Tuesday, September 20th 2016
AMD Vega 10, Vega 20, and Vega 11 GPUs Detailed
AMD CTO, speaking at an investors event organized by Deutsche Bank, recently announced that the company's next-generation "Vega" GPUs, its first high-end parts in close to two years, will be launched in the first half of 2017. AMD is said to have made significant performance/Watt refinements with Vega, over its current "Polaris" architecture. VideoCardz posted probable specs of three parts based on the architecture.
AMD will begin the "Vega" architecture lineup with the Vega 10, an upper-performance segment part designed to disrupt NVIDIA's high-end lineup, with a performance positioning somewhere between the GP104 and GP102. This chip is expected to be endowed with 4,096 stream processors, with up to 24 TFLOP/s 16-bit (half-precision) floating point performance. It will feature 8-16 GB of HBM2 memory with up to 512 GB/s memory bandwidth. AMD is looking at typical board power (TBP) ratings around 225W.
Next up, is "Vega 20." This is one part we've never heard of today, and it's likely scheduled for much later. "Vega 20" is a die-shrink of Vega 10 to the 7 nm GF9 process being developed by GlobalFoundries. It will feature 4,096 stream processors, too, but likely at higher clocks, up to 32 GB of HBM2 memory running full-cylinders at 1 TB/s, PCI-Express gen 4.0 bus support, and a typical board power of 150W.
The "Vega 11" part is a mid-range chip designed to replace "Polaris 10" from the product-stack, and offer slightly higher performance at vastly better performance/Watt. AMD is expecting to roll out the "Navi" architecture some time in 2019, and so AMD will hold out for the next two years with "Vega." There's even talk of a dual-GPU "Vega" product featuring a pair of Vega 10 ASICs.
Source:
VideoCardz
AMD will begin the "Vega" architecture lineup with the Vega 10, an upper-performance segment part designed to disrupt NVIDIA's high-end lineup, with a performance positioning somewhere between the GP104 and GP102. This chip is expected to be endowed with 4,096 stream processors, with up to 24 TFLOP/s 16-bit (half-precision) floating point performance. It will feature 8-16 GB of HBM2 memory with up to 512 GB/s memory bandwidth. AMD is looking at typical board power (TBP) ratings around 225W.
The "Vega 11" part is a mid-range chip designed to replace "Polaris 10" from the product-stack, and offer slightly higher performance at vastly better performance/Watt. AMD is expecting to roll out the "Navi" architecture some time in 2019, and so AMD will hold out for the next two years with "Vega." There's even talk of a dual-GPU "Vega" product featuring a pair of Vega 10 ASICs.
194 Comments on AMD Vega 10, Vega 20, and Vega 11 GPUs Detailed
If AMD selects the dirt cheap 512 GB/s HBM2 then they better not price this more than $899, and I would hope they would choose 8GB to keep prices down. On the other hand if they select the 720 or 1000 GB/s HBM then they could put 16GB on it and charge $1000+.
Also timing is key. It seems like if they can launch an UBER card Before March they may have a full year before Nvidia can respond with Volta, and the worst they would have to put up with is some 1180 with 8GB of HBM in July. But if Vega launches in June it will give Nvidia supremacy of the high end for FAR too long - hopefully a 490 with GDDR5 (X?) will launch this fall either way.
The Fury X also had the correct specs and more-or-less performed as well as it should, it just needed more optimized drivers.
They had plenty of marketshare with the weaker 4870, 5870 (This one did dominate for while), and 6970. Then when the 7970 and 290X were kicking ass they lost marketshare slowly but surely. Technically Radeon's best days were when they purely whent for price/perf and efficiency.
Having said that, the performance gulf between even a cut-down Vega and Polaris will be absolutely massive. I think they will be quite silly if they don't release at least an RX 490 this fall. They wouldn't even need a new die either, as long as 14nm is maturing they could just release P10 clocked at 1450MHz with GDDR5X and an 8-pin connector. That would give it a 20% boost over the 480 and allow it to take on the 1070 in at least DX12 games.
We can see that 1.6GHz GDDR5 (204GB/s) is the right amount for a 750MHz Polaris 10 GPU...
2.2GHz GDDR5 (280GB/s) would be ideal for 1Ghz P10 GPU...
2.8GHz GDDR5 (360GB/s) would be ideal for 1.25GHz P10 GPU
2.9GHz GDDR5 (370GB/s) would be ideal for 1.3GHz P10 GPU.
RX 480 has just 256GB/s of bandwidth... not ideal even at 1GHz GPU clocks.
Vega 10 will need about 75% more bandwidth to feed its CUs... so about 630GB/s would be "ideal."
EDIT:
Also, you can see the diminishing returns from GPU frequency from all of the tests I ran:
1250MHz and 1350MHz are much closer together than 1150MHz and 1250MHz...
I think AMD will have 2 Vega GPU's early next year:
-Cut-down 3584-SP with 768-GB/s HBM - $500 - $600
-Full-die 4096-SP with 1 TB/s HBM - $650 - $800
Keep in mind that nothing is ideal right now, so even if the chips are bandwidth starved AMD has to release something.
The funny thing is AMD has also been able to price gouge -> in the Mid-Range. The 480 is beating the 1060 by quite a bit (It's still constantly selling out). As such the 480 is about $20-$40 more than it was supposed to be, and the 470 is being sold for the same price as lower-tier 480's since stock issues are such a problem. 4GB 460 prices are laughable, but again it is because it has zero competition in the space.
NAVI is still expected for early 2018, a year after Vega.
Vega 20 (if it is to be called Vega) is a real part and is indeed 2019 ... but it won't be HBM2 32GB. Would make no sense whatsoever to have that much VRAM with that many shaders, unless 4K 144Hz / 8K is the norm by then, and the 7nm process take stock clocks to 2.5Ghz - 3Ghz. Neither of which is likely.
Vega 11 is in no way replacing Polaris 10. It will sit at or above Polaris 10, probably at or above Fury X performance. They'll be sold in the same segments that you'd expect them to inhabit today.
Vega 10 details are mostly right.
The thing that bothers me about Vega 10 & 11 is GF's copy (not very) exactly (in fact quite badly) Samsung 14nmFF LP+ process. If they're using it again for Vega (the high performance part), I expect power consumption and clocks may disappoint a little again.
We know Samsung are building something for AMD ... surely AMD should be using their new high performance 14nmFF process(es) for Vega 10, 11 and Zen?
On DX11 the driver does farm off asynchronous tasks to driver worker threads where possible.
NVIDIA's DX11 driver already has multiple threads and asynchronous tasks.Vega 11 seems to be Scorpio's GPU build with it's worst working chip yield around 6 TFLOPS instead of Polaris 10's worst working chip yield of 4.9 TFLOPS for RX-470. Vega 11 effectively replaces Polaris 10.
For Scorpio, MS waited for AMD's medium size GPU chip to reach 6 TFLOPS with good yields.
Scorpio with Polaris 10 at 6 TFLOPS target would have terrible working chip yields.
A lot of reports stated that Samsung is working on making slower (768 GB/s) HBM2 chips that are far cheaper than HBM1. Having one unified architecture means they don't have to segment between GDDR and non-GDDR dies, which is incredibly helpful when it comes to utilizing early yields.
Furthermore it would shave off about 25 - 50w off of all of their models, and it would give them all a nice 20-50% performance boost from last years model.