New Upcoming ATI/AMD GPU's Thread: Leaks, Hopes & Aftermarket GPU's

Likely due to the tensor cores that will not be present in the GV102/4/6.

 

50% as per Nvidia claim, does seem a bit far fetched.
That is basically what Maxwell brought compared to Kepler.


Volta features a major new redesign of the SM processor architecture that is at the center of
the GPU. The new Volta SM is 50% more energy efficient than the previous generation Pascal
design, enabling major boosts in FP32 and FP64 performance in the same power envelope.

Furthermore 50% boost in FP32 does not necessarily translate to 50% gaming boost. I think we already saw some regression here with Pascal.
NV might be taking more AMD approach, aka more raw Flops due to focusing on Deep Learning.

 

I think you underestimate the resources required, especially for a chip that large, even just adding 5billion transitors to an existing design is not a small feat.

In other news , https://youtu.be/XC5Dy_b-kE8?t=8m55s

Amd 's Mike Mantor mentioning that vega can exceed 1700mhz.

 

That and DX12 seems to lean heavily on FP32 performance which is why Fermiand Kepler f***ing suck at it.

Um GV104 will/should be around the same transistor count as GP102 (actually should be less) so not seeing the point there.

Also considering the water cooled Vega is clocked just under 1700Mhz I would hope it could clock past it. However, how much power will it draw there?

 

Not talking about the gv104, GP100 is 15.3~ billion transistors, GV100 is 21billion~.

that is one expensive chip,
link

Also: Power consumption for vega will be high at high clocks, but if the Vega 56 can use modified vbioses, that will be one hell of a deal, especially if the clock scaling issues can be resolved like on fiji.

 

AMD’s Ryzen Threadripper 1950X At 4.1GHz Is Out Of This World

This leak comes courtesy of a redditor who has actually managed to get his hands on a 1950X early on and not only benchmark it but do so after overclocking it. This is one of several reports we’ve seen as of late idnicating that Threadripper is, surprisingly, a better overclocker than mainstream Ryzen.
The consensus seems to be that a 4.0GHz overclock on all cores is within reach on most Threadripper chips.

The owner of this particular 1950X chip reported achieving 4.0GHz on all cores at 1.25v, with a Thermaltake 3.0 liquid cooler. That is simply unattainable by the majority of mainstream Ryzen processors at this voltage. The owner tested the chip once at 4.1GHz in Geekbenck 3 and once at 4.0GHz in Cinebench R15.

The motherboard he used is from Asrock, the precise model wasn’t specified. All 8 memory DIMMs were populated with 3066MHz DDR4 memory.Let’s start out with the Cinebench R15 result first. Here, the 1950X manages to deliver a whopping 3337 points with all sixteen of its cores running at 4.0GHz.
By comparison a similarly priced Intel Core i7 7900X gets 2169 points at stock and maxes out at 2464 with a hefty 4.6GHz overclock. That comes out to a 35%+ difference when both chips are overclocked.
Not even the world record Cinebench R15 run of the 7900X at 5.96GHz with liquid nitrogen cooling is enough to surpass Threadripper’s score.

At 4.1GHz the figures become even more astronomical for Threadripper. The chip manages to score 58391 points in Geekbench 3. Which not only is by far the highest score for any desktop processor we have ever seen but also secures it a 37% lead over a 7900X overclocked to 4.5GHz which manages to score around ~42600 points in the same test.

The Ryzen Threadripper 1950X is by far the most powerful desktop processor we have seen tested to date and once overclocked its performance jumps from the realm of very impressive to, there’s no better word for it, otherworldly.

The embargo on Threadripper reviews goes up this upcoming Thursday and we can’t wait to see what else this chip can do. -> HH

PS. BTW HH when can we have some Good Vega Leaks from you :banana:

 

Then you need to re-read what you quoted. It clearly said, "GV100 will never be a GTX product." The second half of the quote was in reference to the GV104.

 

RX Vega TimeSpy GPU scores w latest driver version ... 22.19.666.1

1630MHz: 7209
http://www.3dmark.com/compare/spy/2192756

1750MHz: 7580
http://www.3dmark.com/compare/spy/2192971

 

So still around 1080 performance my 1080ti is getting 10K graphics in timespy.

 

Yep, stock 1080 scores are around 7400 ...

 

With over 120mhz in the coreclock over the 1630mhz one and only so little increase. There is certainly a bottleneck somewhere and my guess is memory as it was bottlenecking Fury X already.

 

This seems appropriate, lol

https://gfycat.com/UnhappyPalatableIbis

 

This look very bad unless it's the Vega 56 :s

 

I never claimed gv100 would be, and "Also Volta is not a huge departure from Pascal" applies to both the GV100 and the GV104 geforce or not.

My Only point is that GV100 took such a large chunk of nvidia's R&D budget, and manpower, that it may effect the development cycle of other gpus, like the upcoming GV104,

"I think you underestimate the impact of the GV100 on nvidia's total R&D" is what I should have clarified.

performance appears to scale mostly linearly with an Hbm overclock , which For the most part confirms that.

16% hbm clock increase results in a 6% performance increase, however performance scaling is mostly linear, doesn't hit diminishing returns even at 1100mhz in most of the applications this guy tested.

https://youtu.be/9v24mmNCPGg?t=2m42s

 

Highly unlikely as I have read that Vega 56 is launching a month after Vega 64. I would say this is one of the 3 variants of Vega 64. Just not sure which one.

 

Well 945 memory is dead give away since Vega 56 has lower frequency.

@user1

I don't really understand what is going with AMD on that part even Fury X was bandwidth starved. And they went with quite a bit less bandwidth this time.

In reality Vega stock is pushing something like 280GB/s anyway.

 

Exactly, I initially was hoping that this was Vega 56 performance and not Vega 64. But like you said Vega 64 has memory clock of 945 vs Vega 56 of 800. All I can hope, is it's not the watercooled variant. Not good at all.

 

Theres rumours over here that the AMD "Black" reference cards are going to be sold at the same price as the "Limited Edition" cards, ie both at $599 for the aircooled models.

 

AMD Officially Releases Vega Shader ISA Documentation

http://developer.amd.com/wordpress/media/2013/12/Vega_Shader_ISA_28July2017.pdf

 

THX
•
New packed 16-bit math instructions.
V_PK_MAD_I16 V_PK_MUL_LO_U16 V_PK_ADD_I16 V_PK_SUB_I16
V_PK_LSHLREV_B16 V_PK_LSHRREV_B16 V_PK_ASHRREV_I16 V_PK_MAX_I16
V_PK_MIN_I16 V_PK_MAD_U16 V_PK_ADD_U16 V_PK_SUB_U16
V_PK_MAX_U16 V_PK_MIN_U16 V_PK_FMA_F16 V_PK_ADD_F16
V_PK_MUL_F16 V_PK_MIN_F16 V_PK_MAX_F16 V_MAD_MIX_F32
V_MAD_MIXLO_F16 V_MAD_MIXHI_F16 S_PACK_{LL,LH,HH}_B16_B32