What if they actually overclock well? Does that mean that the TIM is still bad? I am not trying to argue just point out that do you think Intel would risk and entire generation of their top end line if the performance was not there?
Assuming that all hwbot submissions for the 6950X are also with unchanged uncore, just like the 7900X, then based on those scores alone, IPC gains are marginal.They don't affect the scores because (1) on client CPUs, the uncore is already 50% faster than HEDT/server, and (2) overclocking 4.4 to 5 GHz is too small to notice any L3 starvation anyway. You are making IPC claims based on OC of over 60%, which isn't a comparable situation at all.
I am using a locked processor, so I can't adjust my uncore frequency to run the benchmarks you demand. On the other hand, since you are such an expert on overclocking, it should be trivial for you to open whatever OC app you have and type in a lower uncore clock. If you don't want to be convinced, of course nobody can convince you.
What if they actually overclock well? Does that mean that the TIM is still bad? I am not trying to argue just point out that do you think Intel would risk and entire generation of their top end line if the performance was not there?
I never implied zero gains, but that when you account for the already increased clock speeds, they don't look that impressive.Yes, and that uncertainty is normally distributed. 10% +/- 5% does not equal zero as you are trying to imply. It means the difference can range from 5 to 15 % with the most likely value still being 10%, i.e. the center of the distribution curve.
It's kind of hard to have more IPC when everything is stalled on L3.Assuming that all hwbot submissions for the 6950X are also with unchanged uncore, just like the 7900X, then based on those scores alone, IPC gains are marginal.
Even Intel's own numbers, despite being useless, point to the same being the case.
Uncore bandwidth limits are visible at stock in AVX workloads, but once you're talking about 60+% OC, you have problems in far more than AVX. Something as simple as cache latency, which is already bad on HEDT/server, is now 60% worse (relative to core), which will of course wreck IPC.I'm no expert in overclocking, never claimed to be one, you OTOH seem to be an expert in understanding cache frequency scaling.
Wasn't your biggest point of contention regarding cache frequency in the context of AVX? Cinebench isn't a good benchmark for testing AVX throughput.
As I said before GFLOPs don't matter for my use case, so I have ZERO intentions in testing something which isn't useful to me at all, and neither am I interested in overclocking, delidding, etc. which people do partly due to bragging rights, but mostly because they can.
What would you consider to be a good overclock for the 10 core chip with "normal" temps? 4.5GHz all cores? 4.7GHz? Just wondering.They will probably still OC, but thermal performance will be bad, especially with all those cores. We'll see soon enough how bad it is for overclocking.
They will probably still OC, but thermal performance will be bad, especially with all those cores. We'll see soon enough how bad it is for overclocking.
What would you consider to be a good overclock for the 10 core chip with "normal" temps? 4.5GHz all cores? 4.7GHz? Just wondering.
Intel is consolidating all their packaging processes to use TIM. Perhaps it is for environmental reasons, perhaps yield, perhaps to save 50 cents per chip on solder. It will not be the end of the world, since Xeon Phi is using TIM and has a 260 W TDP. Sure, extreme OC to 5 GHz or more will require delidding, which is disappointing, but I would expect non-delidded performance to still be competitive with Zen, which has a hard 4 GHz ceiling.Such pessimism! Let's wait for reviews before drawing conclusions either way. I doubt that Intel switched from solder to TIM to save money on these expensive chips, so there might actually be a good technical reason for it.
AVX2 throughout seems to be limited by L2 size, not L3. No idea about frequency, because obviously the scientific establishment doesn't overclock.It's kind of hard to have more IPC when everything is stalled on L3.
Uncore bandwidth limits are visible at stock in AVX workloads, but once you're talking about 60+% OC, you have problems in far more than AVX. Something as simple as cache latency, which is already bad on HEDT/server, is now 60% worse (relative to core), which will of course wreck IPC.
The reason nobody overclocks uncore isn't that it doesn't work. It's simply not efficient when the same power increase can be used to raise core frequency.
Exactly. That reason is simplicity which is mostly equivalent with lower cost. TIM is worse for the common end-user. That's the point. Yet the still do it because it saves money.It is highly probably that they switched to TIM for a real viable engineering reason.
Announced Skylake-X lineup, price per core:If you look at a specific price point, Intel did virtually nothing different here with Skylake. Take the ~$1k chip for example (it is the one of the easiest to compare since Intel has been producing it for a decade).
*snip*
Every 3 to 4 years, Intel has added 2 cores to this ~$1k chip. The base clock is generally in the lower 3 GHz range. The cache has moved around a bit, generally getting bigger, but not always. They did the exact same with Skylake-X, added a couple cores every 3 to 4 years, just like they have been doing like clockwork.
I just don't see a massive change here. Not compared to how much they were usually changing it. What they did do was add more price points, some with less features than before and some with more features. But they started doing that in 2016 with the introduction of a new ~$1700 price point.
Facts:
In my opinion, the decision is linked to cost efficiencies driven out of the manufacturing process or just the reduced cost of raw material.
- Solder provides better heat transfer than this paste/TIM Intel use.
- AMD have shown that there isn't an engineering limitation (Ref Skylake-X) as they use it on Ryzen and so did Intel on previous gens!
Exactly. That reason is simplicity which is mostly equivalent with lower cost. TIM is worse for the common end-user. That's the point. Yet the still do it because it saves money.
Claim 1: TIM is better for long-term reliability(as per speculation)Actually, neither of those are actual facts. Solder doesn't mean better heat transfer than the paste/TIM Intel uses. It *may* mean better heat transfer at BOL, but that says nothing about the heat transfer throughout the life of the chip. Similarly, AMD hasn't shown that there isn't an engineering limitation, only that they may not care about the limitation. As an example, leaded gas works great, but it isn't like it doesn't have issues.
Actually, neither of those are actual facts. Solder doesn't mean better heat transfer than the paste/TIM Intel uses. It *may* mean better heat transfer at BOL, but that says nothing about the heat transfer throughout the life of the chip. Similarly, AMD hasn't shown that there isn't an engineering limitation, only that they may not care about the limitation. As an example, leaded gas works great, but it isn't like it doesn't have issues.
People should be grateful, maybe criticize Intel a bit for taking so long to finally bring TIM in HEDT. The platform was in dire need of long term stability. /sTIM is actually likely better for the common end-user, actually.
You forget other 5 or so materials that are part of indium solder.What IS a fact:
Indium: 81.8 W/m K
Thermal Grizzly Cryonaut: 12.5 W/m K
Conductonaut(tin, indium, gallium): 73 W/m KYou forget other 5 or so materials that are part of indium solder.
Rhys Coleman said:Asus confirmed to us the news of X299 chips’ lack of solder is indeed true. We spoke to one of their reps during their motherboard workshop at Computex yesterday about the overclocking potential of a delidded Skylake-X CPU. They've been seeing a 30°C drop in temperatures by popping the top on the chip and using a liquid metal thermal interface.
He went on to explain Skylake-X shares the same issue as Kaby Lake - temperatures rise at lightning speed when upping the voltage going into the processor, meaning 1.3V of power is roughly the maximum amount of juice you’ll want to pump into it without serious cooling. The Asus team did manage to get 4.3GHz out of the ten-core 7900X with 1.25V of power though, so the chips will still be pretty capable even at default voltages.
Actually, neither of those are actual facts. Solder doesn't mean better heat transfer than the paste/TIM Intel uses. It *may* mean better heat transfer at BOL, but that says nothing about the heat transfer throughout the life of the chip. Similarly, AMD hasn't shown that there isn't an engineering limitation, only that they may not care about the limitation. As an example, leaded gas works great, but it isn't like it doesn't have issues.
Lead thermal material MRB failure investigations & leveraged excursion management to implement process improvements at suppliers – optimized supplier casting & heat spreader plating processes (>$20M cost savings).
Developed & qualified Halogen Free IHS Flux, Elastomer TIM & corresponding suppliers (including one brand new supplier for Intel that resulted in >$10 million savings).