Galax GeForce RTX 5070 Ti 1-Click OC White

[ThomasK]

There were many leaks, it depends what you take into account, for example:

Performance
Price

Also straight from the horse's mouth:

AMD is set to release the Radeon RX 9070 Series shortly, but it probably won't match the performance of the RTX 5070 Ti. Instead, it seems it will be more comparable to the RTX 5070, which is also expected to be released soon. While these new cards cannot rival the RTX 5070 Ti in terms of performance, they are likely to be priced more competitively due to increased competition in this market segment.

Ok, so no facts, only speculation, as I thought.

Next.

 

[ratirt]

$750? Cheapest I found was around $1k.
I expected way more in terms of pricing. I'll pass

 

[10tothemin9volts]

@W1zzard
A question regarding all GeForce Blackwell GPUs' hugely improved AI TOPS number: According to NV's own "Specs" comparison table, the 5070 Ti has 1406 "AI TOPS" vs 4070 Ti' 641. Which theoretical benchmark and practical application can confirm this huge 2.19 times / 119% increase in AI TOPS?
(This huge difference in AI TOPS won't run/inference LLMs faster as for this only VRAM size (so one doesn't have to offload LLM layers from VRAM to RAM) and bandwidth matter.)

[..]

GTX 1070ti was 45% faster than the GTX 980
RTX 2070s was 29% faster than the GTX 1080 [no x70ti for Turing]
RTX 3070ti was 35% faster than the RTX 2080
RTX 4070ti was 17% faster than the RTX 3080
RTX 5070ti is just as fast as the RTX 4080

There's a saying: Names are sound and smoke, aka names don't matter. This is good for historic reference tho. More interesting is price to performance to power efficiency.

 

[W1zzard]

Which theoretical benchmark and practical application can confirm this huge 2.19 times / 119% increase in AI TOPS?

Quantize your LLM to INT4 or FP4

 

[Flogger23m]

Is this better than the ASUS Prime (also $750)? Would like to see TPU do a review on that if possible. Though I don't recall seeing any Galax cards for sale at newegg/Amazon/Best Buy over the last few generations when I was looking for 3070s or 4070s. What stores tend to get them in?

 

[10tothemin9volts]

Quantize your LLM to INT4 or FP4

You mean the quantizing process reflects the AI TOPS increase and is quicker, or after one quantizes to INT4 or FP4, it then reflects the AI TOPS increase? I guess the latter. I thought it would be still rather VRAM bandwidth limited? Can you add this to the GPU' AI benchmarks or post somewhere else where this massive 2.19 times increases in AI TOPS is confirmed in any benchmarks? Most consumer are not going to care about running their own LLM stuff, only if Windows or other apps offer it: Wonder which consumer apps reflect this 2.19 times increase in AI TOPS.

 

[W1zzard]

or after one quantizes to INT4 or FP4, it then reflects the AI TOPS increase?

FP4 is just 4 bits, vs 8 bits on FP8, so half the data = twice the performance, no surprises here. TOPS = "operations", doesn't specify how many bits

 

[10tothemin9volts]

FP4 is just 4 bits, vs 8 bits on FP8, so half the data = twice the performance, no surprises here. TOPS = "operations", doesn't specify how many bits

Yes, it would be no surprise, but then this 2.19x increase would be NV comparing 4 bits vs 8 bits and not putting a disclaimer on their site. A consumer can not know that NV is comparing 4 bits vs 8 bits (should this be the case (it may indeed be the case since NV is using the same node and also the same chips (same transistor count, size, etc.)) without a disclaimer. This means that the AI TOPS are basically the same performance (that's why I'm asking, I already have expected for this maybe to be the case). Putting no disclaimer on their site would be wow.. NV did this 4bit vs 8bit comparison in their Blackwell presentation like a year ago, but there they did put a disclaimer (well, they put it pretty well visible directly on the graph) that it is FP8 vs FP4.