Every now and then, a content piece falls to the wayside and is archived indefinitely -- or just lost under a mountain of other content. That’s what happened with our AMD Ryzen pre-launch interview with Sam Naffziger, AMD Corporate Fellow, and Michael Clarke, Chief Architect of Zen. We interviewed the two leading Zen architects at the Ryzen press event in February, had been placed under embargo for releasing the interview, and then we simply had too many other content pieces to make a push for this one.
The interview discusses topics of uOp cache on Ryzen CPUs, power optimizations, shadow tags, and victim cache. Parts of the interview have been transcribed below, though you’ll have to check the video for discussion on L1 writeback vs. writethrough cache designs and AMD’s shadow tags.
Part of our 4K camera upgrade was for ergonomics – better ability to handle the camera, particularly in show floor environments – with most the other reasons centering around quality. Camera quality is superior in every technical sense, low-light and noise reduction being a major area of improvement, but working with larger files at higher bit-rates means longer render times. We can now capture up to 200Mbps (previously captured 28Mbps) at 4K resolution, and we output at 2x the bit-rate of our previous 1080p60 videos. Render times have skyrocketed, as you’d expect, and have gone from roughly video duration + a few minutes to an hour per 20-minute video.
There’s not a lot we can do about this. Adobe Premiere, sadly, does not really do much with multi-GPU. The GPUs are accelerators, with rendering still falling on the CPU for a lot of the workload. We’re becoming more thread-limited than anything at this point, and really don’t want to build an entirely new production system right now. For now, upgrading the primary GPU to a 1080 Ti will help us out a bit in Premiere and significantly in Blender.
For those who don't follow the YouTube channel as closely as the website, it's possible that you may have missed out on our first two livestreams. Both have VODs up on the YouTube channel over here: GN Live #1 - Seidon Cooler Tear-Down & GN Live #2 - EK Open Loop & Vega Work.
We don't have any plans to start a regularly scheduled stream, but we are working on just streaming when the team is building/unbuilding things anyway. For now, we wanted to take apart and build the stuff shown in streams #1 and #2, so it made sense to set them live.
“Good for streaming” – a phrase almost universally attributed to the R7 series of Ryzen CPUs, like the R7 1700 ($270 currently), but with limited data-driven testing to definitively prove the theory. Along with most other folks in the industry, we supported Ryzen as a streamer-oriented platform in our reviews, but we based this assessment on an understanding of Ryzen’s performance in production workloads. Without actual game stream benchmarking, it was always a bit hazy just how the R7 1700 and the i7-7700K ($310 currently) would perform comparatively in game live-streaming.
This new benchmark looks at the AMD R7 1700 vs. Intel i7-7700K performance while streaming, including stream output/framerate, drop frames, streamer-side FPS, power consumption, and some brief thermal data. The goal is to determine not only whether one CPU is better than the other, but whether the difference is large enough to be potentially paradigm-shifting. The article explores all of this, though we’ve also got an embedded video below. If video is your preferred format, consider checking the article conclusion section for some additional thoughts.
Ask GN returns for its 54th episode – we’ve gotten more consistent than ever – to discuss Noctua fan manufacturing locations (China & Taiwan), thermal pads vs. thermal paste usage on MOSFETs, Vega 10-bit support, and a couple other items.
A few of the items from this week peer into GN’s behind-the-scenes workings, as several viewers and readers have been curious about our staff, whether we keep products, or why we “waste” GPUs by using them for things other than mining.
As always, timestamps below the embed.
Over the last few months, we’ve talked numerous times about the high prices of flash-based storage devices, specifically SSDs and DDR4 RAM. Newegg currently has a “Spotlight Special” sale on SSDs and Memory, and while the prices are not historically low, they are a slight break from the current overpriced market we find ourselves in. All products listed should be available at this price until Monday, July 24th.
This feature benchmark dives into one of the top requests we received from our Patreon backers: Undervolt Vega: Frontier Edition and determine its peak power/performance configuration. The test roped us in immediately, yielding performance uplift largely across the board from preliminary settings tuning. As we dug deeper, once past all the anomalous software issues, we managed to improve Vega: FE Air’s power available to the core, reduce power consumption relative to this, and improve performance in non-trivial ways.
Although power target and core voltage are somewhat tied at the hip, both being tools for overclocking, they don’t govern one another. Power target offset dictates how much additional power budget we’re willing to provide the GPU core (from the power supply) in order to stabilize its clock. GPU Vcore governs the voltage supplied, and will generally range from 900 to 1250mv on Vega: FE cards.
Vega’s native DPM configuration runs its final three states at 1440MHz, 1528MHz, and 1600MHz for the P-states, with DPM7 at 1600MHz/1200mv. This configuration is unsustainable in stock settings, as the core is both power-starved and thermally throttled (we’ll show this in a moment). The thermal limiter on Vega: FE is ~85C, at which point the power and clock will fluctuate hard to try and maintain control of the core temperature. The result is (1) spikey frequencies and frametime latencies, worsening perceived performance, and (2) reduced overall performance as frequency struggles to maintain even 1528MHz (let alone the advertised 1600MHz). To resolve for the thermal issue, we can either configure a more intelligent fan curve than AMD’s stock configuration or create a Hybrid card; unfortunately, we’re still left with a new problem – a power limit.
The power limit can be resolved in large part by offsetting power target by +50%. Making this modification is easy and “fixes” the issue of clock-dropping, but introduces (1) new thermal issues – resolvable by configuring a higher fan RPM, of course, and (2) absurdly high power consumption for a non-linear scaling in performance. In order to truly get value out of this approach, undervolting seems the next appropriate measure. AMD’s native core voltage is far higher than necessary for the card to operate at its 1600MHz target, and so lowering voltage improves performance from the out-of-box config. This is for thermal and power reasons alike. We ultimately see significantly reduced power consumption, to the tune of ~90W in some cases, a more stable core clock and thereby higher performance, and lower temperature – and thereby controllable noise.
This week has been another slow one in the hardware sales department, as we seem to say every week that flash-based storage and GPUs continue to stay high in price. Upgrading or building a new PC is difficult right now; however, this week we found some decent prices on a 16GB kit of DDR4 from G.SKILL and a 120GB SSD from Kingston. With the GPU cryptocurrency mining craze still going on, mid- to low-end cards are all but sold-out and overpriced for gaming, but the GTX 1080 however can still be found for at or just under $500 in some cases. This week also provided us with a slight discount on the new Corsair K63 keyboard with Cherry MX red switches.
We can’t get all the way down to the inner workings of the pump on this one, unfortunately, as all of our source images for the Vega: Frontier Edition – Watercooled card are from a reader. The reader was kind enough to remove the shroud from their new WC version of Vega: FE so that we could get an understanding of the basics, leading us to the conclusion that AMD has built one of the most expensive pre-built liquid cooling solutions for a graphics card.
The video tear-down goes into detail on the images we received, but we’ll revisit most of it here. The card uses the same base PCB, same VRM, same GPU/HBM layout and positioning, and same everything as the air-cooled card. The difference is entirely in the cooling solution, where the Delta VRM fan goes away and is replaced with an additional reservoir (more on that in a moment), while the GPU/VRM cooling is handled by liquid plates and a pump. The die-case finstack atop the I/O is also now gone, and the baseplate is simplified to an aluminum plate with no protrusions.
This week's hardware news recap gives us a break from Vega -- if a brief one -- so that we can discussed nVidia's multi-chip GPU white paper, AMD's Ryzen Threadripper CPUs (1920X + 1950X), the R3 CPUs, and new fabs for Samsung. This discussion also bleeds over into DRAM shortages and NAND prices, particularly relating to Micron's fab "event" from last week.
The show notes are below the embedded video, for folks who prefer the notes and sources.