HardOCP Interviews Scott Herkelman

AlphaAtlas

[H]ard|Gawd
Staff member
Joined
Mar 3, 2018
Messages
1,713
AMD launched the their first 7nm GPU, the Radeon VII, at CES 2019. The company published some internal performance figures to back up their claims, which we analyzed, and the launch itself caused quite a stir, but there are still some details about the GPU that AMD left out of their CES event. So we sat down with Scott Herkelman, the Vice President and General Manager of the Radeon Gaming Business Unit, and got some more details on Radeon VII, the Vega architecture, and the future of AMD's gaming division.

Check out the full interview here.

HardOCP: Is the Radeon VII a true "Vega20" GPU? Beside the smaller process, what improvements are in the new Vega vs the old one? Can we get full specs? Scott Herkelman: Yes, Vega20 is the underlying architecture for this product. We made some surgical enhancements to the Vega architecture to scale to frequencies on 7nm. We also increased the memory interface from 2048 to 4096 bits, all while reducing the footprint from 495mm2 to 331mm2 and we are super happy with the results. This chart shows the full Radeon VII specs...
 
They could have made it faster by making it bigger? I'm not at all sure how this works.
 
They could have made it faster by making it bigger? I'm not at all sure how this works.
Yes, if they actually redisigned the graphics core, crammed in another billion transistors and more rendering pipelines. However, this was just AMD leveradging what they had on hand. Using a die shrink and making the process more efficient and keeping costs down by shrinking it's footprint.

It's not a dumb move. AMD is finally getting some of it's initial Vega features supported by game devs which is a good thing. The fact that they know Ray tracing is important but it's far from mainstream. Thus far it's mostly screen space reflections and even that destroys frame rates. The fact that the card can go toe to toe with a 2080 is decent performance.

I don't think any of us expected AMD to make an Nvidia killer. If these cards don't space invader then I will happily pick one up, seeing as I just gifted my 1080Ti to my cousin. I need something with balls for my front room and this sounds like a decent option.

We will know more after we see actual, [H] benchmarks for the video card.
 
I do find the thing about why they included 16gigs of hbm to be a bit of an eyebrow raiser. If these are salvaged instinct cards (as many are saying) then the hbm would already be on the package, afaik it can't really be physically removed and if they disabled 8 gigs of it in bios it would just hurt the bandwidth (apparently vega 64 was bandwidth starved going by some comments).
 
I do find the thing about why they included 16gigs of hbm to be a bit of an eyebrow raiser. If these are salvaged instinct cards (as many are saying) then the hbm would already be on the package, afaik it can't really be physically removed and if they disabled 8 gigs of it in bios it would just hurt the bandwidth (apparently vega 64 was bandwidth starved going by some comments).
AMD has a bad habit of disabling RAM on cards... Didn't they do that with the RX 4 series? There was a 4GB card that you could flip a BIOS tag and it enabled the other 4GB.

I hear what you're saying tho. They must be going with HBM because they're invested in it... But... Availability and cost gated their older cards time to market . Maybe not that case now, but who in the Sam hell needs the memory bandwidth this card offers? It would be a miracle of this card could saturate half of that pipeline.

Edit:

Sorry, my mind was running away with my typing. You're saying this card is a rebranded MI50... Which looks about right. Still even if they disabled half the memory and the card only had 512 bits of memory bandwidth.. what would use that?

The 2080Ti Is 352 bit and it's good enough .

That's the only thing I see with AMD's marketing... That's fluff. The card isn't fast enough to saturate the memory pipeline.

I learned this a long time ago... You can have the world's fastest memory but if your GPU can't use it... What is the point other than marketing buzz terms and inflated costs?
 
Last edited:
AMD has a bad habit of disabling RAM on cards... Didn't they do that with the RX 4 series? There was a 4GB card that you could flip a BIOS tag and it enabled the other 4GB.

I hear what you're saying tho. They must be going with HBM because they're invested in it... But... Availability and cost gated their older cards time to market . Maybe not that case now, but who in the Sam hell needs the memory bandwidth this card offers? It would be a miracle of this card could saturate half of that pipeline.

Gamers likely won't make use of all the VRAM, but more VRAM heavy professional applications might. If everything pans out the card could be a decent cheaper alternative to professional grade cards in situations where spending thousands of dollars on a single card is either not needed or simply not feasible.
 
If these cards don't space invader then I will happily pick one up, seeing as I just gifted my 1080Ti to my cousin.

I clearly am in the wrong family....

Have a regular 1080 for myself and one for the wife. This is interesting and might be tempting for me. I used to be a huge amd and ati fan back in the day.
 
Gamers likely won't make use of all the VRAM, but more VRAM heavy professional applications might. If everything pans out the card could be a decent cheaper alternative to professional grade cards in situations where spending thousands of dollars on a single card is either not needed or simply not feasible.
You're correct. I wasn't looking at it like an affordable professional card, which it presents a helluva value as .
 
Back
Top