Vega 20 announced with 7.64 TFlops of FP64

[M344587487]

Horizon event announcement (bad audio quality):

https://youtu.be/GwX13bo0RDQ?t=3270


Nice summary:

https://www.youtube.com/watch?v=YmPimQp7xLE&t=350


Highlights:

• 1:2 FP64 performance at 7.64 TFlops
• 32GB HBM2 at ~2Gbps
• ~1TB/s memory bandwidth
• ECC
• Probably expensive as hell as a professional card

What could these specs translate to for prime hunting, particularly PRP? The memory bandwidth is just over double that of Vega64, is that a limiting factor to roughly double the Vega64 performance regardless of what the improved FP64 performance might offer? Am I right in thinking that gpuowl already uses FP64 despite the 1:16 ratio of current Vega, meaning that if it can be kept fed it can be naively said that throughput multiplies by ~8 all other things equal? If FP32 is involved I can't make even an uneducated guess. Anyone knowledgeable want to hazard a guess as to the estimated performance uplift versus Vega64? How about versus a Titan V (which has ~650GB/s memory bandwidth and ~6.9TFlops of FP64 performance)?

[kriesel]

Originally Posted by M344587487

Highlights:

• 1:2 FP64 performance at 7.64 TFlops
• 32GB HBM2 at ~2Gbps
• ~1TB/s memory bandwidth
• ECC
• Probably expensive as hell as a professional card

You're correct about PRP making use of FP64. As does LL, and P-1. That in combination with the large 32GB ram would make this card outstanding for P-1.

[xx005fs]

Originally Posted by M344587487

Horizon event announcement (bad audio quality):

https://youtu.be/GwX13bo0RDQ?t=3270


Nice summary:

https://www.youtube.com/watch?v=YmPimQp7xLE&t=350


Highlights:

• 1:2 FP64 performance at 7.64 TFlops
• 32GB HBM2 at ~2Gbps
• ~1TB/s memory bandwidth
• ECC
• Probably expensive as hell as a professional card

What could these specs translate to for prime hunting, particularly PRP? The memory bandwidth is just over double that of Vega64, is that a limiting factor to roughly double the Vega64 performance regardless of what the improved FP64 performance might offer? Am I right in thinking that gpuowl already uses FP64 despite the 1:16 ratio of current Vega, meaning that if it can be kept fed it can be naively said that throughput multiplies by ~8 all other things equal? If FP32 is involved I can't make even an uneducated guess. Anyone knowledgeable want to hazard a guess as to the estimated performance uplift versus Vega64? How about versus a Titan V (which has ~650GB/s memory bandwidth and ~6.9TFlops of FP64 performance)?

I have both a vega 56 and titan v and titan v never should go above 1300MHz even with 1040MHz memory yielding about 770GB/s BW. Vega is quite limited without ocing the memory at 1:16 ratio so that 1TB/s will def help. I would assume it being around 30% faster than Titan V which put it about 230% faster than vega 56 overclocked.

[M344587487]

Official AMD Horizon video with better audio: https://youtu.be/kC3ny3LBfi4?t=3090

A presentation they did after that talk going into more detail about MI50 and MI60: https://www.youtube.com/watch?v=m0h6-VfH3Xo

Highlights of the second talk:

• MI50 and MI60 are the two GPUs
• MI50 has 60 CUs, MI60 has 64 CUs, which translates to ~10% peak theoretical differences
• Slide shows MI50 has 6.7TFlops and MI60 has 7.4TFlops of FP64
• MI50 has 16GB HBM2, MI60 has 32GB HBM2
• Both have 1TB/s memory bandwidth and 4096 bit bus
• Both have 300W TDP
• ECC extends beyond HBM and includes registers which it didn't before
• Lower latency cache than current gen Vega
• The first we're likely to get our hands on these cards is via renting from cloud vendors

If memory is our main bottleneck the two cards should be functionally the same for PRP, trial factoring and P-1 may be a different matter. The pricing will be interesting, not that they're going to be available to the likes of us for a while.

[xx005fs]

Originally Posted by M344587487

Official AMD Horizon video with better audio: https://youtu.be/kC3ny3LBfi4?t=3090

A presentation they did after that talk going into more detail about MI50 and MI60: https://www.youtube.com/watch?v=m0h6-VfH3Xo

Highlights of the second talk:

• MI50 and MI60 are the two GPUs
• MI50 has 60 CUs, MI60 has 64 CUs, which translates to ~10% peak theoretical differences
• Slide shows MI50 has 6.7TFlops and MI60 has 7.4TFlops of FP64
• MI50 has 16GB HBM2, MI60 has 32GB HBM2
• Both have 1TB/s memory bandwidth and 4096 bit bus
• Both have 300W TDP
• ECC extends beyond HBM and includes registers which it didn't before
• Lower latency cache than current gen Vega
• The first we're likely to get our hands on these cards is via renting from cloud vendors

If memory is our main bottleneck the two cards should be functionally the same for PRP, trial factoring and P-1 may be a different matter. The pricing will be interesting, not that they're going to be available to the likes of us for a while.

AMD didn't publicly disclose their int32 performance, hence trial factoring performance is going to be weak if they stay the same as current gen Vega and any other GCN GPU (compared to Volta and Turing). However, if the memories are overclockable then MI60 definitely have an edge than MI50 because as tested on Titan V at 800GB/s BW (OCed a bit further), it saturates the memory system at about 1230MHz which yields to about 6TFLOP DP. scale it up to 1TB/s (assuming same FFT scaling between Nvidia and AMD arch) meaning that it will use 7.2TFLOP of DP which saturates MI50. However, the difference should be negligible.