Tuesday, June 5th 2018

AMD Demonstrates 7nm Radeon Vega Instinct HPC Accelerator

AMD demonstrated the world's first GPU built on the 7 nanometer silicon fabrication process, a Radeon Vega Instinct HPC/AI accelerator, with a 7 nm GPU based on the "Vega" architecture, at its heart. This chip is an MCM of a 7 nm GPU die, and 32 GB HBM2 memory stacks over four stacks (4096-bit memory bus width). It's also the first product to feature a removable InfinityFabric interface (competition to NVIDIA's NVLink interface). There will also be variants based on the common PCI-Express 3.0 x16. The card supports hardware virtualization and new deep-learning ops.
Add your own comment

100 Comments on AMD Demonstrates 7nm Radeon Vega Instinct HPC Accelerator

#26
Vya Domus
I always knew Vega was held back greatly by an unfitting manufacturing process.
Vayra86However, what we get is ANOTHER product nobody really cares for, which is just a shrink of something that wasn't good to begin with.
You , don't care about it. Come on , I'm sure you where these things are targeted at.
Vayra86It doesn't excel in other areas either relative to the competition's offering.
I am sure you know how this works as well. You can bet this will be at least half or maybe even a third of the price of a V100. It's very much competitive where it needs to, actually. And it's not just performance that matters so does software support , Nvidia for a few years refuses to support anything above OpenCL 1.2 whereas AMD does and for these things that sure as hell matters.
Posted on Reply
#27
Vayra86
Vya DomusI always knew Vega was held back greatly by an unfitting manufacturing process.

You , don't care about it. Come on , I'm sure you where these things are targeted at.

I am sure you know how this works as well. You can bet this will be at least half or maybe even a third of the price of a V100. It's very much competitive where it needs to, actually. And it's not just performance that matters so does software support , Nvidia for a few years refuses to support anything above OpenCL 1.2 whereas AMD does and for these things that sure as hell matters.
I have seriously been looking for stories of large volume Vega purchases and I do not - not a few, but simply do NOT - see them anywhere. Regardless of segment.

I'm sure theoretically this is fantastic and possible. And then there's the real world.
Posted on Reply
#28
enxo218
with the recent news on amd's efforts in cpu and gpu development I'm slowly turning into a fanboy
Posted on Reply
#29
Vayra86
enxo218with the recent news on amd's efforts in cpu and gpu development I'm slowly turning into a fanboy
I can get that... on the CPU side and also in APU AMD is doing some fantastic stuff. Dedicated GPU though... the last announcement I had any hope for was Fury X and it being the 'overclocking dream'... look how that turned out. Since then it's only gotten worse.
Posted on Reply
#30
jabbadap
Vayra86I can get that... on the CPU side and also in APU AMD is doing some fantastic stuff. Dedicated GPU though... the last announcement I had any hope for was Fury X and it being the 'overclocking dream'... look how that turned out. Since then it's only gotten worse.
In some cultures nightmares are dreams too... Topic says HPC, so what is the fp64/fp32 ratio?
Posted on Reply
#31
iO
Such a shame its going to be too expensive for a gaming card. But 2H18 means the manufacturing runs much better than projected as they talked about a launch sometime in 2019 back at CES.
Which makes Turing in 7nm also more likely..
Posted on Reply
#33
jabbadap
It's tiny for sure. On that PoV it might actually come with gaming variant with 16GB of hbm2. But that size makes me very much doubt it will have full fp64 compute power.

Edit: what is the die size estimation? Quickly counted 14*23 ~ 322 mm²
Posted on Reply
#34
ppn
jabbadapIt's tiny for sure. On that PoV it might actually come with gaming variant with 16GB of hbm2...
1,229 GB/s seems a bit of overkill so 16GB 2048 bit is more adequate for desktop users (and keeps the miners away). +50% bandwidth improvement over VEGA10, and 35% overclock to 2.0 Ghz for average of 40% improvement and 1080Ti will be reached at 232mm2.
Posted on Reply
#35
cucker tarlson
JismYeah and what creates this bandwidth? The speed of the ram.
The bus width.
ppn1,229 GB/s seems a bit of overkill so 16GB 2048 bit is more adequate for desktop users (and keeps the miners away). +50% bandwidth improvement over VEGA10, and 35% overclock to 2.0 Ghz for average of 40% improvement and 1080Ti will be reached at 232mm2.
They've got 7nm and HBM2, for crying out loud, they should be running circles around 1080Ti.
Posted on Reply
#36
jabbadap
ppn1,229 GB/s seems a bit of overkill so 16GB 2048 bit is more adequate for desktop users (and keeps the miners away). +50% bandwidth improvement over VEGA10, and 35% overclock to 2.0 Ghz for average of 40% improvement and 1080Ti will be reached at 232mm2.
Well yeah it's so small that they might actually release it with 2 or 3 hbm2 chips for gaming card. That size of gpu is easily mass produced and it should be relative cheap.
Posted on Reply
#37
HD64G
If I was in the lead of RTG team, I would launch a Vega 7nm with +20% FPS from combination of higher clocks for both core and HBM and lower the power consumption to under 250W as possible. This could get very close or higher then 1080Ti in many games, could have competitive price even with the HBM2 added cost and would have no weak points at all (thermals, price, wattage). An even better Vega 56 which didn't have much appeal mostly due to its mining-craziness affected ultra high price (and partially due to the HBM2 shortage at least ultil recently).
Posted on Reply
#38
medi01
HD64GIf I was in the lead of RTG team, I would launch a Vega 7nm with +20% FPS from combination of higher clocks for both core and HBM and lower the power consumption to under 250W as possible. This could get very close or higher then 1080Ti in many games, could have competitive price even with the HBM2 added cost and would have no weak points at all (thermals, price, wattage). An even better Vega 56 which didn't have much appeal mostly due to its mining-craziness affected ultra high price (and partially due to the HBM2 shortage at least ultil recently).
Still too little to late, I'm afraid, with August release of 2080 being planned.
Posted on Reply
#39
T4C Fantasy
CPU & GPU DB Maintainer
medi01Still too little to late, I'm afraid, with August release of 2080 being planned.
1180* lol
Posted on Reply
#40
HD64G
medi01Still too little to late, I'm afraid, with August release of 2080 being planned.
Me thinks that August will be a paper launch or a talk about the next gen features and most possibly will get new nVidia GPU series only in 2019 on 7nm and GDDR6. If we get anything new from them in 2018, it won't be anything great. Just a testament model for the next gen with small performance increase (10-20% max) and GDDDR6. Not easy to make any performance jump when going from 16nm to 12nm.
Posted on Reply
#41
Vayra86
HD64GMe thinks that August will be a paper launch or a talk about the next gen features and most possibly will get new nVidia GPU series only in 2019 on 7nm and GDDR6. If we get anything new from them in 2018, it won't be anything great. Just a testament model for the next gen with small performance increase (10-20% max) and GDDDR6. Not easy to make any performance jump when going from 16nm to 12nm.
On Computex in a closed doors speech Jensen was caught saying the next Geforce 'is going to take a looong time' making a waving motion with his hand. Then again - long is an abstract term.

Logic suggests we may see a refresh of Pascal similar to Kepler refresh.

And for some reason I get this feeling there may just be more in the tank with that refresh than just a 10-20% uplift. Not sure why. Gut sense
Posted on Reply
#42
HD64G
Vayra86On Computex in a closed doors speech Jensen was caught saying the next Geforce 'is going to take a looong time' making a waving motion with his hand. Then again - long is an abstract term.

Logic suggests we may see a refresh of Pascal similar to Kepler refresh.

And for some reason I get this feeling there may just be more in the tank with that refresh than just a 10-20% uplift. Not sure why. Gut sense
Exactly what I said. If launched in 2018, not much more than 10% increase in performance due to the small benefits going from 16nm to 12nm. If launched in 2019 oon 7nm though, it will surely get a big jump in performance. Up to 35% as AMD revealed for the next process.
Posted on Reply
#43
Vayra86
HD64GExactly what I said. If launched in 2018, not much more than 10% increase in performance due to the small benefits going from 16nm to 12nm. If launched in 2019 oon 7nm though, it will surely get a big jump in performance. Up to 35% as AMD revealed for the next process.
I do not think their next launch will be on 7nm. They have no reason to get in the crowded queue for 7nm production lines because Pascal is easily up to snuff for improvements to last another gen on 16nm. From a cost perspective, I would consider that much more logical and also much more Nvidia style. They literally already have product to cover a full new tier up with Volta and kicking Pascal cards down a tier.
Posted on Reply
#44
nemesis.ie
Didn't Lisa imply the gaming version of the new Instinct GPU would follow "soon"? There was also pretty much confirmation of linking chips/cards with infinity fabric.

At the size it is, if power can be kept under control, it should be very doable to pop two of these on one card that behave much closer to a monolithic GPU than multiple cards have done up to now.

If you can then have multiple cards connecting by re-purposing PCIe lanes to IF (like Epyc does) and if they can be made/sold at a reasonable price, we could have a real winner.
Posted on Reply
#45
sergionography
ppn1,229 GB/s seems a bit of overkill so 16GB 2048 bit is more adequate for desktop users (and keeps the miners away). +50% bandwidth improvement over VEGA10, and 35% overclock to 2.0 Ghz for average of 40% improvement and 1080Ti will be reached at 232mm2.
35% more performance i believe is more general about 7nm vs 14nm rather than specifically about vega. Problem here is that normally its not very likely for a gpu of this size to be rated at 250w tdp. So we are more likely to see this at 150-180w with slightly better performance than 14nm vega64 due to more stable clocks and better bandwidth. If anything this is more likely to be made into multi chip modules(MCM) via infinity fabric similar to threadripper, but then for that to happen they need to overcome the vega 4096 core scaling limitation, since i heard vega doesnt scale beyond 4096 radeon cores.
Posted on Reply
#46
londiste
T4C FantasyVega 20

Vega 10
HBM size is probably the best common size here, we can be pretty certain that this is HBM2 on both.
  • HBM2 die size is 7.75 mm × 11.87 mm (91.99 mm²)
  • This puts Vega 10 at around 20.94 mm × 26.79 mm (560.98 mm²).
  • And Vega 20 at around 14.79 mm × 23.72 mm (350.82 mm²).
Numbers do come from fairly small images so error might be around 3-4% and up to 10% for areas.
All in all - Vega 20 is ~35% smaller than Vega 10. Assuming this is direct die shrink (which it likely is), pretty nice.

For comparison, 350 mm² is around the size of Tahiti/Tonga in 7970/RX280/RX380. On the other side of the fence, GP104 in GTX1080 is a bit smaller and GM204 in GTX980 is a bit bigger.
Posted on Reply
#47
T4C Fantasy
CPU & GPU DB Maintainer
londisteHBM size is probably the best common size here, we can be pretty certain that this is HBM2 on both.
  • HBM2 die size is 7.75 mm × 11.87 mm (91.99 mm²)
  • This puts Vega 10 at around 20.94 mm × 26.79 mm (560.98 mm²).
  • And Vega 20 at around 14.79 mm × 23.72 mm (350.82 mm²).
Numbers do come from fairly small images so error might be around 3-4% and up to 10% for areas.
All in all - Vega 20 is ~35% smaller than Vega 10. Assuming this is direct die shrink (which it likely is), pretty nice.
Vega 10 without hbm is 484 i believe
Posted on Reply
#48
londiste
T4C FantasyVega 10 without hbm is 484 i believe
Yeah, that has been the official message. As much as I have seen that, measuring the actual dies always ended up somewhere over 500. So did my Vega. 560 might be a bit too much but I would still put the actual size inside my 10% error margin, at least 520.
Posted on Reply
#49
T4C Fantasy
CPU & GPU DB Maintainer
londisteYeah, that has been the official message. As much as I have seen that, measuring the actual dies always ended up somewhere over 500. So did my Vega. 560 might be a bit too much but I would still put the actual size inside my 10% error margin, at least 520.
Do you still have a vega?
Posted on Reply
#50
londiste
T4C FantasyDo you still have a vega?
Not any more, no :(
Posted on Reply
Add your own comment
Nov 20th, 2026 16:19 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts