AMD Vega Makes an Appearance on CompuBench

[Raevenlord]

An AMD RX Vega video card has apparently made its way towards CompuBench. Granted, the no-name AMD graphics card could be an Instinct accelerator instead of AMD's consumer-oriented RX Vega graphics cards. However, the card did appear on CompuBench under the 6864:00 device ID, which had already appeared under a Vega Linux patch issued by the company. granted, this doesn't necessarily make it a consumer graphics product, so we'll have to look into this with some reservations.





We can at least assume this is an RX Vega graphics card - and if other leaks are true, this could be the top of the line RX Vega Nova graphics card. The GPU as identified by CompuBench carries 64 Cus (Compute Units), which, paired with AMD's typical (and confirmed for Vega) design of 64 processing cores per CU, yields the expected 4096 stream processors from the top end Vega. And it would seem that Vega having been built from the ground up with higher frequency in mind is true, with this GPU in particular carrying a 1600 MHz frequency, a far cry from AMD's RX 580's 1430 MHz.



Coupled with architecture improvements, this should yield a nice performance boost. The only wrinkle in this CompuBench cameo is the fact that the GPU in question appears with a 16 GB HBM2 memory subsystem, which looks a little out of place from what we know of Vega. We know that showcased dies carried only two HBM2 stacks (at an estimated 4 GB per stack), which would yield a total of 8 GB of HBM2 memory. However, AMD slides on the Vega architecture have shown GPUs surrounded by four stacks of HBM2 memory, so that doesn't mean much. But I have to say that such a grand memory pool on a consumer graphics architecture seems at odds with AMD's approach with their HBCC (High bandwidth Cache Controller), which is supposed to significantly shrink memory requirements by allowing a smart, precise asset allocation to the GPU's memory pool.



View at TechPowerUp Main Site

 

[RejZoR]

So, 4 stacks of HBM2 means 4096bit bus, just like on Fury X ?

 

[HTC]

So, 4 stacks of HBM2 means 4096bit bus, just like on Fury X ?

Maybe i'm interpreting this wrong but wasn't Vega supposed to have 2 stacks of HBM2 with double the bandwidth (each) to make the same overall bandwidth as Fury X? I definitely remember seeing Raja holding a "Vega chip" with only 2 HBM2 stacks, unless that particular one wasn't the "big" Vega.

Coupled with architecture improvements, this should yield a nice performance boost. The only wrinkle in this CompuBench cameo is the fact that the GPU in question appears with a 16 GB HBM2 memory subsystem, which looks a little out of place from what we know of Vega. We know that showcased dies carried only two HBM2 stacks (at an estimated 4 GB per stack), which would yield a total of 8 GB of HBM2 memory. However, AMD slides on the Vega architecture have shown GPUs surrounded by four stacks of HBM2 memory, so that doesn't mean much. But I have to say that such a grand memory pool on a consumer graphics architecture seems at odds with AMD's approach with their HBCC (High bandwidth Cache Controller), which is supposed to significantly shrink memory requirements by allowing a smart, precise asset allocation to the GPU's memory pool.

It also doesn't make much sense to have 16 GB HBM2 in big Vega when they're actively seeking to reduce the memory requirements and that tells me that this particular card may be one of those MI 25 (IIRC, that's it's name, right?) deep learning chips.

 

[ADHDGAMING]

Maybe i'm interpreting this wrong but wasn't Vega supposed to have 2 stacks of HBM2 with double the bandwidth (each) to make the same overall bandwidth as Fury X? I definitely remember seeing Raja holding a "Vega chip" with only 2 HBM2 stacks, unless that particular one wasn't the "big" Vega.




It also doesn't make much sense to have 16 GB HBM2 in big Vega when they're actively seeking to reduce the memory requirements and that tells me that this particular card may be one of those MI 25 (IIRC, that's it's name, right?) deep learning chips.

Its a numbers war, just like Intel is doing with the i9s, AMD understands that the average consumer will see the VRAM number and not fully understanding it they will either go with the bigger number or if the numbers are even they will go with price. But, lets be real here they will most likely just buy what their friends recommend which will likely be Nvidia regardless of how much ass Vega ends up kicking. The history of Nvidia V AMD is filled with AMD cards that are just beast but Nvidia's Fanbase is just more hardcore.

 

[FordGT90Concept]

As I said in the other thread, I think this is the Radeon Instinct/MI25. It has a lot of VRAM (16 GiB, needed for compute), it has about 13 TFLOP which matches what Raja said about the Vega Cube, and the ID number (6864) is lower than the other Vega chip we saw (687F) which suggests it's been around a while. I don't think this is the consumer Vega chip.

 

[HTC]

As I said in the other thread, I think this is the Radeon Instinct/MI25. It has a lot of VRAM (16 GiB, needed for compute) and it has about 13 TFLOP which matches what Raja said about the Vega Cube. I don't think this is the consumer Vega chip.

Agreed: the biggest hint to this is the amount of VRAM.

 

[Nkd]

As I said in the other thread, I think this is the Radeon Instinct/MI25. It has a lot of VRAM (16 GiB, needed for compute), it has about 13 TFLOP which matches what Raja said about the Vega Cube, and the ID number (6864) is lower than the other Vega chip we saw (687F) which suggests it's been around a while. I don't think this is the consumer Vega chip.

so if instinct is running at 1600mhz I would assume the gaming chips should run even higher. That is what makes me wonder if this is their limited edition gaming card we heard rumors of with crazy packaging with 16gb of HBM2. I was expecting around 1500 best for the instinct card according to the Tflops rating.

 

[HTC]

Thinking about it some more, there's another reason why the amount of HBM2 makes this card NOT any of the consumer cards: do you dudes honestly believe that with 16 GB HBM2 this card would only cost $599, which is the supposed price of "big" Vega?

 

[mat9v]

Yes, compare HERE that score to Radeon Fury (not even Fury X) that is also in the result database to be "amazed" by the performance. If we were to judge by this benchmark then Vega would be a total failure
After all to reach 1080Ti level of performance it would need roughly 75% speed increase, not the ~30% at best present there.

 

[Estaric]

Yes, compare HERE that score to Radeon Fury (not even Fury X) that is also in the result database to be "amazed" by the performance. If we were to judge by this benchmark then Vega would be a total failure
After all to reach 1080Ti level of performance it would need roughly 75% speed increase, not the ~30% at best present there.

Your right how could be so blind and hopeful that Vega might do something. Clearly Nvidia will always be the best... Okay being a smart a** aside you should never look at the paper specifications and say oh yea can't be as powerful.