Radeon VII - RTX2080 competitor

Yes, they're finally fighting back Nvidia's high end GPU (2080Ti and Titan are still the winners yes I know)

As to previous thread where I said they're going to launch RTX2070, GTX1070 and GTX1060 challangers at lower prices, it's an answer for some folks here that were complaining about the high end cards that are left with no competition.

Cost: 700$

• 7nm Technology Process
• 3840 Stream Processors
• 16 GB HBM2 Memory
• 1 TB/s Memory Bandwidth
• AMD FreeSync2 HDR Technology

Card will be sold from 7th of February with game bundle which will include Devil May Cry 5, Resident Evil 2(remaster) and The Devision 2

its and interesting card can't wait for reviews, but it feels like amd's version of a 1080Ti, too little too late.

watching live stream lisa didn't seem to pleased as i think she knew that this was perhaps the weakest part of her presentation, i think she would have much prefered a better gpu than vega to announce but they had to do something to compete and she kept stressing that amd is for the gaming community. maybe she was underpromising with the hope of over delivering. at any rate i think this Radeon VII like the RX 590 is a stop gap place holder until amd can do a complete redisign of their gpu to offer a really attractive product. she did comfirm in interviews following that amd is developing its own ray tracing gpu.

ray tracing gpu? That reminds me of the ps3 back in the day.

Any advancement from AMD is a good thing. They need to keep pushing performance on their cards. Nvidia has been too dominant and is taking advantage of consumers with that dominance. Price gouging, shady practices, etc.. are not good for consumers. AMD putting out some good cards and Intel getting serious about GPUs in 2020 is all good news.

AMD doesn't need to redesign GCN, AMD needs to increase rasterization, ROPS and improve delta color compression.

1080 Ti has superior rasterization and ROPS when compared to Vega 64.

I would rather see it with less vram at a lower price of 550-600.

And lol at that TDP - a lot of people over here kept saying to me that 7nm will deal with all the power draw issues, but here we are and AMD's 7nm GPU is competing with nvidia's 12nm 2080 and is still higher TPD - just like I said these GPU are going to be. So much for the BS rumors on RX3080.

As for the TDP, AMD is going with 300W for the Radeon VII, 5W higher than the Vega 64 but basically the same. Regarding architectural changes, AMD says Vega 7nm is mostly the same as the previous Vega, with just some small tweaks. Obviously one of the biggest differences is that Vega 7nm now supports four stacks of HBM2, compared to two stacks on the previous iteration.

Kinda sucks that they're back into competition but again comes 2nd place, I don't know about pricing but I'm disappointed by the TDP since it's a very important to me - which means it'll be hotter than Nvidia - a total no go for me.

Looks like those Navi rumors were too good to be true after all. AMD needs a Ryzen equivalent to happen for Radeon. This is basically Vega 64 all over again.

At least Nvidia is planning to support freesync which is nice, so I'm no longer in a need to switch to red team.

According to Anandtech, VII has 128 ROPS which is logical when it has four HBM stacks. VII includes proper 64bit floating point hardware support which is different to Vega 64's silicon.

AMD didn't design high end gaming GPU with near zero 64bit floating point hardware wastage. VII with 60 CU is just a salvage MI60 with lower 16GB HBM v2 memory.

I don't think VII has 8 Shader Engines with 8 rasterization units.

VII's TDP is not a problem since I have 3KW solar cells (planning to increase towards 4KW)

Basically Vega 2.0.

Release a card 6-12 months late to the party and offer similar performance to an equally priced nvidia counterpart while having higher power draw.

@mastershake575: You do realize Nvidia and AMD measure TDP/Power Consumption differently, like how a 185W/250W 1080/1080Ti hits easily 250W+/350W+ power consumption and thermal throttle points (82C/92C) at stock speeds. Or a better offender like how a "95W" i7/i9 drawing more power and heat than a 16-Core Threadripper CPU.

@mastershake575: and vega was already a flop at the consumer level, except for crytominers really. why lisa would double down on this i don't know i hope its a few extra silicon that they have lying around from cards they made for prosumers and content creators and not the little money they made from cryptomining being wasted on this.

@ronvalencia: which is why i think they need to rethink the architecture to take proper advantage of the 7nm process and both DX12 and Vulkan api's. if they don't then amd gpu's are going to have to continue to use brute force to achieve parity with nvidia especially on DX 12 which means cramming more transistors into an already dense space, higher silicon failure rates, more power draw, higher temps. this is not acceptable.

Before Vega IP, AMD's DX12 Async compute is just a smoke screen for AMD's ROPS problem e.g. compute shader runs with TMU read-write path connected to L2 cache instead of pixel shader with ROPS read-write path without L2 cache connection.

Vega v1, AMD introduces ROPS being connected to L2 cache just like NVIDIA's Maxwell/Pascal/Turing ROPS design. Vega 64 is still limited to 64 ROPS while NVIDIA GPUs like GTX 1080 Ti has 88 ROPS.

Vega v2, AMD increased ROPS to 128 units. It's unknown if AMD has increased rasterization units (mass floating point to integer pixels conversion hardware). Vega 64 and GP104 has quad rasterization units while GP102 has six rasterization units. RTX 2080 has six rasterization units with 64 ROPS.

After Vega v2, next stop for AMD should be rasterization units increase.

TFLOPS is meaningless without classic GPU hardware functions to expose it's TFLOPS power.

FYI, VII = Vega II

Never thought of that, nice.

@ronvalencia: all of what you have stated are design changes to the present architecture, and im saying that it wont matter amd is trying to solve a problem by brute force especially with direct x 11 and 12. now either they are going to continue on this path and convince all devs to use vulkan for their games to beat nvidia or they are going to admit that the present architecture is inefficient and make a drastic change. do more with less.

re FYI VII = Vega II, really? isn't that what was said, literally, or maybe @mastershake575 should have said, im praphrasing "basically vega II". II is still 2 in roman numerals rite? im confused

https://tomclancy-thedivision.ubisoft.com/game/en-us/news-updates/342087/the-division-2-pc-features-specs-detailed

Elite – 4K | 60 FPS

• OS: Windows 7 | 8 | 10
• CPU: AMD Ryzen 7 2700X | Intel Core I9-7900X
• RAM: 16 GB
• GPU: AMD Radeon VII | Nvidia Geforce RTX 2080 TI
• VRAM: 11 GB
• DIRECT X: DirectX 11 | 12

Using past history with The Division v1 benchmark

$699 for the same performance of a almost 2 year old 1080 Ti?...

Should we be impressed or excited?... Also usually when AMD or Nvidia show benchmarks they show the best examples, if the best examples of Vega 7 is 2 games out of the three shown being on par with a RTX 2080 which is not much better than a 1080 Ti then I can imagine that when reviews hit you will find that in a benchmark run of 10+ modern games we will find the Vega running worse in 50% of the games.

This GPU would be impressive a year ago at $699... Now at that price range you need a selling point like Ray Tracing other wise why would I get the same performance or slower to a card with Ray Tracing for the same price?... You would have to be a complete fan boy to buy the card.

Were are you getting that 128 ROPS claim from? It's been confirmed it's still 64.

https://www.extremetech.com/gaming/283649-the-amd-radeon-viis-core-configuration-has-been-misreported

https://www.anandtech.com/show/13832/amd-radeon-vii-high-end-7nm-february-7th-for-699

https://www.pcgamesn.com/amd/amd-radeon-vii-release-date-specs-and-performance

I made "IF" statements with 128 ROPS with other topics.

extremetech

The good news is, this explains some things. It wasn’t at all clear why AMD would double up ROPs but leave GPU core count and texture unit resources unchanged.

There's a precedent with Vega M GH with 64 ROPS read-write paths and 24 CU(with 96 TMU read-write paths) .

If AMD is pushing async compute shader with 240 TMU read-write paths optimizations, then 64 ROPS read-write paths are seriously bottlenecked which has less consistent performance results across many games i.e. games with heavy TMU read-write optimizations gets closer to NVIDIA's TFLOPS counterparts while heavy ROPS bound being the worst cases. There's a reason for VII landing on RTX 2080 level which is the fastest 64 ROPS Turing card. In other topics, I made comment on rasterization problem with VII despite "128 ROPS".

READ my comment on "It's unknown if AMD has increased rasterization units (mass floating point to integer pixels conversion hardware)" .

1. READ the topic's thread posts for context.

2. I have my own doubts with VII rasterization power.

Rasterization has two hardware components which are

• ROPS read and write unit
• Rasterizer Unit

Again, READ the topic's thread posts for context.

AMD needs to re-design their GCN architecture. This will probably be a great card, but running hot and showing more and more the limits of the architecture.

@ronvalencia: amd has been working with ubi and bethesda for awhile now and as usual when it comes to framerate some titles will favor nvidia and others will favor amd. this division title should amd, becuz it will take advantage of async which amd has an edge in, rapid packed math, which favours amd, the 16GB vram becuase ultra detail 4K games should demand more at that rez. none of this solves the issue that i mentioned.

AMD needs to scale GCN's Shader Engine (SE) count similar to R9-280's dual SE to R9-290X's quad SE move and that's before "Mr TFLOPS" joined AMD in 2013.

Mr TFLOPS = Raja Koduri

Increasing SE also increase rasterizer unit count.

GTX 1080 Ti and RTX 2080 Ti has six GPC (equivalent to SE) with six raster engines.

Vega 56 at 1710 Mhz with 12 TFLOPS beating Vega 64 at 1590Mhz with 13 TFLOPS shows improved rasterizer on Vega 56 at 1710 Mhz clock speed has better results.

Fury X added 20 CU on existing R9-290X's quad SE layout which should have been two extra SE units with 11 CU per SE. Six SE units could have yielded 96 ROPS.

The alternative is to unify read-write TMU and ROPS design. Vega 64 has 256 TMU and 64 ROPS read-write units up to 1536 Mhz. GTX 1080 Ti has 225 TMUs and 88 ROPS with up to 1800Mhz. TFLOPS is meaningless without read-write units.

Vega 64 design is bias two towards compute/TMU read-write path.

Pixel Shader has ROPS read-write path

Compute Shader has TMU read-write path

Games with compute/TMU read-write path optimizations are closer to NVIDIA's TFLOPS counterparts. Cryptocurrency uses compute/TMU read-write path.

Turing has Rapid Packed Math for INT4, INT8 and FP16

Pascal GP102 has Rapid Packed Math for INT8, instructions scheduling issues which integer instructions blocking floating point instruction issue at SM level.

-------------

Vega 64 has Rapid Packed Math for INT8 and FP16

Vega II has Rapid Packed Math for INT4, INT8 and FP16

The main reason for async compute is for TMU read-write path which Vega 64/VII is competitive against NVIDIA TFLOPS GPU counterparts e.g.

Vega 64 has 256 TMUs up to 1536 Mhz, 64 ROPS up to 1536Mhz, 4MB L2 cache

GTX 1080 Ti has 225 TMUs up to 1800 Mhz, 88 ROPS up to 1800 Mhz, 3MB L2 cache

RTX 2080 has 184 TMUs up to +1900 Mhz, 64 ROPS up to +1900 Mhz, 4MB L2 cache

VII has 240 TMUs up to 1800 Mhz, 64 ROPS up to 1800 Mhz, ??MB L2 cache unknown. At least 4MB L2 cache.

----

RTX 2080 Ti has 272 TMUs up to +1900 Mhz, 88 ROPS up to +1900 Mhz, 6 MB L2 cache

@ronvalencia: amd wasn't comparing the VII aka II to 2080Ti they were comparing it to the 2080, the only reason why II aka VII is even mentioned with the 2080Ti in the division 2 game 4K rec specs is becuz VII aka II will be the best card that amd has in the enthusiast market.

why would amd work with ubi to produce the game and not have any cards make enthusiast the rec specs, i mean you help to make a game to take advantage of your hardware but can't recommend any of your hardware to play said game at its best settings, that makes no sense dude, they had to put something otherwise amd would be a laughing stock.

edit: and as i said amd in has an advantage in RPM

2080 Ti's TMU count is 272 hence I doubt AMD's VII async compute/TMU read-write path could catch RTX 2080 Ti.

From my 1080p-60Hz range experience, RX 470 4GB put a lot more stress on my PSU than GTX 1060 6GB that I currently own. It also ran hotter and louder (it did have smaller fans, but still, GTX 1060 never reaches 60C no matter what I stress it with, RX was running at 70-75C). They are both 120W cards officially.

@PfizersaurusRex: Radeon R9 300, RX400-500 are failure of series, let's hope Navi will change some things.

With RX-480, AMD wasn't able to master 8 ROPS read-write per 32 bit memory channel like in Pascal GP104. AMD was stuck with 4 ROPS per 32 bit memory channel design.

This is why AMD is pushing for async compute shader PR which shifts read-write operations to TMUs.