Discussion Nvidia Blackwell in Q4-2024 ?

[blackangus]

Neither is delayed.

Nice, just seeing rumors for the opposite saying consumer is pushed back to 2025.

 

[adroc_thurston]

saying consumer is pushed back to 2025.

It was always somewhere H1'25.

 

[jpiniero]

It was always somewhere H1'25.

That would be a surprise that it was always 1H 25 given that they've stuck to 2 years. And with no mobile refresh, that's a long time to go between new stuff if you assume that mobile won't be first. Even with AMD basically wiped out.

 

[Tigerick]

Guys, I have updated the frontpage with the GDDR7 info as standard memory config for RTX5090 series. As per MLID, NV has been stockpiling GDDR7 in order to prepare launching of Blackwell series:

• With GDDR7, RTX5090 still offer same memory bandwidth as 512-bit GDDR6X. Plus there is option for 36GB version (most likely going to be called RTX5090Ti, want to guess how much would it cost????)
• The rest of the lineup should continue using GDDR6X, that's mean RTX5080Ti has similar specs compared to current RTX4090.
• AMD has rumored to source 20Gbps GDDR6 for their upcoming RDNA4 GPUs, yeah pretty much confirm RDNA4 cards are mid end cards at best.
• It also means RDNA5 GPUs are not ready yet...

 

[jpiniero]

Guys, I have updated the frontpage with the GDDR7 info as standard memory config for RTX5090 series. As per MLID, NV has been stockpiling GDDR7 in order to prepare launching of Blackwell series:

• With GDDR7, RTX5090 still offer same memory bandwidth as 512-bit GDDR6X. Plus there is option for 36GB version (most likely going to be called RTX5090Ti, want to guess how much would it cost????)
• The rest of the lineup should continue using GDDR6X, that's mean RTX5080Ti has similar specs compared to current RTX4090.

While I wouldn't be surprised if the GB205/6/7 desktop products use GDDR6X, I would have to think that they would use GDDR7 with the GB203 ones. For 256-bit, you'd need 32 gbps to get similar bandwidth to the 4090. And since I assume it will have less L2 cache, it's going to need all the memory bandwidth it can get.

 

[Tigerick]

For 256-bit, you'd need 32 gbps to get similar bandwidth to the 4090.

256-bit GDDR7 provides 1TB/s memory bandwidth. If using 384-bit 24Gbps GDDR6X, it provides 1.15TB/s memory bandwidth, yeah NV definitely needs as much bandwidth as possible. Therefore 384-bit GDDR6X seems better choice for GB203.

As for L2 cache, RTX4080 is having 64MB L2 with 256-bit memory bus. If GB203 is having 384-bit bus, 96MB L2 cache is possible cause full die AD102 is actually having 96MB L2....

 

[jpiniero]

256-bit GDDR7 provides 1TB/s memory bandwidth. If using 384-bit 24Gbps GDDR6X, it provides 1.15TB/s memory bandwidth, yeah NV definitely needs as much bandwidth as possible. Therefore 384-bit GDDR6X seems better choice for GB203.

As for L2 cache, RTX4080 is having 64MB L2 with 256-bit memory bus. If GB203 is having 384-bit bus, 96MB L2 cache is possible cause full die AD102 is actually having 96MB L2....

It's much more likely for GB203 to be 256-bit and 64 MB L2 (or less). Remember that N3E has no cache scaling at all.

 

[Tigerick]

It's much more likely for GB203 to be 256-bit and 64 MB L2 (or less). Remember that N3E has no cache scaling at all.

And I have calculated 384-bit GDDR6X has bigger memory bandwidth, bring up your Excel to calculate...

And check out how big the AD102's L2 full die has...god

 

[Tigerick]

I think I know why NV has intentionally crippled amount of L2 on AD102 from 96MB to 72MB..it is all related to upcoming Blackwell series. L2 cache is crucial to GPU performance, NV won't make it lesser, and I know N3E has no cache scaling, it just means NV has to use bigger die area if they have to...

That's why there was rumored about 128MB L2 cache associated with 512-bit memory bus. Now that NV has chosen to use 384-bit GDDR7 bus, I have reduced amount of L2 cache to 96MB; same amount as AD102 full die. Still bigger than RTX4090...and that's how NV ups the performance level of Blackwell series.

Same as RTX4080Ti Super, NV has used bigger die AD103 but crippled amount of L2 cache. The reasons are the same, make road for upcoming RTX5070Ti which amount of L2 cache will bump from 48MB to 64MB. And that is part of the NV's recipe to fight with AMD with GDDR7 memory...

 

[jpiniero]

I think I know why NV has intentionally crippled amount of L2 on AD102 from 96MB to 72MB..

4090 you mean, and probably yield and/or segmentation. There's a Quadro which has the full AD102 configuration including 96 MB of L2.

 

[Tigerick]

4090 you mean, and probably yield and/or segmentation. There's a Quadro which has the full AD102 configuration including 96 MB of L2.

Yeah, part of the reason but NV has not crippled AD103 on RTX4080. Go figure...And I think NV has every intention to use 384-bit bus which associated with bigger L2 cache

 

[jpiniero]

I see it like this:

GB202: 384 bit
GB203: 256 bit
GB205: 192 bit (or less)
GB206: 128 bit
GB207: 96 bit, maybe 128

 

[Tigerick]

I see it like this:

GB202: 384 bit
GB203: 256 bit
GB205: 192 bit (or less)
GB206: 128 bit
GB207: 96 bit, maybe 128

NP, why not you create a table like I did and list down your speculated specs and see whether it make sense or not? Trust me, it will be fun

 

[Tigerick]

I have updated the table to include possible choice of GB206 die use as upcoming RTX5060Ti. Cause RDNA4 seems like supporting 192-bit 20GBps GDDR6, if NV continue using smaller die like 4060Ti, they will lose to AMD like 7800XT. Knowing Jensen, he most likely uses the GB206 (same one as RTX5070) with 192-bit GDDR6 memory which provides much higher memory BW...we shall see

 

[jpiniero]

NP, why not you create a table like I did and list down your speculated specs and see whether it make sense or not? Trust me, it will be fun

No. But I think the faster GB203 will be a bit faster than the 4090 in raster and the slower GB203 being slower than the 4090 but decently better than the 4080S. The slower being $1199 or so.

GB205 is going to be the real question since I think it's going to be a lot slower than any of the GB203 SKUs.

 

[Tigerick]

No. But I think the faster GB203 will be a bit faster than the 4090 in raster and the slower GB203 being slower than the 4090 but decently better than the 4080S. The slower being $1199 or so.

GB205 is going to be the real question since I think it's going to be a lot slower than any of the GB203 SKUs.

Then why do NV release something slower to compete?? That's why I assume GB205 comes with 256-bit memory bus which compete with RDNA5 with 256-bit GDDR7. If NV is not bring up the best of BW, then they will lose out, do you think NV, the AI company don't know the competitors??

And why did NV release RTX4070Ti Super with 256-bit 16GB then downgrade to 192-bit???

 

[jpiniero]

And why did NV release RTX4070Ti Super with 256-bit 16GB then downgrade to 192-bit???

NV has options. They could just sell both G6X and 3 GB G7 of the faster GB205 if they think there will be haters.

 

[jpiniero]

Oh, forgot to add. I think the faster GB205 model is going to struggle to beat the 4070 Ti Super in raster. I think clocks would have to be really high or there be big improvements in the SM. That obvs would mean there would be a huge gap between it and the slower GB203 if that's going to be a tad slower than the 4090.

There would be room to do something inbetween with AD103... but the marketing would be kind of weird if they do greatly improve the RT with Blackwell.

 

[Tigerick]

Hoho, RGT has posted video about Blackwell lineup, he also think whole lineup will use GDDR7. Does he really count the memory config and size? I have created a table on the frontpage to compare my version versus RGT's version. I really hope after seeing the table no members are fool enough to believe such config make sense...

 

[jpiniero]

Hoho, RGT has posted video about Blackwell lineup, he also think whole lineup will use GDDR7. Does he really count the memory config and size? I have created a table on the frontpage to compare my version versus RGT's version. I really hope after seeing the table no members are fool enough to believe such config make sense...

Believe there was an actual legit leak that suggested that GB202 is indeed 192 SM but is still 12 GPC.

My totally accurate calculations has it as:

GB202 - 192 SM (12*8)
GB203 - 112 SM (7*8)
GB205 - 48 SM (4*6)
GB206 - 24 SM (3*4)
GB207 - 18 SM (3*3)

With an Ada rebrand (or two) filling in the gap between GB203 and GB205. Yeah I know the lower tier products have less SMs, I am accounting for a big drop in die size in those to make up for N3E's prices and lots of room taken up by Tensor/AI. Perhaps the clock speed gains will be there.

 

[maddie]

Will most of the older PC owners with PCIe 3.0 have problems with GB107 and the x4 connection?

 

[jpiniero]

Will most of the older PC owners with PCIe 3.0 have problems with GB107 and the x4 connection?

Don't worry, RGT so it's fake nooz

 

[SteinFG]

With an Ada rebrand

There's no ada rebrand. Nvidia haven't mixed 2 gens in one series since Geforce 700 series ten years ago. AMD also stopped doing it after GCN 3.0 8 years ago.

 

[TESKATLIPOKA]

My totally accurate calculations has it as:

GB202 - 192 SM (12*8)
GB203 - 112 SM (7*8)
GB205 - 48 SM (4*6)
GB206 - 24 SM (3*4)
GB207 - 18 SM (3*3)

With an Ada rebrand (or two) filling in the gap between GB203 and GB205. Yeah I know the lower tier products have less SMs, I am accounting for a big drop in die size in those to make up for N3E's prices and lots of room taken up by Tensor/AI. Perhaps the clock speed gains will be there.

Even If GB could clock higher, It would also apply to GB202 or GB203.
Difference between GB206->GB205 is 100%.
Difference between GB205->GB203 is 133%.
That's just too much.
Your GB205,206,207 is just too cutdown, It could end up even slower than Ada104,106,107 for example.

In my opinion Nvidia will have to price It very competitively compared to previous gen or add >=20% more SM in combination with higher clocks(10%?).
And of course more Vram.

 

[jpiniero]

It'd be a lot faster in AI