Turing disappointing

[Earnist_]

Nobody has pointed this out, the 1650 super has same cuda cores (1280) as 1060. Its based on the turing architecture and is hardly faster than the 1060. Comparing the 2 shows how much disappointing turing was this generation 

[Sakkura]

The 1650 Super only has 32 ROPs where the 1060 6GB has 48. It's not just about the shaders.

[i_build_nanosuits]

Actually isn't the 1060 a fair bit stronger than the 1650?

[i_build_nanosuits]

nevermind i missed the ''super'' lol

[TMobile]

I think they perform about the same, but new the 1650 super would be a bit cheaper 

[Marbo]

If you're comparing gens shouldn't you be putting it against the 1660?

[Zodiark1593]

5 hours ago, Earnist_ said:

Nobody has pointed this out, the 1650 super has same cuda cores (1280) as 1060. Its based on the turing architecture and is hardly faster than the 1060. Comparing the 2 shows how much disappointing turing was this generation 

The 1650 is a lower tier card with a much lower launch price. If you wish to compare identical tiers, you’re comparing the 1060 against the 1660 Super or TI. 

[SpaceGhostC2C]

33 minutes ago, Marbo said:

If you're comparing gens shouldn't you be putting it against the 1660?

 

24 minutes ago, Zodiark1593 said:

The 1650 is a lower tier card with a much lower launch price. If you wish to compare identical tiers, you’re comparing the 1060 against the 1660 Super or TI. 

But gens and tiers are bit architectures. I think what OP wants is a GPU equivalent of IPC, and for that you'd want to compare two products where the only difference. 

 

The problem, as @Sakkurapointed out, is that there are several elements that go into a GPU design, so at some point you have to decide which ones are supposed to remain constant for the comparison, and which are "part of the architecture" be hence allowed to vary.

 

No one would say Coffee Lake is farther than Kaby because the K i7 does better than the other K i7 at cinebench, ignoring the core count increase. But what about cache, memory controllers, I/O...? 

[Zodiark1593]

13 minutes ago, SpaceGhostC2C said:

 

But gens and tiers are bit architectures. I think what OP wants is a GPU equivalent of IPC, and for that you'd want to compare two products where the only difference. 

 

The problem, as @Sakkurapointed out, is that there are several elements that go into a GPU design, so at some point you have to decide which ones are supposed to remain constant for the comparison, and which are "part of the architecture" be hence allowed to vary.

 

No one would say Coffee Lake is farther than Kaby because the K i7 does better than the other K i7 at cinebench, ignoring the core count increase. But what about cache, memory controllers, I/O...? 

Very much so. Bottlenecks also exist within the GPU depending on the graphical load presented. In shader bound scenarios, the two should be pretty evenly matched. In fill bound scenarios (think smoke, big explosions, full screen effects, etc), the higher ROP count of the 1060 would allow it to pull ahead, especially at high resolutions. Everything else appears to be a pretty even match, clock speed aside. 

The architecture it layout of the shaders themselves may also be different as well, as we’ve seen between Fermi to Kepler, and from Kepler to Maxwell. 

6 hours ago, Sakkura said:

The 1650 Super only has 32 ROPs where the 1060 6GB has 48. It's not just about the shaders.

Memory bandwidth is also a pretty big deal in regards to graphics performance. Though the 1650 Super and the 1060 6 GB are matched here as well. 

[Earnist_]

4 hours ago, Zodiark1593 said:

The 1650 is a lower tier card with a much lower launch price. If you wish to compare identical tiers, you’re comparing the 1060 against the 1660 Super or TI. 

I'm comparing 1650 super to 1060 because they both have same number of cuda cores but different architecture.  Its a good way to base architectural improvements 

[Zodiark1593]

16 minutes ago, Earnist_ said:

I'm comparing 1650 super to 1060 because they both have same number of cuda cores but different architecture.  Its a good way to base architectural improvements 

There doesn’t appear to be a lot of improvements made to the CUDA core that really increases IPC. Much of the performance gains from the past couple generations have been from cramming more cores onto the chip, new memory, and upping clock speeds, along with the inclusion of Tensor cores for RTX cards. Which is fine tbh, as shader code is extremely easy to parallelize, and mucking around with the inner workings could break compatibility with other things and require fixing, like CUDA. 
 

In purely shader-bound scenarios, the slightly higher clock speed of the 1650 Super should allow it to pull ahead. CUDA cores, even if equal, are only a part of the story as there’s more to a GPU than just shader hardware. ROPs, TMUs, memory bandwidth, geometry engines, etc will all play a factor. For a long time, Radeon cards had a lot more shader performance, but was limited in other ways compared to their Nvidia counterparts. (Not the case anymore btw). 
 

The 1650 has fewer ROPs than the 1060, which will come up in fill-bound scenarios though. Think smoke, thick grass, explosions, anything that is layering lots of translucent/transparent elements or otherwise causes lots of overdraw. Resolution influences the performance impact of overdraw. If fill limited, the 1650 will probably actually be slower than the 1060. 

[BuckGup]

Most people just fixate on the latest gen but I somewhat agree. A 980ti OCed matches a 1080FE which matches a 2070 give or take depending on what cards you use but people think last gen is obsolete or not powerful at all but that's not the case

[Marbo]

12 minutes ago, BuckGup said:

Most people just fixate on the latest gen but I somewhat agree. A 980ti OCed matches a 1080FE which matches a 2070 give or take depending on what cards you use but people think last gen is obsolete or not powerful at all but that's not the case. 

Like most things with pcs, there isn't a one size fits all. 

 

A 1080 ti would beat my 2060 super hands down for gaming. But my 2060 S is about 15-20% faster rendering in Daz studio. In iray benchmarks it's around the same speed as 2 X 980 ti's 

[Roswell]

4 hours ago, Earnist_ said:

Its a good way to base architectural improvements 

No, it isn't. There's a litany of contributing factors when it comes to GPU performance, it's not a CPU.

 

4 hours ago, Zodiark1593 said:

There doesn’t appear to be a lot of improvements made to the CUDA core that really increases IPC.

IPC for a GPU isn't really a thing. The strategy isn't really to keep cramming more cores either, my old 780 Ti had 2880 CUDA cores, the 2080 Super has 3072. 

[The1Dickens]

5 hours ago, Zodiark1593 said:

There doesn’t appear to be a lot of improvements made to the CUDA core that really increases IPC. Much of the performance gains from the past couple generations have been from cramming more cores onto the chip, new memory, and upping clock speeds, along with the inclusion of Tensor cores for RTX cards. Which is fine tbh, as shader code is extremely easy to parallelize, and mucking around with the inner workings could break compatibility with other things and require fixing, like CUDA.

 

1 hour ago, Vitamanic said:

IPC for a GPU isn't really a thing. The strategy isn't really to keep cramming more cores either, my old 780 Ti had 2880 CUDA cores, the 2080 Super has 3072. 

This blog just posted an article about the birth/evolution of the CUDA architecture, and how it greatly differs from a CPU. The blog is a wealth of information, from dissecting id game engines, to writing a business-card-sized ray-tracing ID thing. Fabien Sanglard

 

Edit: Figured it might be something either of you find interesting, so tagged you both.

Edited May 9, 2020 by The1Dickens
Friendly tag reason

[Swallowtail]

About 5 to 10% faster than the GTX1060 in some titles and also at a much lower launch price . So i'd say that the 1650 super isn't disappointing at all. Though i would've liked if they ship this at 6gb vram, 4gb is gonna be obsolete in 2 years time max.