Nvidia Blackwell Architecture Speculation

  • Thread starter Deleted member 2197
  • Start date
homerdog said:
Is there a reason IHVs have mostly avoided odd numbers of memory channels? Are they better in pairs or something?
Click to expand...
I found this answer:
forums.developer.nvidia.com

Memory transaction size

That’s a lot of questions External Media I’ll just try to answer two… On G80 to GT200 architectures, 32-bit GDDR3 RAM chips are grouped in pairs, and every transaction is routed to a pair of chips put in parallel. This is why bus widths of NVidia GPUs are all 64-bit multiples. Then, each...
forums.developer.nvidia.com forums.developer.nvidia.com
 
TopSpoiler said:
I found this answer:
forums.developer.nvidia.com

Memory transaction size

That’s a lot of questions External Media I’ll just try to answer two… On G80 to GT200 architectures, 32-bit GDDR3 RAM chips are grouped in pairs, and every transaction is routed to a pair of chips put in parallel. This is why bus widths of NVidia GPUs are all 64-bit multiples. Then, each...
forums.developer.nvidia.com forums.developer.nvidia.com
Click to expand...

I believe this was changed with Pascal.

GM204 -


GP104 -


Hence likely also why the we saw 11 channel configurations used for Pascal (1080ti) and the next generation in Turing (2080ti) as well.
 
homerdog said:
Is there a reason IHVs have mostly avoided odd numbers of memory channels?
Click to expand...
The only reason is that it leads to rare configurations like 10 or 11GB. Anything rare is in danger of being omitted from developers optimization and QA processes.
I was thinking recently that with 3GB G7 modules there are now more options for such configuration than previously. You could have a 224 bit GB203 with 21GB of memory for example. Or a 160 bit GB205 with 15GBs. The disabling of one MC and associated L2 partition could offset the cost of denser memory modules. But such configurations could be a challenge for many developers who still optimize to some fixed memory amount progression like 8/12/16 GBs.

arandomguy said:
I believe this was changed with Pascal.
Click to expand...
I think the change is more related to memory type which is being used. At some point 64 bit memory channels became 32 bit ones.
 
That’s not the reason. It would particularly complicate any logic involved in request-to-channel balancing, which is critical for performance. So unless the odd number is 1, it’s then very easy to end up with lower utilisation of the memory system. Design and validation in hardware and then any tuning in software is not juice that’s worth the squeeze (otherwise you’d see it more often).
 
Rys said:
That’s not the reason. It would particularly complicate any logic involved in request-to-channel balancing, which is critical for performance. So unless the odd number is 1, it’s then very easy to end up with lower utilisation of the memory system. Design and validation in hardware and then any tuning in software is not juice that’s worth the squeeze (otherwise you’d see it more often).
Click to expand...
Doesn't look like 1080Ti had any issues with request-to-channel balancing.
 
Seanspeed said:
Everything else will be a fair bit lower than that.
Click to expand...
Let's wait for benchmarks. The selection which Nvidia used for comparisons without FG remains very weird - two AMD sponsored titles, one console title w/o RT and one with RT shadows. I will be surprised if other titles like those which make heavy use of RT will show the same results as Horizon for example.
 
DegustatoR said:
Let's wait for benchmarks.
Click to expand...
You can see the specs and understand what's coming the same as I can. Nvidia isn't exactly shy about touting raw performance gains if they've got something great, either. But it'll probably be easier to bury/deflect such talk with other things about the reviews when they come, complaining that some of them weren't fair to Nvidia or weren't testing frame generation and Nvidia's advantages enough, blah blah. I can already see the HUB topic going crazy.

Would be very glad to be wrong, but the writing has been on the wall for a good while, and these recent leaks about performance are pretty much confirming what could reasonably be guessed beforehand.
 
Seanspeed said:
You can see the specs and understand what's coming the same as I can.
Click to expand...
No, I can't. The architecture is different enough to avoid applying previous patterns to how it will perform. Shading core seem to be different, RT units are different, tensor h/w is different, even memory controllers are different again.
 
Seanspeed said:
It's gonna be a very disappointing generation by most standards.
Click to expand...

It’s shaping up to be a non-event for most owners of current generation cards. Maybe people shopping in the $400-$600 range will get something enticing from AMD with FSR 4.

Lots of RDNA 2 and Ampere owners out there though so everything will still sell.
 
DegustatoR said:
Doesn't look like 1080Ti had any issues with request-to-channel balancing.
Click to expand...
Any is a strong word. The product did very well but I’d be surprised if there were zero issues from a product bringup and tuning perspective. My point is more that the GP102 in that product wouldn’t ever have been designed with 11 channels natively.

There’s a reason we’ve rarely (never as far as I can remember, at least in the last 20 years) seen a native odd numbered channel count in a GPU before.
 
I’ll wait for reviews but my enthusiasm is definitely fading. I’ll probably still get a 5090 for nvenc, better rt, framegen etc but it feels like I could’ve just bought a 4090 2 years ago instead. Oh well.

I think Ollie mentioned that the 4090 encoder still doesn’t hit 4K 120fps. If the 5090 also doesn’t get there it would suck.
 
trinibwoy said:
I’ll wait for reviews but my enthusiasm is definitely fading. I’ll probably still get a 5090 for nvenc, better rt, framegen etc but it feels like I could’ve just bought a 4090 2 years ago instead. Oh well.

I think Ollie mentioned that the 4090 encoder still doesn’t hit 4K 120fps. If the 5090 also doesn’t get there it would suck.
Click to expand...
yeah im just staying with my 4090 this gen and maybe the 6090 will replace it in the future
 
I think if you have a 60Hz display (yikes) or even a 120Hz display, then 40 series to 50 series looks like a bad upgrade. It's betting a lot on the adoption of all of the new neural shader stuff. Another two year wait is probably the best option. I have a 3080 and a 4080 would be, at worst, a card that gives me maybe 75% improved frame rate, which is a nice upgrade. Then I also get frame gen, which I currently don't have in a good form. Makes ray tracing viable etc. But 40 series to 50 series is maybe the weakest upgrade ever. The only good use case for that upgrade is really high refresh displays like 240, 360 and 480Hz. But that's basically one use case that relies on frame gen and someone having one of those displays or planning on buying one. They're expensive.
 
GB203 in particular looks like a straight rehash of AD103 -- similar die size, transistor density, clock rates and a minor bump in SM count. The GDDR7 support is a bright spot for bandwidth bottleneck situations.
TSMC wafer rates for latest nodes must really eyewatering even for Nvidia.
 
You must log in or register to reply here.
Back
Top