-story/blackmagic-design-announces-blackmagic-egpu-pro.html
 
The Blackmagic eGPUs are also not upgradable. From the FAQ:
Can I upgrade the GPU in Blackmagic eGPUs?
No, the design has been optimized for quiet operation so it’s better suited for creative customers. Blackmagic eGPUs combine the graphics processor, electronics, mechanical, cooling, power and extra connections all in a single unit.
 
Back
Top