DavidGraham
Veteran
Also Radeon VII is PCIe 3.0, not 4.0.
AMD Radeon VII Announcement and Discussion
- Thread starter Deleted member 13524
- Start date
-
- Tags
- amd
itsmydamnation
Veteran
Its a smart move, limit the ability to scale compute performance past one card to the MI range.
Someone on another forum had the idea they may just announce/market it as PCIe 3.0 at first, because some people might get confused, and may think PCIe 4.0 is not backwards compatible, or might complain because it isn't using PCIe 4.0 on PCIe 3.0 platforms.
They probably just want to charge more for it on professional cards.
They didn't announce any Radeon Professional cards so far, and it would be stupid not to limit the DP of a gaming card to 1/8 (like Hawaii was limited), if they plan to sell Pro cards.
This whole VEGA 20 adventure is strange.
A super crazy theory says: the original Greenland - highend chip - complementing Ellesmere and Baffin launch might actually have been VEGA 20@14nm. Things didn't go smoothly (TDP?) and instead we got cut-down and rushed variant known as - VEGA 10 - 14 months after Polaris. The original plan would put a bit slower VEGA 20 to end of 2016. Is there any hard evidence contradicting this?
A super crazy theory says: the original Greenland - highend chip - complementing Ellesmere and Baffin launch might actually have been VEGA 20@14nm. Things didn't go smoothly (TDP?) and instead we got cut-down and rushed variant known as - VEGA 10 - 14 months after Polaris. The original plan would put a bit slower VEGA 20 to end of 2016. Is there any hard evidence contradicting this?
Sorry, I forgot to mention the suggested cut-down concerned the chip as whole - this affects all target markets: HPC, AI, PRO graphics/content creation and gaming.
The supposed Greenland HPC APU concept had limited details besides probably having 1/2 DP and 4 xGMI links. The DP peak numbers seemed in line with the other unit counts being like Vega 10, and Greenland had the same two stacks of HBM2.
The concept had 4 xGMI links with a total of 100 GB/s bandwidth. This is notably slower than the xGMI links on Vega 20, although it's possible this was more for the benefit of matching up with the x86 dies and their more limited memory subsystems.
Those two differences seem like modest changes. DP is easily changed at design time, or could be throttle back after manufacturing readily. The cost of that functionality when it's not used fully seems modest, going by examples like Hawaii.
The xGMI links could have been jettisoned once the APU MCM concept was ditched for Vega and Zeppelin, although it's not a major disruption to most of the chip to remove them. Disabling them was simple for Vega VII.
Anarchist4000
Veteran
I'd guess to protect pro products with higher margins. VII would be a rather beefy video editing product like SSG with 16GB, 1TB/s, and PCIe4 feeding it. Not to mention that INT8 performance.
An now they claim they spoke to someone from AMD who told them FP64 would not be unlimited:
AMD’s Radeon VII GPU Will Not Support Uncapped Double-Precision (FP64)
AMD’s Radeon VII GPU Will Not Support Uncapped Double-Precision (FP64)
D
Deleted member 13524
Guest
The specs chart that Scott Herkelman gave to HardOCP also says the card is not a "Double Precision PCIe Accelerator".
Maybe they're still undecided, lol.
Regardless, if FP64 is axed then the card makes less sense as it corners the card's addressable market into an even smaller niche. And since the Instinct MI50/MI60 AFAIK can't be bought on retail, the 5 year-old FirePro W8100/W9100 are the only FP64 cards that AMD offers.
Unless they're launching a more expensive FP64-enabled Vega 20 down the road.. But if they are, why not announce it at CES?
The downside to AMD keeping their cards so close to their chest is that too few people know what's going on, it would seem. It goes without saying that this is not the answer I got last week.
